They've got, ballpark, $5t to $10t to make back in the next 5 years, or the hardware buildouts will start getting written down.
This means we're going to need $1t+ per year in spending, per year, on tokens. 200m knowledge workers in the world, 30m developers. We're talking about a world where you need 5% of every knowledge workers salary to go into tokens. 20% if you're a developer.
That's a _huge_ shift. Most people I know cite +20%-40% velocity with these tools, against the actual work their company cares about doing. +20% speed for +20% spend isn't going to motivate a trillion dollars a year in spending.
We're not there yet. This is still the upswing of the hype cycle, and unless we figure out how to make developers 2x, 5x, 10x as productive on stuff that matters, this isn't going to play out well.
- The publicly available information about how inference costs compare to training costs is conflicted. EEs involved in datacenters talk about power usage spikes during training runs as if they were a major factor in the designs, but academic papers discussing cost-optimal scaling confidently treat inference-time compute as a major factor.
- On the side of the balance indicating that training is more compute-intensive after amortization than inference is that Chinese providers, constrained primarily by access to compute, have nearly unlimited token availability at a lower price than US providers (inference), but poorer model capabilities (training). That would make sense only if US providers are inflating inference costs by 20-30x due to amortized training costs that overseas providers were not able to take on (there are other factors too).
- If training >> inference, they're in a prisoner's dilemma that far exceeds the ordinary zero-marginals model of competition between firms (due to its huge discrete stepwise nature). On the other hand, if inference>>training, the high-level analysis popularized by certain thought leaders, that it's like a utility, would be true. You'd tend to count this as a vote for inference>>training, but the CEOs saying it at least have a huge incentive to agree because the alternative, the prisoner's dilemma, would stop investment very fast.
- The only voice in the story that I just told you to have anything to do with fact (as opposed to high-level analysis and ivory tower armchair management of a secretive business) were the rumors from facilities engineers. That shows you the state of our understanding...
- If we don't even know the ratio between amortized capital expenses and operational costs, outside investor analysis is impossible. It doesn't matter how finely they divide the accounting buckets for office ferns and indoor ferns if the single biggest part of their business is obscured for trade secret reasons.
I'm about to leave a shallow comment, but I am a bit skeptical of the supposed drop in inference costs. If AI labs saw a lot of potential there, they'd surely be bragging about it non-stop? So the fact that publicly available information is conflicted is probably a sign that at the very least, the numbers aren't amazing.
Yes I know there's no evidence and this is lazy reasoning. But there's probably a bit of truth to this line of thought.
Why on earth would AI labs be bragging about how little the product they sell actually costs them to make? You don't want to do anything that reduces it's perceived value to the user, that might make them less willing to pay for it.
Also, inference costs are bound to go way down with more optimized architectures. GPUs are fundamentally not great at inference. No platform where the weights are streamed from a large pool of memory is. If the models ever quiet down, there will be massive step changes in cost/token, energy/token and tokens/second, as models are etched into silicon ala https://chatjimmy.ai/
A couple of years ago Altman was saying the price of AI compute is going to drop 90% year over year or something like that, so I don't think they're nervous about talking about lowering their costs. They probably just haven't been able to lower their costs.
You have to keep in mind that about 99% of their announcements are targeted towards investors (their most important revenue source..), so they're not going to be afraid to mention metrics that make the business look better.
> Why on earth would AI labs be bragging about how little the product they sell actually costs them to make? You don't want to do anything that reduces it's perceived value to the user, that might make them less willing to pay for it.
Wouldn't they be bragging about it to investors? It feels like something that would matter a lot to them, and at least OpenAI kinda feels desperate to find them.
There's also the small question about whether a drop in inference cost would actually change anything about profitability, when training seems to get exponentially more expensive.
In the US, companies are not allowed to unfairly privilege some investors over others by giving them access to secret information that would let them judge the future prospects of the company. (Except in all the ways they can, but these usually involve some kinds of insider trading rules.) Private companies can handle giving out secrets to investors by literally writing and memo and mailing it to all their investors, if they want to give out some secrets to one of them.
Public companies cannot do that, even if they knew who all their investors were, but must instead consider every member of the public a potential investor, even if they don't already own the stock. Because of this, when public companies want to reveal material information about their future prospects, they must reveal it to everyone.
Besides the legal requirement, the reason these companies go public is often to provide liquidity for early investors or employees. So they do want to have as good of a margin story that they can, at least in terms of unit margin.
This is an interesting anomaly in the US. In the civilised world all corporations have to file public accounts, as the price for their limited liability. The detail and audit requirements depend on the size, turnover, staff numbers etc. This is because the shareholders are not the only stakeholder. The companies creditors, for instance, who are exposed to the limited liability have a right to see what they are lending to.
To answer the sibling comment, all of these public accounts follow local GAAP or IFRS.
The US still astounds me with its willingness to allow corporations to rip people off!
Creditors in the US can make visibility into financials a requirement for financing if they want. Protecting creditors isn’t a good argument for public reporting.
As a consumer you are often sending deposits or even the full cost of goods to companies some time before you receive those goods (in effect you become a creditor). You are also dependent upon some of those companies for service and repairs. It seems reasonable that you can check the finances of a company you are creating a business relationship with, I know in the past I've checked company statements.
You are unlikely to have significant enough sway to force that kind of disclosure. Small businesses as consumers have less legal protection and are similarly unlikely to be able to make disclosure a precondition of a deal.
Isn't there a limit on the public markets where if a company has less than a certain percentage of its ownership traded publicly then it is no longer a public company and therefore de-listed?
I remember hearing about a guy trying to squeeze out short sellers of his own company but ended up effectively taking his company private because he bought out like 95% of all the shares.
I wonder how that aligns to these small releases of stock for the public.
It's not a big secret. If you just do the math yourself, it's easy to compute that inference doesn't cost all that much. People just see all the capital investment going around and all the new data centers being built, see that it's spent on "AI", put two and two together and get a three, or "clearly serving AI requests costs an arm and a leg".
The 1 they were missing is that AI requires both training and inference, and training is by far the expensive part. And that in principle you can stop training at any point and keep using the models as they are. (But that means that if other companies keep improving their models, you'll be left behind...)
In contrast, inference is fairly cheap and all the providers have great margins on it. Eventually either investment in training stops having commensurate impact on model quality, and people stop doing that and instead concentrate on making inference faster and even more efficient. Or if that doesn't happen, things will get very weird very quickly.
Just look at large open weights models being served by inference providers.
Kimi 2.6 is a 1 trillion total / 32B active parameter model that's something comparable to Sonnet. Sonnet's API pricing is $5 in, $15 out per million tokens. Deepinfra serves Kimi at $0.75 in, $3.50 out, and about the same at openrouter. So you're looking at a 4-7x multiple that Anthropic is charging compared to market rates that any plebe can get with a credit card.
I'm not sure just how good that looks for Anthropic/OpenAI.
4-7x isn't a tiny markup, but how does that compare to high-margin internet businesses like AdSense? Meta and Google do hundreds of billions in ad revenue a year, and after taking out the publisher's portion (60-80% per some searching), I wonder what the ratio of the remaining tens-of-billions is against the compute cost and headcount required to run it.
And how much room for maintaining or improving that margin do they have if the cheap competitors also continue getting better? Is there a "good enough" point where the easier inference tasks are all moving to vendors massively undercutting them, and then they don't have the volume necessary to justify spending on further cutting-edge development?
> Kimi 2.6 is a 1 trillion total / 32B active parameter model that's something comparable to Sonnet.
No it's not. On some rigged paper maybe. Some such benchmarks say all models group together, which they clearly do not.
> Sonnet's API pricing is $5 in, $15 out per million tokens. Deepinfra serves Kimi at $0.75 in, $3.50 out, and about the same at openrouter. So you're looking at a 4-7x multiple that Anthropic is charging compared to market rates that any plebe can get with a credit card.
That's not saying much. You can get "cloud" at AWS and you can get a VPS. There is likely a 10x difference. It's not "same". Whilst AWS costs more they also don't have 7x margins similarly.
Growing companies don't brag about their margins, they brag about their growth and revenue. Margin talk is for when you're a mature company squeezing out every bit of profitability you can - if anything it would be a negative sign to be worrying about your margins when you're supposed to still be growing and innovating.
Because companies that want to go public need to look profitable or potentially profitable. And before they go public they have to release real, actual, legally demonstrable numbers for their costs and revenue anyway.
Because the most important thing for any pure play AI company right now is to prove they are a viable company. And sure they have proved they can make billions, but also that they can lose billions more. They are going to need even more money and to prove to the next round of investors at an even higher valuation that they are a viable business they need to show not that they can generate revenue, but that they can one day turn a healthy profit. And that is the trillion dollar question.
Not super clear from the site itself, but this LLM is running on specialized silicon implementing just it. So has super low energy use and blazing speed.
Because they can think more than one quarter into the future? Why on earth would someone adopt something into their core workflow that was fantastically unprofitable? Uncertainty and business don’t mix. Most people aren’t hype-eating bacteria that only care about maximizing their next paycheck.
One reason is that all the code you write with this goes in your private git. If using AI no longer is possible because of cost, you can still profit a lot from what you did with it before.
If inference costs drop 90% or whatever, that would be a massive write-off of hardware even before they gave any returns for it?! Given Chinese and others are snapping at the heels and would also benefit from such reduction in cost.
> Why on earth would AI labs be bragging about how little the product they sell actually costs them to make?
Investor confidence. They have a bit of a need for cash (also an interesting part of the profitability discussion of course).
> Also, inference costs are bound to go way down with more optimized architectures
I agree. Jimmy is incredible, I wonder what non-toy use cases they have. Surely they’ll come out with updated chips soon.
That said, I was apparently a bit over-excited for Groq and Cerebras. I thought they’d quickly dethrone Nvidia for inference, but not so far. Even the GPT spark trial isn’t seeming to go far.
Inference has traditionally been far less expensive than training. One public example is the fact that hobbyists can run StableDiffusion ($600k training costs[1]) on their personal computers.
Speaking to your point, inference being dramatically less costly than training would not be seen as a delta from the norm. The model of providing inference for anything near the operational costs (like a utility would), would the delta from the norm if it were true.
The difference between training and inference is 1) one have to keep intermediate results for backward pass in training and 2) computation for training double because of the backward pass.
Training is also done over batches, which increase memory requirements by several orders of magnitude. This is why training needs costly compute.
One of the ways out of this unfortunate situation is to use something like Stochastic Average Gradient Descent [1]. Examples there are mostly concerned with regularized logistic regression, which makes problem more or less convex. Neural networks are inherently non-convex. Still, maybe some ideas from there can be utilized in the context of neural networks, like use of estimated Lipshitz constant to derive curvature and appropriate learning step.
It's all got much more complex than that in recent years. Training now involves large amounts of inference for RL rollouts and similar. You can't disentangle them computationally like that. "Inference" is just the word used to mean serving customer traffic now, and "training" means creating the model you serve.
Multiply "inference + backwards pass (~2x inference cost) + activations (vram overhead)" by batch size (thousands) to get to the actual RAM and compute cost. Optimizer like ADAM adds only two or three model-sized overhead.
And last, but not least, you need only one hidden layer kept in RAM for inference, but you need all of them (61 for Deepseek models) kept in RAM for computing gradient for one sample.
Microbatch size is a hyperparameter, it can be set to 1 and work just as effectively. With gradient accumulation it's equivalent even. Large batch sizes are used to increase parallelism, and sometimes to reduce variance in the loss signal (at the cost of increased bias).
Batch size is frequently limited by compute bottlenecks well before memory.
And of course you do all of this for every object in your training set, which is going to be larger than the total number of uses for any individual user.
I think in your StableDiffusion example, a lot more than $600k will have been spend on electricity alone for inference (on those personal computers you mention). So inference is more expensive then training.
For equal capability tokens, there has been about a 10x drop in cost every 6 months.
We are still chasing the best because the best is moving rapidly, but it’s a simple thought experiment to work out what the cost to serve an 8B model from 2 years ago is in a world of 2T models.
Note: parameter counts are illustrative. Concretely, qwen3.6 27B delivers opus 4.5 capability at 1/27th the cost on openrouter. Single chip llama3 8b performance can exceed 17k tokens/sec.
> For equal capability tokens, there has been about a 10x drop in cost every 6 months
Is this still happening? Opus 4.5 was six months ago, can you get its capabilities for 1/10 cost now? Are we on track to get the same for 4.6 in a couple months?
8B models would be consider obsolete in the world of 2T models, at least if we're talking about the competitiveness of OpenAI/Anthropic. The only reason why they are valued so highly is their supposed dominance at the top end.
The main story of agent use cases is in enterprise so far. An enterprise will only pay for a model capable of handling the task and no more. Most enterprise's see no need to hire PhDs as factory line workers.
Coding is an interesting case as [1] the pace of progress has been absurd and [2] it's hard to put an upper bound on required capability. However hard to put a bound on and will are different, it's quite possible that the average engineer will cease to see the benefit of rapid progress - or that their employer will be satisfied with lower tier models.
How smart of a model do you need to build a high quality CRUD app for internal users? Or build a scalable web service?
> If AI labs saw a lot of potential there, they'd surely be bragging about it non-stop?
Google seems to pretty regularly post about how their TPU and algorithm advancements have been decreasing energy costs for both inference and training.
Small alternative potential future changes that alter this analysis:
* At some point model capability reaches diminishing returns. Then inference >> training in the future but training >> inference now. It’s not a prisoner’s dilemma but a land grab to solidify market position and be one of the 2-3 firms left standing as dominant in the space. The model companies aren’t super sticky yet but they’re working on it.
* even if training remains >> inference, it’s possible to have multiple price points like they do today. If you need the most capable model you’ll be paying exponentially more per token to supplement the training cost even though the serving cost is marginal because most people will be satisfied with cheaper / less capable models for most tasks.
I buy that inference is a dropping line item while training is a growing one. There’s all sorts of things on the horizon that’ll be order of magnitudes improvements, from startups burning models into ASICs to get order of magnitudes more performance to alternate architectures like diffusion transformers that have orders of magnitude structural optimizations. It’s inevitable that it’ll come down even further from where we are. It’s possible model training also will go down but I’ve not seen any compelling research suggesting major “easy” reductions here.
The issue is that most tasks do not require frontier-level intelligence, but companies like OAI can really only profit off of the frontier. Capabilities from a year or two ago are so outdated that even OpenAI gives it away for free and there are many other models biting at their heels. In other words they are spending huge amounts of money to cash in on a depreciating asset.
So one possible future is that frontier-level training becomes so expensive and the use cases so sparse that it simply isn’t viable to keep going bigger.
We have GPU costs, power costs, and how many token/s models can generate on those GPUs. It’s possible to figure out the marginal cost based on this. The current estimate is about $0.40 per million tokens for gpt4 equivalent model. Sonnet 4 is $15 per million tokens, so they are charging high margins on inference. The issue is how large of a margin is needed to recover their costs before the GPUs age out, and how high of a margin can be charged before it’s not economically viable.
I skimmed the article, but couldn’t spot any details on their estimates. They mention 70b+ params as being large in several places. But we’ve had several 100b+ param models that trail Sonnet.
Why would power spikes from training runs imply training>>inference? The cost of a training run scales with energy, whereas power is energy per unit time. All that tells you is that they're speeding up their training run so it will take less time overall (probably chasing some first-mover advantage, where they're out with a given model before their closest competitors), whereas they obviously can't do that for inference (which is a steady flow of requests over time).
I don't see how it would be possible for inference costs to dominate training costs, even after amortization.
Training involves multiple passes over the entire training dataset, ideally in large batches where you can perform inference on as many samples as possible simultaneously and then perform backpropagation to adjust the model weights (which is about as expensive as inference).
Let's consider the size of the dataset we're dealing with here. The dataset likely consists of practically every piece of digitized text they can get their hands on (including that extracted from audio and video). We know Google has digitized a large portion of the books in existence as part of their "search book contents" feature and we have no reason to believe they're not using it alongside their cache of 90+% of the internet to train their models. We're talking about 100s of millions of books each with an average of 100,000s of tokens. The internet has 10s to 100s of billions of pages on it with who knows how many tokens on average. This is a huge dataset that we've got to go through hundreds of times.
Second, let's consider the effect of batching and how it sets requirements for our hardware. We know that larger batch sizes converge faster, are more stable, and produce better models. So if you want a good model you need large batch sizes. This means that you need machines several orders of magnitude more powerful than you use for inference. From what I heard Google uses clusters of 100s of the their TPUs all located in a single rack for training. These clusters are organized in a customized computing architecture to maximize memory locality between cores (really critical for efficient back-propagation). Further, you can't use reduced precision weights for training like you can for inference, so there are no shortcuts.
Finally, the initial training stage is followed by reinforcement learning stages - this is key development in how AI models have improved in the past year. This may mean going through a curated set of traces (either synthetic or captured from users) and adjusting the weights based on experienced outcome.
Overall there's so many orders of magnitude more work and more hardware requirements for training that I find it improbable that inference dominates. The number of "inference" steps in training is freaking ridiculous and includes such factors as the "number of words ever written".
It's been a while since I saw a detailed paper on a high end training run, but extrapolating from what I remember, it seems those training runs are in the 10s of trillions of tokens. This already accounts for potentially sampling tokens multiple times during the training run.
That seems like a large number, until you realize that OpenAI claims to have almost a billion weekly users. And OpenRouter shows many models at over a trillion tokens per week.
So in pure token terms, I'd say it is in fact extremely plausible that inference dominates, at least for the popular models.
Yes the huge discrete stepwise training spend is critical.
Maybe investors will realise that "the only winning move is not to play".
And so we are left with (as was) frontier models getting more and more out of date as whoever their post bankruptcy custodians are tries to eek pennies on the dollar for inference on their decaying property. Perhaps along with local and/or highly specialized models still feeding on the after-glow of the huge amount of training that was (and is no longer) done.
The next AI winter is going to be deep, savage, and long.
> frontier models getting more and more out of date
Why are they getting out of date? Is it because we have new content from the internet that the older models did not have? Or are we simply trying to increase the size of the training data? In other words not more up-todate in terms of time the content was created vs. wanting to use bigger training-input-sets?
lack of new content from the internet will make them go out of date. Not just facts and figures, but (for example) new programming languages/techniques.
I work for a tiny little company ($150MM annual rev with 9% net) and we are already looking at dropping $100k on hardware to run local models because, for us, they're "good enough."
Our estimated spend for AIaaS would exceed that cost in less than a year.
In a few years, there will be hardware capable of running frontier models good enough for most things at accessible prices for even tiny companies.
Yeah, that's the part that just seems to be wildly under-discussed to me.
If open source models are ~3-6 months behind SOTA, and ~opus4.6 capabilities are good-enough for product market fit, do the frontier labs have half a decade to catch up on their prior burn?
AI cost ballooning faster than companies can afford is becoming a very common topic in my circles right now. The era of "I'll pay infinitely more for marginal gains" is over from what I can tell.
> If open source models are ~3-6 months behind SOTA, and ~opus4.6 capabilities are good-enough for product market fit, do the frontier labs have half a decade to catch up on their prior burn?
They know they do not and that’s why they’re all trying to IPO right now, so they can pass the bag to consumer investors
Open source models that you can run locally are much more than 3 to 6 months behind. 6 months was the November inflection for Claude. No open source model is as good as Claude Opus 4.6.
It depends what you mean by locally. I don't foresee running a model on my laptop anytime soon to power a coding agent. Far more likely is an infra team at my company operating an open source model on cloud infrastructure. When they're already paying $1000 / month / dev, it starts to pencil pretty quickly.
Not OP but I've been thinking about this a lot (like everyone ha) and I think my answer is, yes?
I hope there's a "good enough" point but I don't think we're there yet. Like for me hardware got good enough several years ago. But while opus 4.7 is really good compared to everything else, it's not so good that I would use it at a discount over whatever is available in a few months. The improvement in quality, speed, and daily frustration is worth it to me... Spoken as someone whose employer is footing the bill, so take that with a grain of salt.
I want to run my own local models, but I don't think that's feasible without lots of frustration until a few generations of frontier models are so good that they're almost indistinguishable for common tasks. Kind of like how MacBook pros have been for a while.
While I can imagine that I'd want to use Opus 4.8 over 4.6 for a fair number of things (at least if they can avoid further speed regressions), I also have noticed that certain types of failures seem to be systemic. Bigger context has been helpful for bootstrapping, but still doesn't fix problems of getting stuck on the wrong things - you can toss more things in the blender, but you don't necessarily know which way it'll slice them up in advance, or which things from them it'll latch onto. And output still seems to get into "blindered" states where important details get dropped - even though it'll agree very quickly when you point that out. As long as we're in that sort of "spit something out in local targeted manner, and then do a revision loop until tests are green" style of execution, bigger models haven't shown me the ability to really avoid finding non-optimal / subtly-broken outputs for complex problems.
Using Cursor to hop between models, I've found Opus to be generally better at really tricky debugging than GPT 5.5 or earlier models, but not reliably better at execution because of these things. I'm not sure Composer 2.5 is quite there yet for the execution side, but it's getting pretty close to those other ones, such that I'm definitely still in a "debug and plan with slow, execute with faster ones" operating model for working on hard shit.
I'm very happy to have multiple sessions open (and do) and switch between fast and slow models, and if there were a batch mode in codex or Claude code I would use it. (Just like I sometimes use codex fast mode)
But at the moment, I can't imagine why I wouldn't be spending the majority of my time with the best models. I'm spending a lot of time with them! Reducing the number of back-and-forths is extremely valuable to me.
I expect in two months I will still want to spend >80% of my time prompting the best models, and that's true if I were spending my own money on hobby projects, too.
Something that's under appreciated right now is when designing systems and proposing solutions, my colleagues and I do a lot of brainstorming with llms. The core architectures have come out of that, but the best pieces of that architecture are still coming from humans.
These are ideas that simplify the design, reduce future work and tie together the entire system. If in two months I can arrive at ideas of that quality with normal brainstorming with llms that will be extremely valuable
would you be dissatisfied by Opus-4.6-level open-weight
models, just because Opus 4.8 will be out?
Well, I see what you mean, but two big concepts...
1A. Models get stale pretty quickly w.r.t. new developments that occur past their cutoff date. "But you can just keep them current by linking them to never documentation, etc!" Well, no, you sorta can't -- at least not in perpetuity. Those search results fill up your context window real quick. So that gets unsustainable real quick.
1B. Even when your context has plenty of free space, the results you get from "here's a link to the documentation for this new framework that released after your cutoff date" absolutely pales to the results you get from knowledge that is fully baked into the trained model as opposed to your context window. For one thing, that documentation link you pasted into your context might link to... a dozen code examples. Whereas if that was baked into the model itself, the model might have been trained on many thousands of examples in Github etc.
2. It's also a reality that most professional engineers have to keep up with their peers and competitors. We can maybe say it shouldn't be that way, but it is. So if $SOME_NEW_MODEL is significantly better than 4.6... and my peers and or competitors are using it, then yeah I might but really feeling the need to match them. And I'm not even necessarily talking about some kind of cutthroat dog-eat-dog stack-ranked workplace.
These limitations aren't relevant for all use cases or careers but they're hiiiiiiiighly relevant for professional software engineering.
I image that'd be handled via a fairly regular minor bit of additional fine tuning to update them with new information rather than polluting the context space.
For coding don't think so, but they are very close. I code with sonnet mostly because I think opus is just useful if you fail to dissect problems adequately, but anyway.
Kimi is close for example regarding SWE bench for code. For reasoning there are open models that surpass opus by quite a margin already.
The future I see isn't most companies buying hundreds of thousands in hardware to run models, it's them adding a line item to their AWS bill. Inference costs on the larger hosted open source models are dramatically lower than the frontier labs API pricing.
The future I'm seeing is AI coprocessors running inference locally in most devices that today have a CPU. Just look at how powerful your mobile phone has become compared to your desktop computer 15 years ago and compared to a main frame 30 years ago.
The days of requiring a data center to run anything resembling opus 4.6 are already counted. (But the industry will fight hard to get people to keep paying the Claude tax.)
I'm already running a google TPU over USB on an otherwise very cheap board to do local computer vision on a front-door camera since I wanted to get away from Ring and other cloud services for that use case.
And yeah, that may be the ~decade world, but we're in the mainframe era of the frontier models. It's going to be more economical for basically any consumer, and most businesses, to pay someone else to host models for quite a while.
My experience over the past decade has been being subsequently burned by being reliant on one provider's ecosystem after another. This is great until Reolink starts doing something shady to pad the bottom line and then it's on to the next.
I wanted the ability to run whatever cameras on a VLAN and own the stack.
I'm guessing that they are using Fargate which is an OSS NVR. It supports a little addon USB stick you can buy for about $30 that will run common computer vision tasks for object detection. Stuff that we've been able to do with WebAssembly and Canvas for a long time now.
A gaming PC can already host models that perfectly serve casual users who just want recipes, todo tracking, picture identification, etc. E.g. Qwen 3.6 35b which will run on a $650 GPU at 75 t/s (Nvidia 1660 ti 16GB).
Said model will also run as a tool-calling coding model excellently (it's no Opus, but for a thing that once set up is just the cost of energy, it's incredible). It can type faster than you can, probably 10x faster, so with guidance it'll make you faster. And it's free.
It's here. If folks want ChatGPT without a subscription, they can have it today on their computer. The only money to be made is in the high end models doing "serious business" work spanning 1M+ token contexts and massive uncertainty. Everything else is already set to be eaten by today's local models.
The problem with models like Qwen 3.6 35B (which really is an excellent model) is that my expectations of what a model can do have gone SO high now.
Here's a prompt I just ran against Claude Opus 4.7:
> Use python3 to experiment with whether the SQLite3 authorizer mechanism can be used to detect an INSERT OR REPLACE based just on running an explain query without examining the SQL string itself
I tried the same thing against qwen/qwen3.5-35b-a3b running locally in lmstudio, with the Pi coding agent. At first it looked like it was going to do great! And then it fell apart over the course of several tool calls: https://gisthost.github.io/?8ae2f842df619fb7fd8f1ccd82fe41c7
I'm used to GPT-5.5 and Opus 4.7 handling that kind of prompt without any problems at all.
This worked for me with qwen3.6-36b-a3b even at a q4 quant. I ran pi in a docker container and it had to figure out how to install python as well. I used the same initial prompt you had without any additional. You talked about Qwen 3.6, but then said you tried Qwen 3.5 in lmstudio. Not sure if you meant Qwen 3.6. I ran with llama.cpp llama-server with the recommended settings from unsloth.
I'm not an expert in SQLLite so I can't say if this is 100% correct, but it seemed directionally similar to the conclusion from claude.
### TL;DR
- Authorizer + EXPLAIN: No — authorizer only sees SQLITE_INSERT, not VDBE opcodes
- EXPLAIN opcode analysis alone: Yes — Delete opcode at position 10 is the unique signature of INSERT OR REPLACE / REPLACE
I can't help but think the not-so-distant future will see language models expected on commodity personal computing devices.
So one of the prominent LLM advocates known for testing every model shared a prompt intended to exhibit Opus 4.7 capabilities, and Qwen 3.6 sorted it out okay? Interesting.
Not saying they're equivalent, local models still decohere much quicker as the context grows in my experience. But... Interesting.
Something is definitely going wrong with your Qwen setup, in the link you posted it starts and ends with a compaction step due to a 4k token context limit. Qwen 35b supports I think up to 200k+ context limit (though I run only with 128k), that seems to be a major source of the problem.
> a thing that once set up is just the cost of energy
I don't think we can discount this, frankly. Newer electronics are energy efficient, but older devices are more energy-intensive, and unless configured well, a gaming PC can easily use a few dollars a month in electricity, so now you're approaching subscription territory. A subscription comes with no upfront cost, higher reliability, no wasted space in your home, mobile apps, etc. (and less privacy).
Even when run on datacenters, it would be like current day webhosting. It is hyper competitive and it will be a race to the bottom. There is money to be made but not as much as investors hope. There will be datacenters in random countries like Kazakhstan because some oligarchs have found a free energy glitch (like with bitcoin mining).
Magical thinking. I guess if your phone is going to have 128gb of dddr5 then sure. You people fundamentally don't understand the memory requirements for running inference. Your cute local models seem good enough because you have no standards and anything an LLM produces seems like magic to you.
> Magical thinking. I guess if your phone is going to have 128gb of dddr5 then sure.
Why would it not? The typical new phone today has 16gb of RAM. 20 years ago that was somewhere around 32mb. Factor 512. It's not hard to see that we'll get there rather soon, especially if there is an application that provides demand.
> You people fundamentally don't understand the memory requirements for running inference.
You seem to be overlooking how fast things change in this industry, especially if tons o money can be made as a consequence.
> Your cute local models seem good enough because you have no standards and anything an LLM produces seems like magic to you.
Please don't generalize. I'm an expressed AI skeptic and have to deal with the bad consequences of AI slop every day. But you can't deny that there are enough applicationn areas where people have use cases and those will be much easier if things don't need a few round trips to a data center that sucks all the electricity and water out of neighboring communities.
Eh, you're off by an order of magnitude or so on both ends.
The iPhone 17 has like 8 gb, the Pixel 10 12.
The original iPhone was 128mb, and the iPhone 6 from 2016-2018 was around 1gb; that puts the iPhone at around 8x RAM per decade, and puts us at 128gb in our pockets at around 2036 or so.
(Incidentally, the big news in phone RAM is that a lot of new phones are dropping back to 4gb because of RAM shortages.)
Buying "hundreds of thousands in hardware" sounds like a lot but many companies - especially software companies - already do that if they have 100+ employees.
Running software in the cloud gives you certain reliability and scaling advantages that would be very hard to replicate locally. Running some code agents in the cloud vs local hardware, if the local hardware gets "good enough," breaks the other way - offline usage, alone, would be hugely valuable to many people and companies.
It'd be very interesting to see where various players would decide to make a call "local is good enough" though. Buying the hardware isn't a small bet, if it's not something that ends up as part of your standard computer.
Many business tasks do not need the latest frontier models. I have a production system running since early GPT-4o. It now runs with GPT-5.2, not for improvements, but because it is cheaper. I could invest in switching to a local model, I tried and it works well enough, but api costs for this task are so low, it barely scratches $30/month. So I am using the local machine for other things and leave the inference on OpenAI, for now.
I've been doing my work with OpenCode Go, with Kimi2.6. It is not as good as Claude Opus, but it's good enough to get the job done, and I never run out of tokens.
This project argues that with appropriate harness, the performance gap between frontier and much smaller open weight models shrinks dramatically: https://github.com/antoinezambelli/forge. I haven't kicked the tires yet.
Opus 4.6 is a February model. Every time this subject comes up it seems like people post intentionally misleading things and move the goalposts.
The goalpost we've been bludgeoned with over and over again is that, in particular, Everything Changed in November 2025. That GPT 5.2 and Claude 4.5 were the inflection point. That is actually 6 months ago. And DeepSeek 4 is already there.
> run locally
You can't run DeepSeek locally on consumer hardware[1], but you can on enterprise hardware, and enterprise spend is the subject of this conversation -- and even if you aren't self-hosting, it doesn't matter, because you can just get your inference from one of the the many companies serving DeepSeek, who trivially undercut the pricing of OpenAI/Anthropic because they didn't have to spend hundreds of billions on training frontier from scratch but instead only invest in supporting inference, which is already profitable.
[1] Since this misconception comes up all the time, I'll go ahead and pre-empt it: no, training a 32b parameter model on outputs from DeepSeek and running that locally is not "running DeepSeek", despite the hundreds of stupid articles and Youtube videos making that idiotic claim that they're running it on a 5090.
> You can't run DeepSeek locally on consumer hardware
I'd qualify that by writing that you can't run it with ordinary, real-time speed and throughput. If all you care about is slow and high-latency inference, there's no reason why that shouldn't be feasible even on the cheapest miniPC around, as long as it can literally store the model weights and keep around the (rather small) context.
> You can't run DeepSeek locally on consumer hardware
Maybe not DeepSeek v4 Pro, but I've run DeepSeek v4 Flash on my 128GB MacBook Pro using antirez's carefully quantized https://github.com/antirez/ds4 and it's impressive.
Oh sure, yeah, that's nothing to sneeze at either. I think unqualified "DeepSeek" should generally refer to the main model, though, especially in the context of GPT5.2-grade quality.
I keep hearing about this "inflection", but it feels extremely exaggerated to me. And yes, I was using it at the time. It got incrementally better, it wasn't that amazing.
But one will be in few months. And then you have choice of paying say $100k for hardware and pay just power cost (or pay someone to do that for you), or pay way, way more for your team to have access to marginal improvement.
And 5% worse model for 10% of the price of the bleeding edge will be worth it for majority of people
This is totally valid and I don't agree with the downvotes you're getting. Someone coming out with a 10x improvement is possible and would change the game immediately. The thing is, we really have been seeing marginal gains with shifting leaders in who's got the "best" since GPT3, and at least as a user of these tools that pace has been slowing, not accelerating. Subjectively it feels like we're in the back half of an S-curve.
We're 3.5 years into this current AI wave, and a lot of the valuations have been predicated on what you're arguing here -- that essentially should one of the labs make an order-of-magnitude improvement or hit escape velocity on recursive self-improvement they'd become the most powerful economic chokepoint in history.
The reality has been that given access to compute + capital all of the labs can stay pretty competitive with each other. Someone does a bit better on coding, someone else does a bit better on tool calling, and then they swap after each spending another $100bn.
The market looks like a commodity market where the commodity is intelligence, not a winner-take-all market with massive margins. Plenty of people get rich in oil and airlines, but they notably don't tend to be the innovators long term, they tend to be the operators. Obviously if the machines become sentient tomorrow, turn on their masters, and hit world-dominating intelligence, that assessment changes, but after several years of that narrative while objective reality looks quite different I think the more sober voices are starting to gain a foothold.
I agree with most of what you're saying, but I think the point I was trying to make wasn't as high-flying as you and others understood it.
I'd pay a premium for even just a model that's 20% better, no ASI required, and I think a lot of people would. I wouldn't call that marginal, if it means I'm getting frustrated on 20% fewer tasks.
A recurring pattern that I've seen in myself and others is to at first be very impressed by a new model's coding capabilities, and then desensitize quickly and start being frustrated by the shortcomings.
> I'd pay a premium for even just a model that's 20% better
The point I'm making is that I think we're rapidly hitting levels where corporate buyers aren't willing to pay multiple-times-more for marginal gains, and I expect that to become more the case over time, not less. You, and a small % of other power users in the market might tolerate a $400/month pro-supreme-plan for access to Mythos or whatever, but I don't think that's going to scale up in quite the same ways we've seen so far.
Even a year ago paying multiples times more for a 50% gain was very sensible for a lot of workflows. But if we're getting to "good enough" for things like coding, justifying to your CTO/CFO why the org should go from spending $1m/year to $5m/year for a 10% higher hit-rate on one-shot prompts from the engineers is a much tougher sell.
The leap from GPT-4 to GPT-5.5 has been astounding in my opinion. There is no way GPT-4 could run a coding agent harness like Codex at even a fraction of the quality that GPT-5.5 does.
I don’t think that’s exactly indicative of GPT-5.5 being an astoundingly more intelligent model, however. An alternate interpretation is that GPT-5.5 was trained on tool usage/harness patterns and has been optimized for this use case.
I remember that even when GPT-4 was king, the Gorilla paper showed that Llama 7B could be fine-tuned to outperform GPT-4 on tool calling.
On domains that don’t involve agentic tool calling*, I haven’t found the frontier to have advanced that much.
Edit: I should broaden this to domains that naturally lend themselves to RLVR training. Models are drastically better at math now.
None of this matters in the product: it either is capable of agentic loop workflows or it isn’t. A 10% improvement in probability of single task success makes or breaks the use case.
Open source models, especially qwen are pretty dang good. But its not opus 4.6, the evals dont tell the full story. I question the assumption open source models are 3-6 months out.
Its not just about the quality of output, but you also can finetune them to proprietary needs, if the skillsets are their internally, to make them better without governance risks. So being SOTA doesn't matter as much, since generalized tasks are not what matter most to companies, its the specialization relative to business need or internal datasets.
The larger point I'm making is I think models are rapidly becoming commoditized. There is probably a small market long term that's willing to pay 10x for 10% marginal gains, but the majority of the buyers in the market will be economic and we're likely to have a lot of folks willing to spend 1/10 the cost for 90% of the performance, and plenty of companies that haven't raised hundreds of billions-trillions who can provide that.
A lot of the frontier labs valuations has been based on an assumption that 1-2 companies would get break-away intelligence that basically made them economic chokepoints indefinitely into the future. The reality that's becoming increasingly clear is that model quality is a pretty linear function of (cash burned - ability to copy other's homework) and the economics are starting to look a lot more like airlines than online advertising.
You have to think about why open models are behind. Exfiltration is a big part of it. So you could change the Nash equilibrium by increasing your security, or other multilateral approaches.
Better now than ZIRP for me - at least people are asking timid questions about the unit economics and how long the runway is _early_ while also spending absolutely insane amounts of money on this bet. During ZIRP, these companies would have turned down any investor asking questions. Less contagion when rates aren't zero hopefully? :grimace:
> ...we are already looking at dropping $100k on hardware to run local models...
Just think how much further that $100K would have gone if the hardware market wasn't so screwed-up.
Anecdote: I priced-out adding 1TB of RAM to a four node cluster a couple months ago. The cluster was purchased in fall of 2024 w/ 4 nodes, each with 256GB RAM. The nodes cost just over $14K apiece back in 2024 (entire box, not just the RAM).
Dell wanted >$90K a couple months ago to add 256GB to each node.
> Dell wanted >$90K a couple months ago to add 256GB to each node.
RAM is expensive, but not THAT expensive. I just bought 128Gb for about $5k for our build cluster (it's not even for AI, sigh). Even if you need larger-sized DIMM sticks, it's still going to be in the vicinity of ~15k tops.
It was crazy. I found the part on the open market for a lot less but the edict from the Customer was to buy from Dell to keep the support entitlement intact. That inflated the price to an astronomical level to be sure.
I haven't had problems w/ Dell support and 3rd party memory, personally, but given the machines' application I understood the concern.
Yes 100% this. A lot of people keep talking about how OpenAI and Anthropic will need to raise their prices. What is less discussed is how they CAN'T raise their prices because competition exists, and sure it's not SOTA, but it's literally an order of magnitude cheaper in many cases and the drive to figure out how to make it work well enough is going on right now (and will only intensify when the SOTA models raise their price).
It's a given that the SOTA models need to raise their prices. It's also a given that they can't. The more they raise the more customers will move to their competition.
So what happens next? Well I think it will suck horribly if you can't move off of SOTA sooner or later, because the Big Two are going to lose customers, and therefore have to raise prices on the locked in customers even more than these projections suggest.
Beyond that if you're looking to start a business, figure out how to use cheap models in new scenarios. Build software which does that and license it. This is kind of contrary to the idea that you shouldn't over optimize for deficiencies in the models that will likely go away in the next generation - for instance a lot of problems were solved when context windows got way bigger. So it's a thin line to walk but I think it's there because a lot of orgs are using Claude today for pretty basic tasks.
The dev who's addicted to SOTA models honestly is going to have to settle for less or get totally screwed. Most applications within business from what I see aside from complex research do not require SOTA. They summarize, they classify, they transform, and doing that accurately has been cheap for a while.
I get the impression the hive mind hasn't come to terms with the point that a model is optimised for certain tasks. It's like having someone ask you "is that a good hammer?". Good for what? There are claw hammers, sledgehammers, ball-peen hammers, club hammers, mallets, .... Yes, in a pinch, they can all bang in nails, but you wouldn't choose a dead blow hammer for that if you had a choice.
The Gemini Flash is very good at searches. Just about any low end model can toss out a poem. All the higher end models (open source and otherwise) seem to be able to churn out code that passes tests. The smaller, "less capable" ones are much faster at it, which means in the hands of a skilled practitioner are the best choice for that task. But they rapidly fall apart where there isn't a hard source of truth (like a good test suite) to grind against. Because of that you have to use a bigger model for bug finding. In that task the open source models tend to fail on larger code bases, where something like Opus still shines. I gather Mythos is an absolute monster, and unparalleled, and unavailable. I'm sure one of the reasons for that is it's so expensive to run.
Or to put it another way - you don't use a 100 tonne crane to pick up the shopping. And ... the smaller models will happily run on in-house hardware. You may not do it today because of the current DRAM price and integrated NPUs have just started shipping, but in 5 years time models will be running on your phone.
On prem AI makes sense for more than just the cost. More control, IP, model improvements you can keep, data privacy to name a few. People will realize that AI is not like compute the moment they get their own knowledge sold back at a premium.
What are the advantages to on-prem for a company that's already in the cloud and trusts it with their IP? That company can just rent GPU instances from the cloud if they want to train/fine-tune their own models and keep avoiding CapEx.
Agree. You have these tipping points when a model is good enough to do some task. Yes, a better model will further improve your capabilities but the unlock is at a certain intelligence level.
We see this also with humans. People with very low intelligence can't learn to read. Once you cross a certain threshold of intelligence you can learn to read. More intelligence doesn't really help you in the task of reading. A person with an IQ of 160 is not substantially better in reading than someone with an IQ of 85. If your IQ is 50, you might not be able to learn to read at all.
Depends on the task and the writing though doesn't it?
There's not that much depth in a lot of 'everyday' writing. For many tasks that means that you don't need to be hyperintelligent - reading a recipe or a shopping list, reading a newspaper article, etc.
By my calculations 100k could get you 18 5090's + compute to host them, or 18 96gb Mac mini's. You can get a lot of context window and users out of that setup.
I doubt it. Companies that have moved to the cloud are already trusting the cloud with their IP. You can rent time on a high end Nvidia system from various clouds. OpEx means there's no write down in three/five years as that system goes out of date so it would only make sense if the performance/$ is there, or the company is highly protective of their IP and doesn't trust the cloud, at which point they're not on the cloud anyway.
DeepSeek v4 Flash, various quantised versions of Kimi K2.6, MiniMax 2.7, Qwen 3.5 “full sized, with a dual spark setup you can fit some decent setups on here
My single spark has me running Qwen 3.6 27B and antirez’s specially quantised DeepSeek v4 Flash (which is shockingly impressive)
Have you tried it? It would be slow for sure, but the main limitation AIUI would actually be storing the context in RAM - models like Kimi and GLM have high demands there which limit your ability to get meaningful aggregate throughput via large batches.
My much larger company has got people already using various models through Bedrock because the Claude and OpenAI limits are too harsh and it's too expensive.
It might be possible that in a few years someone will be able to engineer a reasonably priced machine to run today's frontier models (hint, your price is an order of magnitude off). However, they won't be able to run the frontier models that will exist in a few years.
I’m curious: are you spending on beefy developer machines, or some kind of shared local inference server? Would be interested to know more if it’s the latter.
.> Hardware costs haven't exactly been cratering recently.
No, but local models have been booming in performance/quality improvements. The RAM shortage won't last forever (more supply will come online when if demand doesn't diminish), and then the math would be pretty easy.
same, but you need more then 100k of hw to run something like kimi k2.6 for a bigger team. on the other hand there is a ds4 flash that you can run on a macbook with 128gb ram. an that one is perfectly usable for a lot of tasks.
That’s exactly where the market is heading and it’s going to have to reckon with this fact
My guess is there’s gonna be some legislation or something “you can’t share anything over this level of complexity” and I think that that’s what a lot of that mythos rattling was all about
> there will be hardware capable of running frontier models
The current frontier? Sure. The frontier then? No - obviously that frontier is going to keep consuming available datacenter compute capacity, which will be better
You people are delusional. How many times a day am I going to read this fiction of "good enough in a few years for most things".
There are physical limits to how much you can compress data and how much is needed for a capable model. If by hardware capable for running SOTA you mean a 7 figure investment for a company, than sure. But how come these companies didnt do the same thing for cloud? There's been this option for self hosting infrastructure for a decade but companies don't use it, they pay AWS.
> In a few years, there will be hardware capable of running frontier models good enough for most things at accessible prices for even tiny companies.
I was going to say - the models are just going to keep growing at a pace exceeding the pace of hardware pricing/availability
But then I realised that, far more likely, there will be a plateau reached (again) where nobody is seeing gain, and at that point hardware will catch up
I was in college in the late 1990s/early 2000s and I distinctly remember an econometrics professor state the following:
"As cable TV and Pay Per View came out, there were studies done about how many movies people would watch if given unlimited access to films. The results were bandied about as proof that we should build out all this infrastructure to support this line of business. When the data was further analyzed by statisticians etc, it turned out that people claimed they were going to watch films 10-12 hours a day, every day of the week. Impossible."
I feel like we are in a similar boat here where some people are assuming:
- EVERYONE is going to be using max tokens
- tokens will NEVER get cheaper due to improvements in hardware, software, design, market forces etc etc
>I feel like we are in a similar boat here where some people are assuming:
>- EVERYONE is going to be using max tokens
>- tokens will NEVER get cheaper due to improvements in hardware, software, design, market forces etc etc
I feel like the reverse assumption is being made, that the current model looks like IBM doubling down on Mainframes soon to become cheap enough to deploy everywhere, when the real action is that the costs coming down represents cheaper hardware or more efficient software, and that a big chunk of "cheaper" AI will be eaten by smaller products deployed by individuals. Whatever the Personal Computer of AI looks like is going to be more disruptive than just an API endpoint you can fling tokens at.
We already see this with things like chrome auto installing an LLM.
You cant tell me with complete certainty that theres a moat here for the people spending 1 trillion + on this infra.
>When the data was further analyzed by statisticians etc, it turned out that people claimed they were going to watch films 10-12 hours a day, every day of the week. Impossible.
I also think this applies to people suggesting that companies will sack workers for AI, when the costs of replacing everything someone does in a day is more expensive in terms of tokens (likely even at a reduced price) than just hiring a bloke.
> it turned out that people claimed they were going to watch films 10-12 hours a day, every day of the week. Impossible."
I realized it long ago: one needs output to make meaning. Input can only be the cherry on a cake in one's life. That, actually, makes FIRE or Fat FIRE not so sustainable unless one has other hobbies.
What people: I'm sure some people are watching 10-12 hours per day - in places like nursing homes or hospitals. I know a reasonable number of people who watch a film nearly every day: 2-3 hours. Most people watch something every few days - often a tuesday night movie night for the family (or something like that). There are some who never watch anything. I don't know what the statistics are on this.
My friends in day care tell me the kids hate "movie day" because movies are all the get at home and they are sick of them - they want to play all day. (but I'm not sure if this is representative of anything other than the types of people who put their kids in that particular daycare)
> they were going to watch films 10-12 hours a day, every day of the week. Impossible.
A lot of these LLM demand scaling scenarios make broad "up and to the right" assumptions about things which in practice have finite limits. Only some percentage of knowledge work benefits from acceleration, optimization or other improvements, and even then the amount of economic gain is capped.
It's vaguely disturbing that people "watch" films 10-12 hours a day. Many of them are using it as a radio, for background noise, without really caring what the program is beyond vague genre, tuning in and out without particular regard to the plot… and yet we have all the cost of transmitting high-resolution video point-to-point.
Surely we could just put better stuff on the radio, and accomplish most of the same goals for a far lower price?
My Dad was in the hospital, and just wanted to watch the Pirates play. The TV was filled with apps, some of them free to watch, others demanding a subscription and log in once you selected something.
None of them had the Pirates game.
I was thinking how the transistor radio was a far superior experience for this use case. Just tune to the channel broadcasting the game.
Radio has not gone anywhere you know? There is of course podcasts, but for instance Radio France has amazing music services like FIP: https://www.radiofrance.fr/fip
Then there’s NTS, BBC… Ypu can listen to them from online service, but at least in Europe there’s amazing national FM broadcastimg services.
anthropic already hunts down OpenClaw users for using too much on their plan.
I'll give different example: When LED lights started to be more popular, the power usage didn't drop by the amount of power saved
>- tokens will NEVER get cheaper due to improvements in hardware, software, design, market forces etc etc
Well, first, improvements in computing stalled or even rolled back just purely because price of everything compute shot up cos of AI and that will NOT be fixed for a while and ESPECIALLY if AI usage will continue to increase
Second, the token per model might go down in time but better models have more expensive tokens, so we quickly get into spot when:
* price increase in token might not be worth marginal improvement next, better model brings
* more and more models are passing "good enough for the task" threshold so for less and less companies there is any economic sense to pay for the "best" instead of paying deepseek or some other company to run "previous gen" models
I don't think the maths works like that. They have raised ~$200bn so far and need to make that back. Saying they need to make $5 to $10tn isn't really real. They might need that to meet some extravagant Altman projections but not to justify what they have actually spent.
Yeah, claiming “product-market fit” on coding assistants for this multi-trillion dollar capital expenditure seems premature. Anthropic will post one and only one quarter of “operating profit” (aka losses after taxes and debt obligations) on the back of free-for-all spending by enterprise and engineer tokenmaxxing, neither of which will last. The investment was commensurate to a world-eating AGI, and if all that comes out of it is coding agents and slightly better enterprise software, I don’t think that makes up for the money spent.
The real question is: can you incentivize a non-tokenmaxxing Uber to spend the same amount on AI as they were when tokenmaxxing, just with fewer tokens and higher per-token costs? Even with plateauing improvement in frontier models? I think the answer may be yes.
And part of my reasoning for this is: the only system capable of actually fixing bugs in vibe-created code is an LLM. If we humans couldn't write it without assistance, we certainly won't be able to debug it without assistance. So there's a real stickiness here.
We're signing pacts with demons - we have to, if we want to outcompete the other warlocks - and those pacts are written in the very size of our codebases.
The bottleneck has moved from producing a thing that works to knowing that the thing was the right thing to build. The more of the latter they can take on, the fewer knowledge workers are needed at all. So rather than 5% of every knowledge worker's salary going into tokens, 100% of the knowledge worker's total employment cost goes into tokens and you get a 20x productivity boost as a theoretical minimum across those tasks.
That's the game. There's a view you could take of this that this is just a growing of the pie: with those cost dynamics a lot more "small businesses" get a vast amount of leverage, so the overall economy grows without replacing the knowledge workers. I'm not sure I trust the MBA class to have that view.
>The bottleneck has moved from producing a thing that works to knowing that the thing was the right thing to build
I would argue that that's been the case for quite some time before AI. As an example, what innovative amazing world-changing products have Google or Meta launched in the past decade with their very high numbers of very talented and highly-compensated engineers? The issue with most big tech companies are leadership, strategy, and product direction. I'm not saying that they don't make any profits, just that they probably aren't "building [the right thing]".
AI for product development and management would be far more impactful than automating rote coding tasks / building React UIs that mirror API structures IMO.
> AI for product development and management would be far more impactful than automating rote coding tasks [...]
Yeah, if this stuff actually worked that well already, OpenAI et al. would just run AI CEOs and engineers. Why get some other company to pay you at all when you can automate every other company out of existence and take all the money they make?
The fact of the matter is that while the tech has some uses, it sure as hell isn't a full scale replacement and you almost always actually have to massage the input into LLMs to get anything decent back out in practice. Some CEOs and managers can learn to do this, of course, and some already are... but that quickly turns into a second full time job. A "programmer" is still needed. The job might change from mostly hand-writing C++/JS/Python to prompt engineering + some manual coding to fix all the stupid fuck-ups that the bots can't solve themselves, but you still need someone to actually prompt the bot.
When that changes, it won't just be engineers losing work; there will be no reason to even have a human CEO any more.
> When that changes, it won't just be engineers losing work; there will be no reason to even have a human CEO any more.
The human race isn’t ready for that world IMHO. The only reason there is a middle class is because people have leverage in the form of their labor. When that becomes worthless … the people who own stuff and make their living from doing so won’t hesitate to get rid of everyone else - whom are now worthless to them.
I don't know, if you've ever tried to build something at companies of that scale you run into incredibly boring problems "what data table do I need for X" and "who is the right person to reach out to for Y" and "they aren't answering me I guess I'll have to escalate"
I don't think there is any shortage of great ideas at these companies, they are just extremely bloated. And I don't think its something like indecision or bad PMs, it's "we have a finite amount of time and resources so we need to be conservative but also not too conservative"
If you have AI systems that can simply build out POCs in days, backtest on real data, show reliable results and numbers, you get a suite of product options you were never able to get before. If you have coding agents that can speed up implementation, you can build more stuff and choose the things that stick.
It changes the cost/benefit calculus of the entire business. I think you are exactly right in that: PMs/leadership are by their nature orchestration machines. Other roles are as well, but I think PM's are at a particular advantage here in that it will be quite awhile I would expect before core product decisions and creativity can be delegated to an AI, but not quite awhile until virtually everything that they're blocked on (legal approvals, POCs, wire frames, etc etc etc) will become less and less of a blocker
>If you have AI systems that can simply build out POCs in days, backtest on real data, show reliable results and numbers, you get a suite of product options you were never able to get before. If you have coding agents that can speed up implementation, you can build more stuff and choose the things that stick.
I'll also add this: within a large organization, you often need to interact with many different codebases owned by many different teams. Agents have made it much easier to wrangle by having the ability to deploy one to scope out your web of dependencies to learn about what would be needed for feature X, and how that integration can happen.
We've been doing far more away team work simply because it makes things move faster. It's easier to convince a team to sign off/review something than it is to get them to commit to the planning and eventual work.
It genuinely is helping things move faster inside large organizations. Or at least, it is for us, particularly since we're getting organizational prioritization to actually build the scaffolding to make those agents more effective at search.
> It's easier to convince a team to sign off/review something than it is to get them to commit to the planning and eventual work.
1000x yes: you have touched on what I think is the single biggest factor here, that is the humongous value of POCs. they are gnarly to build without agents, and so we used to have to get everyone on board so we didn't get screwed in performance reviews, which was monumental task because that means convincing very busy PMs who have a lot on their plate and dont want to take risks on things they don't understand, and now it's like "can we scale this out" and you have a very nicely formatted proposal and POC. It de-risks things very quickly
Absolutely but you want to package it to them nicely and efficiently. The biggest blocker is legal and everyone else speak two completely different languages and we often don’t know what’s important to flag and legal doesn’t know enough to ask all the right questions. Also, many things can be templated, and in an industry where regulations and precedents change so quickly, agents are at the very least a good tool to flag issues (e.g. we were approved to use data X for Y but now decision Z negates this). The propagation of this information is not very effective now and legal review at tech companies, while absolutely essential, is somehow a worse experience than going to the DMV when it’s crowded.
Pieces of concept and other prototypes have always been cheap (see hackatons). The main issue is that as soon as you’re touching customer data or modifying process they’ve paid you for, you have to be really careful. No one wants to be responsible for an outage that cost the company its biggest customer.
Yes, but at scaled companies, where building a simple POC hooked into real systems is nowhere close to easy. To the point where it means that you might as well just do the full thing. That's where the enterprise spend and the impact is.
historically it's been a matter of an absolutely horrific amount of Kafka-esque problems.
Say I want to build a feature in a product.
- DS has to do a deep dive (need buy in) to opportunity size and derisk with data. That DS has to work with other DS (people may have left or moved teams) to figure out how to get the right data and figure out what the difference is between 10 different tables that have overlapping but inconsistent data.
- Eng has to build up an actual simple demo (need buy in)
- Design has to make it not hideous (need buy in)
- Legal has to review what you're doing; POCs should involve real data where possible because otherwise no one will trust it, even if its just for user analysis on existing products
This plus about 6 internal system bugs for custom tools that are flaky and who's team has long been re-orged or laid off, 8 people who won't answer you, 2 PTO's for the stakeholders, 6 weekly meetings
no one did POCs, they just had ideas and tried to get PM's to put it on the roadmap so if it fell through at least it was bought into
Yes, that exists at the wider business level. No question. I think what needs to get asked is are we talking about a bottleneck within the business as a whole, or a bottleneck within the scope of the knowledge work in question. Within software delivery there's a very clear shift when it's suddenly trivial to drop a 100kLoC plausible-looking PR into code review within an afternoon. Producing working code with a whole bunch of tests which make a very clear assertion that it does, in fact, work has had (if you're going that way) all the human-scale thinking time taken out of it, down to a rounding error. It still needs to be checked by a human, which was previously assumed to be a comparatively quick task in comparison to producing the thing. At least, it does where I am, and I don't think that's a silly position today at all.
If they can crack that latter review/spec-check/assurance step, checking that what was built was what was demanded of the problem such that we don't have humans in the loop at that step either, then the bottleneck moves again. Then I think it moves to requirements capture and to product development, but that might depend on the industry.
Numbers I see put Pixel at less than 5% of iPhone sales. Not nothing, but world-changing? I doubt the world would look significantly different had Google not done Pixel.
If you really think this you simply have no theory of mind for this stuff. There are tons of immensely successful products in the ad space that both of those companies have launched. They don't need to innovate in the product or technology space (doing so certainly makes a big difference in having more placement for ad real estate), but to suggest there have been no real innovations (specifically engineering specific innovations) related to ad tech would be completely ridiculous to suggest. You don't need to change the world to get rich, just look at wall street where major innovations have been made in the pricing models of fixed income securities.
Second to this are countless other areas that have a major impact on the companies bottom line that are entirely engineering driven, especially at google given they are a cloud provider and have meaningfully grown the workspace business and launched waymo in this time.
Google's internally developed and sometimes even launched plenty of innovative new products in the past decade. Stadia, Fuchsia, federated learning, and the whole transformer architecture that underlies this AI boom are good examples.
The problem is they get killed by some other executive who is afraid of their department looking bad by comparison.
I think this is fairly illustrative of the challenges in AI becoming as impactful as the Internet. The bottleneck is not making things. There are plenty of people who are really good at making things and can easily be 10x or 100x as productive as the average corporate worker. YCombinator was founded on that premise - small teams of founders and early employees could be orders of magnitudes more productive than the 1000s of corporate employees at their competitors.
The bottleneck is on bringing your product to market. If your innovative new product is built within a corporate environment, it'll get killed unless the executive you work under can get a promotion out of it, and you'll be denied all sorts of help with approvals, launch process, PR, marketing, branding, etc. If it's a startup, they'll try to shut you out with exclusive distribution deals, legal threats, lobbying efforts to change the legal environment, PR campaigns, FUD, etc.
The Internet was revolutionary because it let millions of people bring products to market without asking permission. Instead of having to bid for retail shelf space among dozens of entrenched competitors that all had sweetheart deals with the retailer, you could just put up a website and sell it to anyone across the globe. Instead of following hundreds of regulations that governed existing commerce, you could just launch something and sort it out later. AI doesn't really have that property - if anything, it makes things more centralized, with more gatekeepers, and so seems more likely to destroy economic value than add to it.
Google does not follow through in the long run on many of their pet me-too follow projects, however they do not stray away from their core remit making their real customers happy the ones who buy the ads…
Obviously that includes whatever needs to be done to hoover in data from their marks and Meta also does the same thing without fail and both are really good at it. But outside their remit not so much.
What I think is happening is that the scale of thing you can hope to build at a below-corporate scale should radically grow. Corporate environments should suffer for this, being that inefficient.
> YCombinator was founded on that premise - small teams of founders and early employees could be orders of magnitudes more productive than the 1000s of corporate employees at their competitors.
I think this is still true, but the theory is:
1. You don't need YC-type funding to do YC-type business any more;
2. You don't need to scale the business past those small teams any more, you just buy more tokens.
For clarity YC still obviously has a place as an incubator, mentoring, and networking function. I just think that what was previously the inevitable conclusion that you have to hire all the people the second you hit PMF to keep up with scaling the business as you scale sales is no longer inevitable. If you didn't want to go that way before AI, you were a "lifestyle business" and not worth investing in. As more and more knowledge functions get capably implemented by AI, it's the preferred position: humans are vastly more expensive than tokens, so you want them doing the stuff the AI still can't do.
I don't think this necessarily translates to mass unemployment. I think it translates to masses of smaller businesses that are radically more efficient because the handoffs between business functions are tool calls, not emails to someone who doesn't want to help.
> The Internet was revolutionary because it let millions of people bring products to market without asking permission.
Think about it this way: if I am a small business owner but I think it makes sense to do something that previously only a team in a corporate environment could do but is now within the reach of AI, not only can I do it now, but I also don't have to ask anyone for permission! Who wins between the corporation and the small business in that scenario?
> AI doesn't really have that property - if anything, it makes things more centralized, with more gatekeepers, and so seems more likely to destroy economic value than add to it.
I think this will turn out to be backwards. I can see a version of this where the number of things you can do without needing to turn to a gatekeeper for help increases to the extent that the balance completely inverts.
The vast majority of businesses are small, and AI can give them tools which previously required corporate scale to make sense, without the inefficient hand-offs between busy, political humans. Which is also something that the internet did! Getting an advert in front of a national market pre-internet was Hard but sometimes you had to do it because your target market was "all Canadians who buy toothpaste" or whatever and that meant saturation-bombing the physical environment with physical billboard ads, posters, flyers, and so on. So you only did it if you were P&G-scale. Now you, personally, can do it, trivially, for better or worse.
I dunno if the employees were ever really needed for scale. WhatsApp famously had 300M users and 13 employees at the time of acquisition; Instagram was something like 50M users and 55 employees. If you know what you're doing software scales basically infinitely, and the employees are there to make the software just slightly more tailored to specific user populations (and because upward career mobility for managers involves having more headcount). Yeah, building a revenue model takes people, but Valve employs only about 400 people and makes billions, as do various quant hedge funds like DE Shaw or RenTech.
The insta/whatsapp/plentyoffish model works if you're very lucky with both product-market fit and the technical constraints of the product itself. If you have something that technically scales extremely cleanly, it basically sells itself, and it doesn't need feature churn to retain or gain users, you're golden. I do think more businesses could do with checking whether they do in fact have that lottery ticket before hitting the scale button; there aren't that many examples around.
> Valve
Arguably a monopoly. They've got a product that sells itself with very low infra overheads for the income.
> Hedge funds
Very different model. I don't think the same intuitions apply.
Most of the people they have on staff are there to support their real customers, and I don’t mean the marks out in Internet land Google and Meta‘s real customers are the people placing ads and giving them money, most of the staff is dedicated towards servicing them, that again is where most of the money goes to support their real customers.
Google & Meta are illustrative of late-stage capitalism -- it's all about distribution, not innovation. Their job is (mostly) to just acquire the products that have passed the gauntlet, then scale up their monetization through their distribution-focused machine. The same dynamic plays out in virtually every industry (not just tech).
You'll find that most internal "innovation" teams are just lip service. In most cases, the "mothership" will be incapable of reproducing true innovation -- from a statistical perspective, culture perspective (mega corps are anti-scrappy; internal politics), and motivation perspective (startups aren't 9-to-5). It's much easier to have big M&A budgets, a VC arm, and some handwavvy internal innovation group.
Every now and again, you'll get real innovations (Waymo, transistors, GUIs), but even those have a spotty track record of commercialization when created internally.
This is the same argument that has been historically made for outsourcing developers. Get 20 more devs for the cost of 1 dev in the US.
I suspect that AI will fail to pan out to the same extent for the same reason why outsourcing hasn't fully panned out (even though every company tries it after getting big enough).
The problems that will come up will be and always have been ongoing maintenance. AI is great at writing new code without a brain behind it, but once you get to the point where you need to refactor code, you start really needing someone with coding experience to guide the AI or veto it's mistakes.
I don't think that's really fixable even with a lot better AI. It's not something that ultimately comes out of the likes of github data.
I'm not saying that AI isn't going to make things better, btw, I just don't think we'll see a 20x improvement. Probably more like 1.5 or 2x.
Outsourcing of knowledge workers didn't work out because at large enough scales, the geographic arbitrage disappeared. Companies mostly always got what they paid for.
The determinant of success was only whether the task needed American-tier labor or could make do with sub-American quality labor.
I am not sure this feels right. I agree that the US currently has leading minds in terms of tech, but I am not sure it is too big of difference with the EU knowledge workers. EU is still a lot cheaper then US in terms of wages you would need to pay.
EU workers themselves get a lot less, but the EU is expensive because of 1) the huge payroll tax (45% in France) and 2) the challenges with hiring and firing mean you are carrying people that aren’t contributing.
Sure is an interesting thought. None of this is sarcasm: why do US companies deal with the time zone differences and language barriers they won’t need to bother with so much by outsourcing to say, Ireland?
The mechanism is often that they'll actually outsource to someone like Accenture, who have teams everywhere, and whose contract managers will try to get their cheapest viable team onto the contract to maximise their margin. If the buyer can't judge the quality of what they're buying, or doesn't know why the resulting hand-offs, delays, mistakes and rework will cost them more than keeping everything in-house ever would have, they're going to have a bad time.
Ireland isn't that much cheaper than, say, Oklahoma. And the cultural differences with Ireland are not a lot smaller than those with India or the Philippines or what have you, once you try to actually start working together.
(Yes, all the good developers from Oklahoma move out, but the same is true of Ireland)
Basically every big tech has large offices and employ a lot of people there.
The limitation is that Ireland is a relatively small country, and most Irish developers are already employed (which is why Ireland end up being one of the main destinations for tech workers being hired from abroad).
That's certainly part of it. But the other part that I've heard time and time again is that in order for outsourcing to be successful you basically needed an american engineer in the mix hand holding everything, clarifying requirements, and vetoing bad code.
That part of dev work, the requirements gathering, attention to details, clarifying requirements, is something AI also struggles with. A lot of companies basically waste time and money on outsourced devs because without a clear path forward they effectively will sit and do nothing, waiting for a prompt.
I would not agree on that point. It really depends on company's structure. I mean it also depends with people that makes the team. I would say there are a lot of unknowns but I would certainly not generalize.
How I find your argument is that one distinguished engineer from US could do the same with the use of AI.
I worked with both and I know great and bad engineers from both sides. Only thing is that US has a bigger pool of great engineers.
I think the mechanism here isn't that American engineers are magic. It's that you need that contextual knowledge really close to where the work is actually being done, so that the turnaround for questions, blockages, clarifications, "we've got no work to do", quality level-setting and so on is on the scale of minutes, not time-zones.
It doesn’t have to be an American but it does have to be a direct employee of the company ideally working in the same time zone as management and the people defining the requirements.
Outsourcing of knowledge workers is still ramping up. The issue in the past was the skills were few and far between internationally. Facilities were also not built. That has changed now in a lot of fields. New campuses have been built in places like Bangalore and Hyderabad, even Singapore. The skills are there now, the training is decent, and you can see that the hiring is going on for very skilled titles in these cities. It is a different animal than just 10 years ago in this.
The “American tier” labor of course is distributed across the world and the top performers in every nation find ways to get paid at something approaching American salary levels. Look at all the international FAANG offices paying high salaries, in purchase pricing parity terms.
> I suspect that AI will fail to pan out to the same extent for the same reason why outsourcing hasn't fully panned out
My mental model for that is that outsourcing fails where the work is being done organisationally far from the knowledge needed to do it. We know that's true of teams inside organisations, there's been a lot of research on how distance in the organisational tree negatively impacts productivity. Outsourcing is a pathological worst-case of that.
The promise (promise! We're not there yet!) of AI is that I can have a cross-functional team on my laptop. Organisational distance is zero. Where previously the outsourced team has to wait for the time zones to roll round so I can answer their blocking question when I get to my email STRICTLY AFTER I have had my coffee, now it's a prompt in a chat window with a button I can click to make a choice in 5 seconds. Delay is gone, cost of delay is gone.
> The problems that will come up will be and always have been ongoing maintenance. AI is great at writing new code without a brain behind it, but once you get to the point where you need to refactor code, you start really needing someone with coding experience to guide the AI or veto it's mistakes.
Oh, absolutely. That's a minefield. Today. It will be, right up until it isn't. There are ways to set up agents and projects right now that make a dramatic difference to how this part of the picture plays out, but those will sink into the harnesses as time goes on.
But also the big problem with maintenance and outsourced teams tends to be the commercial structure around the contract. You get a Build team, who Build the Thing and then: no more features for you, anything you want to add past the original spec costs extra. They hand over to the Run And Maintain team, who get to fix all the bugs that the Build team left but without the knowledge gained from building the thing, but are scaled and located to be absolutely as cheap as the supplier can get away with so probably don't have the skill, inclination, motivation, or permission to take on any restructuring to make the bug fixing easier and they're on the wrong end of the globe so there's a 24-hour latency on any queries. It's a terrible way to set teams up, but it looks good on paper.
Again, that's peculiar to outsourcing and completely goes away if I have the same team that built the thing own the thing long-term. That's true if it's humans or AI!
> I don't think that's really fixable even with a lot better AI. It's not something that ultimately comes out of the likes of github data.
No, it's a harness problem. You need to start from a maintainable point and keep standards in place. It'll take work to get the harnesses there and it's not ubiquitous. You might also need better models, but I've already personally seen big differences in outcomes between projects that took certain steps and others that didn't; it's nothing revolutionary, mostly stuff that works for humans also works for AIs but you need to know to ask for it.
> I'm not saying that AI isn't going to make things better, btw, I just don't think we'll see a 20x improvement. Probably more like 1.5 or 2x.
I think people radically underestimate the cost of delay. I don't know if 20x is realistic for the AI itself, but I think it's not impossible once the inefficiencies of having to go to other humans is factored in.
> mostly stuff that works for humans also works for AIs but you need to know to ask for it
I'm most curious about this sentence. What have you noticed about the similarities? I'm getting really good at asking for confidence levels, tests and pushing back, but I'm curious what you found
Outsourcing also fails because it’s a pathological case of adverse selection. The businesses that outsource projects are ones who are organisationally incapable of managing those projects well internally. But, that inability extends to their oversight of outsourcing shops as well.
End result is that many outsourcing firms are borderline fraudulent in the way they treat their customers.
Yes, but I do worry about junior knowledge worker job loss. These models are very good (and getting better) at the vast dark matter of "donkey work" that happens in knowledge-based industries -- work typically done by junior devs / analysts / lawyers / consultants, paralegals, admin assistants, customer success / support, etc. -- and those roles comprise the bulk of the workforce.
And worse, these are the tasks that help the junior people eventually grow into the skilled knowledge workers required to operate models, so there's a pipeline problem too.
I do too, but I think it currently has a lot more to do with the quasi-recession we've been in since the end of ZIRP and AI is a better excuse to stop training juniors than telling investors it's belt tightening, just like layoffs.
I'm already seeing tech execs/hiring managers getting very frustrated at the lack of new-senior-engineers to hire. The market will correct for this in time.
Curious if you can share any backing information from your last statement? As a senior engineer (well, that's my job title anyway), I find it encouraging.
This doesn't break it down by experience, and I can't find specific data on that, but the recent spike in demand for engineers + subsequent drop in unemployment this year is well documented [1].
The demand for senior+ engineers has remained steadier through this downturn from my anecdotal observations, with new grads being by far the most negatively affected, but even that seems to both be shifting from talking to people a handful of years younger than me + CS enrollment has already precipitously declined [2] as the narrative that programming is dead because of AI has spread rapidly.
All that leads me to think it's going to be a junk-show over the next decade for people trying to hire as the pipeline was destroyed.
To follow on from that comment, if the growth in breadth of capacity of AI leads to a decrease in the risk of running a smaller business, which I don't think is an unreasonable prediction, then it's not inevitable people do lose their jobs. Employers get smaller, higher-leverage, and more plentiful.
Back then people were mostly farmers, but we already automated that job away.
Not completely, but compared to the middle ages we 50x'd their output. Which is a great illustration what it means to make a job 50 times more productive. We went from 80-90% of the population being required to barely make enough food for everyone to survive, to 4% of the population producing such an abundance that consuming too much food has become a systemic health issue
At the mere cost of destroying soil, and polluting water and the atmosphere in only 200 years! Possibly this will also play out well, and there is a huge market of... maybe social media influencer economy to pick up those being automated out of their previous work... or rather identity, as actually much like in the middle ages, the modern world also makes the profession largely the identity of the individual.
I'm pretty skeptical on the outcomes and the costs also (natural and social as well), but possibly we can have 50x or even more software in the end! The phrase will be truer than ever:
Maybe ironically, but software and robotics should allow us to scale regenerative agriculture in a way that doesn't leave everyone in poverty. We already have lasers mounted to trailers doing precise weeding instead of broad herbicide usage.
Next innovation could be to scale succession planting, which keeps the ground from being exposed in between crops and lets you transition from nitrogen fixers to users quicker, getting more food out per acre while reducing fertilizer usage. But you can't do that with current harvesters and human labor is too valuable to spend on this.
Also take broccoli harvesting, typically you get a few big heads, then it keeps producing smaller heads, but it's not economical to harvest the smaller heads with human labor. Robotic harvesting lets the same plant produce more food per acre and uses the energy needed for new plants instead to keep producing food.
Masses will be unemployed, due to robots displacing them, but human labor will also be too costly. We won't be able to afford a person shepherding, but we will need to produce "meat" (substitutes) in plants, or in inhumane animal-jail, and we'll need robot-weedkiller lasers to produce the feedstock instead of letting animals graze... and we'll give the food produced this way to people on UBI...
This is where this is going, the whole industrialism is totally self-serving, and for every problem its answer is digging deeper in the rabbit hole, and creating 2 more problems in addition to solving the initial problem only half-way.
I don't want to say what you are suggesting is not possibly useful, I just want to emphasize how stuff works out in reality, in addition to doing some nice stuff like what you called out (the halfway solution to the problems). All we get is more alienation and humans getting depressed and feeling a lack of purpose... but somehow we cannot afford to pay fair prices for the agricultural work, and pay fair prices for the food, and not overproduce and overpollute... and the same thing is happening in every aspect of the human condition, not only food production, which is the most basic and ancient activity we have been doing.
Cattle grazing is helpful for fields left fallow, but succession planting is far superior in so many other ways. You can mix plants which repel particular pests with those susceptible to them (and other beneficial strategies), topsoil is grown instead of depleted, flowers are present for wide range of season so bees naturally thrive with food always available, you don't need a significant generator of greenhouse gas running around (cows), and it gets more vegetables per acre so it would be good if vegetables were cheaper because we don't eat enough of them.
I have done succession planting in my home garden, but it's definitely not worth the time investment for the food alone. But it's real neat to see your aphid problem disappear as the nasturtiums pop up without any pesticides needed. You can even feed the world with it, if most everyone wanted to be farmers... (as opposed to some Organic practices which is the same mass farming but the pesticides are "naturally-derived")
Farming has been destroying soil and polluting water for thousands of years. The Tigris & Euphrates used to be crazy fertile, now it's desert. Yes, the destruction has accelerated but farms now feed 8 billion people.
There definitely were what could be considered knowledge workers in the (high) middle ages, it just wasn't the majority of work like today. The knowledge workers then were just a tiny, elite faction, mostly employed by the church or directly by nobility. Kindgoms were still big bureaucracies and needed scribes, theologians, academics, lawyers.
Relatively few anyway. Monks (who wrote and edited books and managed libraries, and also taught), artists and musicians, bookkeepers/treasury/exchequer, scribes/chancery (who were the administrators of the kingdoms), and bankers all existed in European "middle ages". But a significantly smaller part of economy/society compared to "western world" now, yes.
Are you sure? Any functional organization requires keepers to oil the machine. First the government. The best examples were the chinese empire, the catholic church, and the various kingdoms. Or do you think that everyone was either fighting or farming? Stewardship is knowledge work. Bookkeeping is another.
> Who pays for that value, and from what, if all knowledge workers lose their jobs?
They do not care unless these companies can get a bailout.
UBI only exists for companies that are too big to fail. Case in point, 2008 and SVB when there was too much money on the line.
One of the AI companies attempted to guarantee themselves a way for the government to bail them out if they were close to defaulting on the debt from the data center build out.
SVB didn't get bailed out, their investors and creditors were wiped out. You could argue depositors were bailed out -- as they took the undue risk of keeping more than $250k in a single bank (though as part of a requirement for getting a loan from SVB, you had to have your operating accounts with them. So lots of companies had no choice, as SVB was one of the few banks that would lend to them).
Arguably, the main impact of securing SVB depositors above the $250k limit is that it prevented thousands of people from being laid off that week, as their employers wouldn't have had the money to make payroll the following Wednesday.
Thank you for saying this. Continuing to point at SVB as a bailout is annoying. They were not bailed out. Anyone with deposits in an accredited bank should be made whole - always. Without trusted banking we have no economy.
> Anyone with deposits in an accredited bank should be made whole - always
Sure, but is that the case now? Is everyone made whole when a bank fails and they have more deposits than the insurance limits? Or only when it's the well-connected / too-big-to-fail?
So I don't think it's unreasonable to describe SVB as a bailout. Not for the investors, but for the depositors. Has anything changed to reduce the moral hazard / make it less likely to recur?
In early 2023, within the span of two months, the United States
experienced three out of the four largest commercial bank failures in
U.S. history, as Signature Bank, Silicon Valley Bank, and First Republic
Bank all toppled.1 Yet, despite these banks having roughly $300 billion in
uninsured deposits at the time of their failures2 and despite the failures
costing the Deposit Insurance Fund (DIF) of the Federal Deposit
Insurance Corporation (FDIC) an estimated $38 billion, uninsured
depositors took no losses in any of the failures.3 While these results were
striking, they were far from unusual. Since 2008, uninsured depositors
have experienced losses in only 6% of total U.S. bank failures.
...
Formally,
the United States caps deposit insurance at $250,000 per account,6 but,
in reality, the post-2008 financial system comes close to providing de
facto total deposit insurance covering all amounts in all accounts.
So we all now know that a bailout DID occur with the SVB depositors who had all their money in the bank and most deposits were over the FDIC insurance limit. The FDIC insurance rules somehow didn't apply here because there was too much money at risk. (And too big to fail).
But if there was a bank failure at a regionally smaller bank with a regular customer or startup depositing the same amount of money over the insurance limit, their money is gone.
Just like Intel got a "bailout" from investment as chosen by the US government, AI will eventually have a very similar story.
Producing a thing has always been cheap since personal computers existed. From mail-order software companies' times to SaaS times, producing a sellable MVP was an initial cost that is relatively small compared to the later cost of expansion and maintenance. Marketing and selling was and still is the hardest part.
Why do you think of knowledge workers as a fungible commodity?
What makes you think the people who used to build (or would have built) software will switch into the industry of "knowing that the thing was the right thing to build", as opposed to something cooler like surgery, city planning or experimental physics? The roles within a tech company are not the only jobs in the world.
> Why do you think of knowledge workers as a fungible commodity?
I don't.
> What makes you think the people who used to build (or would have built) software will switch into the industry of "knowing that the thing was the right thing to build", as opposed to something cooler like surgery, city planning or experimental physics?
Because it's probably already part of the job. It's a change of emphasis, not a change of career. Your boss can already ask you to do it. If you're producing code, you're probably also reviewing code, checking it matches the acceptance criteria, testing it, sanity checking that it was the right code to have been written, today.
> The bottleneck has moved from producing a thing that works to knowing that the thing was the right thing to build
“There’s more capital than good ideas to fund” has been a complaint from the likes of A16z & other VCs for a long time now. It’s why we ended up with stuff like NFTs getting funded.
I will also tell you, as someone who works at a company that's trying to remain profitable, that token spend has caught the eyes of finance and much like cloud spend they've already started applying pressure to control costs. This May my team is protected to use 30% fewer tokens than we did in April - this was by intention. I suspect we'll drop more in June.
I expect in the future, when these AI companies stop subsidizing costs, the idea of spinning up 20 agents to work on some brain fart idea that you throw out after looking closer will come to an end. It'll be seen like assigning developers on work that hasn't been properly planned for or reviewed.
I think the original argument is too limited in its scope. The wider AI market, which is primarily fueled by OpenAI, Anthropic, Google and the large frontier labs (are there any other in the West, except for these 3?) is spending how much... $900bn this year in DC buildouts? After the spent $500bn last year and they're probably planning to spend just as much the next few years if things go remotely their way.
So yeah, I wouldn't be shocked that in the 2023 - 2033 timespan total AI investment worldwide will be around $5tn, maybe even going towards $10tn.
All that money will have to be repaid, and it will have to be repaid 10x, otherwise heads will roll.
The enshittification we've seen so far is nothing compared to what's coming.
> ... against the actual work their company cares about doing. [...] stuff that matters
This is a key point. Some engineers are having fun doing e.g. greenfield stuff with AI that they never would have had time for otherwise. Whether the company cares about that is another question.
It's related to Goodhart’s Law. If AI token usage is a target, then you're going to get a lot of token usage, but it's not likely to correlate well to improved business outcomes.
Anecdotally, my take on this is that biggest value lever is strategy and alignment, not implementation. The typical company is dozens of little vectors pointed in different directions, and they cancel each other out. Scaling up the magnitude of each is still net zero.
I was recently consulting at org where two separate engineering teams were all in on two different, incompatible deployment platforms and using AI to accelerate adoption of each.
Management was mystified why their engineering leads kept telling them they couldn’t deploy a complete implementation of their solution.
> Why didn’t delta airlines get significantly more operationally efficient in the last 3 months due to the introduction of better software?
The coding agents got good in November. Most individual engineers didn't fully clock this until January/February. This means that companies didn't really figure it out until March/April.
Assuming companies like Delta have adopted coding agents (which would be pretty fast) it still takes months from adopting a new tool to the code results of that tool rolling out to production.
I expect (and would hope) Delta's software development culture is very conservative. Since nobody can confidently tell Delta "here are proven practices for using this tech to produce high quality, more secure code" yet it would be surprising if they were blasting full-steam ahead.
I expect that even companies that got on board with coding agents in January will only just be starting to ship user-facing features that benefited from those new tools. Shipping software takes a long time, no matter how much faster the "typing the code in" bit gets!
Why? All their money is going to Apple Silicon and the five ecosystems, so far in Apples entire history, the largest acquisition has only been $3 billion dollars, OpenAI is currently getting nothing and they gave Google a measly $1 billion refund per year for the use of Gemini.
If John Ternus wants to spend some money, spend it on bringing memory in house. Apple has the money and the engineering talent to do so, have it fab/made onshore in partnership with TSMC.
Do it Apple because you have to not because you want to the Chinese probably will be taking over the memory industry, worldwide, by taking advantage of the greed from three memory companies and their AI overlords.
Seems roughly right, that does seem to be about the boost in the most well-suited cases where you essentially know exactly how to solve the problem, the problem won't change much, and it's truly a matter of just churning out the implementation.
In that case precisely prompting, doing the review & nudge loop, can be a pretty nice (nice, still not game changing) speed boost over literally typing out the code to match the design in your head.
The less optimistic view though is that most things you build aren't like that. Even if they seem like it first. These things get booked as a nice speed boost, but you'll only find out much later they weren't.
A confounding factor is that it seems like many people not in the detail of building software do seem to think of most to all things are like that, even before AI assisted coding. Not much need to say more - see the entire history of the 'agile' movement for evidence of this.
And because most things aren't like that, I actually struggle to see fundamentally how more than 20-40% will ever be achieved (short of the ever-present deus ex machina of AGI argument), simply because the generation is already really good for these types of things. So since things like this aren't going to increase in overall proportion of things to be done, I don't see where the overall extra gains come from by models improving at this point.
I mean this case with AI-productivity fires itself back when we talk about GDP.
The more AI causes productivity increases, the less and less number of workers will be needed. This will heat up the job market even more and bring salaries down.
Net effect of this productivity increase: less consumption by the masses, even though you may be producing more good and much more efficiently.
A third effect also comes into play that once all this starts to happen, common people, who are generally living paycheck to paycheck, will now start to hesitate towards making any long term investment, housing included. And that indirectly will end up impacting financial and banking sector, which will then impact existing savings, bonds yields and retirement funds, and the recession-like cycle starts.
This productivity increase only makes sense if it is capped to a very small number.. like 20% max. Beyond that, who these companies will even be selling to?
> The more AI causes productivity increases, the less and less number of workers will be needed.
That only holds if companies have a fixed need for "productivity" which is met by their current employees, such that their employees becoming more productive means they need less of them.
Every company I've ever worked for has wanted to achieve way more than they are able to get done with current resources.
But generally yes, the biggest open question about all of this is how the impact will play out on the economy, job opportunities etc. I've not seen anyone come close to a confident prediction about how this will play out.
It's hard to imagine how making insurance sales cheaper for the brokers, churning out astrology apps faster, AI boyfriend bots or running ad campaigns with fewer and lower paid designers is going to drive 10% GDP growth in developed and middle income countries, that's the sort of figures you see when very poor countries finish rolling out electrification, sanitation and transportation.
>The more AI causes productivity increases, the less and less number of workers will be needed. This will heat up the job market even more and bring salaries down.
>Net effect of this productivity increase: less consumption by the masses, even though you may be producing more good and much more efficiently.
Big tech companies can't even create login flows and account recovery flows that work for everyone yet. There are countless stories of folks losing access to business Instagram accounts that get hacked, Google support from a human to fix a problem that is outside of their help articles is non-existent, etc etc. There's still so much "low-hanging fruit" IMO that isn't particularly fun or exciting to fix, but ask your average non-tech friend or family member what they think of the Facebook + Instagram security settings pages / sites / desktop-only settings.
Who is going to pay for all of these subscriptions that will power this GDP increase when average purchasing power of those outside of the top ~10% of earners is decreasing YoY? We're headed toward food and water shortages next to sprawling datacenters, not shared societal prosperity and a healthy middle class.
First of all, common people are not living paycheck to paycheck in the sense that they're at risk of not having money[0]. This is corporate content marketing that has entered the collective memory of people, not anything close to reality.
Secondarily, reducing the cost of making a thing doesn't always mean you get less of a thing. For me, certainly, what happened is that I write way more software than I originally did. When we built compilers, the amount of human engineering effort required to do things plunged, but the amount of software engineering jobs didn't go down.
This is as bad as models will ever be. That part is true. And it's entirely possible we go foom. But it's also possible we don't, and then it depends on where the asymptote lands.
Respectfully, that is truly ignorant. The vast majority of humans do not have any savings and would be in big trouble if regular income ceased. No paycheck no food. It’s wage slavery and it’s pervasive.
Nope, if AI were to realise the hype, you have to take into account macroeconomics. Usually this isn't a problem for most businesses
>The more AI causes productivity increases, the less and less number of workers will be needed. This will heat up the job market even more and bring salaries down.
People also underestimate that the reason why companies are so excited about AI isn't to increase productivity, its to fire workers and crack down on worker rights. They won't lay people off because AI means they don't need as many people to get the job done, they'll fire everyone while doing a much shittier job, because they hate having to abide by worker's rights and pay people
> The more AI causes productivity increases, the less and less number of workers will be needed.
Why does this have to be the case with AI but it didn't have to be (and wasn't) the case with the steam engine, electricity, the automobile, or the computer & internet?
Certainly, AI could be different.
It's curious to me why the vast majority of people on here think it must be different.
Because the previous revolutions only automated a small subset of jobs, it didn't automate manual work.
Some people take the view that AI could make knowledge work largely irrelevant. Any niche humans could carve for themselves would only live long enough to generate training data for the AI to automate.
> The more AI causes productivity increases, the less and less number of workers will be needed
This might not necessarily be true. Increased efficiency creates induced demand to the point where more workers are needed. Because the new capabilities unlock more value to extract and the economy rushes in to get it. The steam engine is a huge example of this
I dont exactly know what new value genAI will unlock but i think its more likely than not
And yet the job everyone loves to hate, the humble "burger flipper", continues to resist automation yet command minimum wage labor rates. This future of either being a CEO of a company consisting primarily of AI agents building some monthly subscription-based solution to some trivial digital chores OR manual labor that isn't [yet] fiscally viable to automate seems quite bleak. We'd also need a ton of robot technicians and manufacturing that the US has neither the educational and training institutions to support nor the will of the population to fill. Given the ongoing war on immigration, visas, and foreign-made hardware, if this continues, good luck.
Hey, I wrote this down one time.
I estimated way higher yearly revenue required, to be adversarial. And you can keep the "cost per unit AI work" a parameter and play with the results.
But the point is that if people are willing to delegate part of their salary (e.g., buy consumer products), vs requiring employers to pay for the tokens, then it's quite possibly a net win. Something like "I pay a largeish fee every month to make my own job much easier", similarly to how we buy a car to make commuting easier.
You are making the assumption that the models are only used / paid for by 2.5% of the population (your knowledge workers value). There will be new value created by these models which people are happy to pay for which simply did not exist at all before. It is also naive to say that the hyperscalers are going to be expecting a return on this in 5 years, it will be entirely propped up by investments / IPOs as has been the case with any tech company for decades now to reach scale. The hyperscalers are currently spending ~650b combined annually, which they have the cash for and can sell in future compute instantly.
I'm sorry, what the feck does "value creation" mean here? I live in a place where people are so, insanely squeezed from every angle. Wages are stagnant, prices rocketing. Where is the money to pay for this value going to come from?
No one I know feels richer than they did a decade back. I've not been able to meaningfully put up my prices for a decade. People are tired and stressed and scared, particularly scared of a technology everyone keeps telling them will make them redundant.
There is no rising tide lifting all boats, just most of us drowning whilst a few whizz past in their yachts.
I honestly hope these guys faceplant ASAP. Couldn't happen to a nicer bunch of people.
Feelings aren’t fact. A lot of data shows the doomerism is not reflected in the actual numbers and much of it has to do with rapid inflation and continued vibes.
Consumption has risen, inflation adjusted wages have risen for blue collar and white collar alike. Most social mobility has been the middle class moving into the upper middle class, not moving to the lower class.
The main thing holding people back is the housing crisis. This is orthogonal to the value creation of businesses.
Value creation is growth. If it didn’t exist the S&P would still be 42.55$.
> The main thing holding people back is the housing crisis. This is orthogonal to the value creation of businesses.
This feels wholly at odds with saying most social mobility is upwards. So most of the social movement is into a class where a home and vacations are a given, but we also have a growing class of people who can't afford a home? Per BLS, average real wages are down 0.3% YoY https://www.bls.gov/news.release/realer.nr0.htm .
> Value creation is growth. If it didn’t exist the S&P would still be 42.55$.
This reductively assumes "value creation" is the only effect on the S&P pricing. You'll note a ton of graphs correlate with it, e.g. https://tradingeconomics.com/united-states/inflation-cpi is the US inflation rate, which also tracks the S&P pricing. Ie if a company is worth $100 a year ago and inflation was 4%, I'd expect to pay $104 for their stock with 0 value creation whatsoever.
Home ownership is not the definition of economic class.
The s&p has vastly outstripped inflation, this isn’t even an argument. It’s a very bizarre and uninformed opinion to say “inflation is correlated with s&p value”.
In economics, you deduct the inflation rate from growth to get the real rate of return.
I wonder why so many people with such little understanding of financial markets make comments like these.
The upper decile of income earners account for more than half of all consumption in the US. Household balance sheets and wealth have never looked stronger, again when you account for all the appreciated stocks and properties owned by the upper quartile. True incomes for the lowest decile rose significantly for the first time since 1970 in 2022 and then sort of stayed flat again. Sure, statistically significantly, not "significantly" as in personally meaningful after figuring in rising consumer costs. There is a narrative where you can see all this as hugely positive but this is also largely a "vibes" based narrative. I don't know why you'd expect most people to care about what the "vibes" are like for the best off in society, that's a bit removed from their daily concerns.
> Consumption has risen, inflation adjusted wages have risen for blue collar and white collar alike.
My wages haven't risen for nearly 5 years, while inflation has occurred over the past 5 years. Why the blanket statements?
> The main thing holding people back is the housing crisis. This is orthogonal to the value creation of businesses.
Are you suggesting a "housing crisis," in your words, wouldn't impact consumption? I'm watching my spending (and living like a child in his parent's house, except it's not my parent and I have to pay for it) in the hopes that in about a decade, I'll have saved up enough of a down payment for a home somewhere in my state that I could actually afford the mortgage on the remaining amount. There are plenty of things I'd potentially spend money on but won't as long as I feel like I'm economically stuck and have a chance in hell of saving my way out of it. So this feeling translates to fact.
If you think my personal experience is just an anecdote and doesn't count because it's not being told through the lens of large-scale numbers, fine. But I really agree with the person you replied to that you're gonna have to be a whole lot more specific than "value creation" if you want people to spend money on your AI products "in this economy," whether it's because they're actually strapped for cash or just pretending like you seem to think they are.
Sounds like internet sentiment and not research data.
It's kind of become socially taboo to not be suffering "in this economy", but on paper it's hard to see weakness in places that there isn't always weakness. As long as the 65-95% are doing well, there isn't going to be a collapse.
The most recent U Michigan 'Survey of Consumer Sentiment', which is THE authorative source in the US, shows consumer sentiment at the lowest levels since the survey started in 1977
A literal example is that I can use AI to file my taxes instead of spending a weekend and hundreds of dollars to have an accountant do it for me. It costs me like $5. that 245$ delta is the value of that output to me, as long as I am confident it is correct.
Seems to be a thing in the US to need specialised software, an accountant or AI to file taxes.
In most of Europe individuals at least don't need any of that.
I'm in France and it's just a connection to a government run website to enter a few figures, takes less than an hour most of it is already pre-entered (salary etc), the main thing to add manually is charitable donations.
If you're running a business then yes an accountant can be good (or be required depending on the legal form of the business) but not for individuals.
In the US you don't need software or an accountant either. People have been convinced it is hard, but the reality is taxes are basic math for most people. It isn't hard to copy the few figures from a couple forms and add them up, the look things up in a chart. It is tedious, but not hard. That said, I use the software anyway because once in a while I make a human mistake and those are annoying when the IRS catches latter.
There are a few people who have a complex situation who need an accountant. Most do not, but since they have never looked at what is really involved they don't realize what is going on.
Taxes are one of those things that seem difficult and people reach for tooling or expertise without trying initially without, but are pretty easy to do yourself just filling out the forms.
It really depends your situation. If you’re entrepreneurial, self employed, others things become messy fast, and you cannot just know what you need from filling forms
Part of the value of paying an accountant is that you can get representation in case you are audited. Though I guess you did say you were confident it is correct.
I think that to sum things up, we will have to wait until we can evaluate the cost of the mistakes. You could be lucky but you could also end up with a very negative output value in the longer time frame.
Not speaking to the wisdom of filing taxes using LLMs, but just FYI (assuming US here) taxes being accepted doesn't mean they were correct. It just means the IRS hasn't found anything major wrong (e.g. SSN used on multiple returns). Even being approved isn't a guarantee, an audit could come later.
Audits often find incorrect in favor of the person audited as well. If you are not audited that is a bad thing as you have no idea if things are wrong, and it could be costing you a lot of they are wrong.
Thats the thing; the "increase in productivity" isn't being felt by the general public, the end user. If your "increase in productivity" just means more money being shifted around at the corporate level then it is meaningless.
> There will be new value created by these models which people are happy to pay for which simply did not exist at all before.
True, but I think the GP's point was that what consumers will pay won't be nearly as profitable as what enterprises will pay to increase the output of their developers and knowledge workers. ChatGPT is currently the overwhelming leader in consumer AI usage but only ~5% pay $20/mo.
As a recently retired serial tech founder, I'm now one of those consumers. I use AI webchat daily for general search, Q&A and even to write little automation scripts for myself, yet I haven't paid anyone anything for AI yet. Even after being heavily restricted and performance nerfed to hell in recent months, free webchat AI is still fine for everything I do, and I'm not remotely price sensitive.
Even as AI compute costs fall over time, I doubt serving ads against AI webchat to consumers will generate the kind of high-margin, sustainable growth VCs get excited about. It's so undifferentiated I bounce around between all four leading providers because there's virtually no moat locking casual consumers to any chatbot beyond a single question thread. I guess if it had a nearly infinite context window seamlessly integrated across all sessions, that might be somewhat sticky for some consumers but it could also get creepy for some others - and it would devour gobs of the scarcest resource in AI. Beyond Maslow's Hierarchy of Needs, the mobile phone is the largest revenue, long-term mass consumer product ever but I just got a new flagship phone from a top-tier provider for $30/mo over 3 yrs. IMHO, even an all-you-can-eat, infinite context window, next-gen Mythos couldn't reach and sustain mobile phone levels of global consumer adoption at ~$20/mo. Unlike professional developers and knowledge workers, consumers don't have any "job to be done" big enough for an LLM to command that much of their zero-sum discretionary spend.
100%, a driving factor will likely be how good we can make models that are so small they use almost no compute. Until then it is a race for adoption and moat-building (or screwing people over?) once you have users
> a driving factor will likely be how good we can make models that are so small they use almost no compute.
That will certainly help but it doesn't move the fundamental limit because resource efficiency is a cost driver not a demand driver - and my argument is against the thesis that lying beyond professional devs and knowledge workers, there's an untapped trillion dollar industry serving LLMs to mass global consumers.
Using Simon's cost estimates, I agree that halving the current $1,000 - $1,200/mo MSRP to profitably serve frontier inference to professional developers and knowledge workers (PD&K) will help Vendor A steal share from Vendor B or C. It will also increase LLM sales penetration into the segments of the global PD&K TAM which can't afford ~$1K/mo for every seat. A fair chunk of the PD&K workers in many SMEs aren't included in today's ~$1K/mo per seat license pool, especially in 2nd and 3rd world geos. When the price falls to $500 and $250 most will but that's still just saturating the existing PD&K TAM - not pushing into mass consumers.
While the PD&K TAM is big, justifying Trillion+ dollar capex spend requires believing the TAM is much more than PD&K and eventually grows into converting a couple billion non-PD&K consumers into ~$20/mo subscribers. I don't buy it for two reasons:
1) The Comps: There are vanishingly few examples of long-term, mass consumer adoption of a discretionary technology at that scale. Mobile phones at ~$15 to $30/mo are the obvious one but LLMs are nowhere near being that valuable to the average plumber in Des Moines, baker in Jakarta or retired nurse in Hamburg. Pondering it, I just imagined forcing any of those people to choose between their mobile phone and an LLM chatbot. Sure, some who are flush with cash might choose both but for most consumers in the world ~$20/mo is big enough they'd have to pick one and ~zero percent would choose the LLM over their phone. After mobile phones, the second comp for discretionary tech spend I thought about was XBox and Playstation monthly gaming subscriptions but combined they have less than 90M paying subscribers and the ARR is just under $10/mo. As an industry, "Big LLM" is spending well over a trillion dollars every five years. XBox and PS ARR doesn't even cover paying the interest on that capital, much less the 3 to 5x returns hedge fund investors are betting on.
2) The Alternative: It's useful to doubt my own intuitions and one counter to my skepticism is to assume "But LLMs aren't finished yet, they're going to get much better." How much better could an LLM which can be profitable at ~$20/mo get than Claude Mythos in the next five years? Instead of debating future unknowables with myself, I've found it's better to just imagine the most perfect future product I can that's still realistically plausible. So, let's imagine we're willing to spend a million dollars a month to very unprofitably deploy a prototype to test the consumer demand for "Tomorrow's Awesomest $20/mo LLM" today. So we gather a few hundred super smart, broadly knowledgeable intellectuals together at one top-tier university research library, where they'll have access to every commercial database and unlimited Claude Mythos 2.0 and ChatGPT 6.0. Since our experimental budget is $1M/mo we can afford to add in several Nobel prize and Fields Medal winners too. They'll work together manually reviewing and improving not only every LLM answer but also our test user's prompts - and of course our test chatbot will have human-level real-time speech recognition and vision (via Zoom and screen-sharing with actual genius-level humans), making this truly a test of the "smartest, most accurate, best consumer chatbot" we can imagine.
Now, let's run the test by having one thousand mass consumers try it out and see how many Des Moines plumbers, Jakarta bakers and Hamburg retired nurses we can convert to a 1 year @ $20/mo subscription for our $1M/mo ultimate chatbot simulation. Playing this thought experiment out in a bunch of ways, I find some percentage of outliers, iconoclasts and closet intellectuals would go for it but... the vast majority just don't find it enough better than "free" chatbot alternatives AT&T includes with their phone subscription or Samsung bundles with Galaxy phones - despite only being ChatGPT 5.4-level. It turns out, most plumbers, bakers and ex-nurses don't have a compelling "job to be done" in their daily lives that even an MoE panel of actual Nobel and Field's medalists with ivy league professors can make enough more valuable than an inferior but free-to-me chatbot, in the judgement of our Des Moines plumber. While the world's smartest chatbot is nice, when it comes time to pay, he prefers having one additional premium football match on TV and a six pack of cold beers every month.
I'm having a hard time understanding this huge post that doesn't talk about enterprise users. I'm convinced that the consumer isn't going to be coming up with enough money to justify AI valuations... but doesn't this just mean that we expect the money to come from large enterprise users?
A recent post here said AI spend could be "20% of every software developer's salary"... and that seemed plausible based on productivity improvements. That's not about a phone bill.
> doesn't this just mean that we expect the money to come from large enterprise users?
Yes, if you agree that not enough consumers will value "smart paid chatbots" over dumb free chatbots at ~$20/mo, then, as you say, the money has to come from enterprise developers and knowledge workers (PD&K). The big problem with that is the numbers don't work. There aren't enough PD&K workers that are highly-paid enough to justify $12,000 to $16,000 a year to cover the astronomical spending run rates.
Without recapitulating all the various scenarios AI CEOs like to hand-wave, my take-away is the scenarios which show today's frontier AI vendors earning the returns they've promised investors to get those trillions of dollars they're spending ALL require truly extraordinary deltas far above current historical actuals. Whereas ALL the scenarios which I find more plausibly realistic, even if still quite bullish, never even get near the required ballpark.
I've tried to make various scenarios fit but inevitably, when the PD&K demand side starts getting implausibly inflated, I start pushing the cost side down to keep the adoption rates remotely plausible. For example, assuming things like the number of PD&K workers will grow at 2X the highest rates ever seen before or that the percentage of the tasks knowledge workers actually do at sufficient frequency to really matter and which are also "LLM improvable" is implausibly vast.
It gets challenging because one quickly hits finite limits on the increase in tangible economic value that an LLM can possibly deliver on many common knowledge work tasks like drafting an email or a product proposal. Even if we assume next year's frontier LLM is so good literally ZERO slop is even possible so that our knowledge worker doesn't ever have to review anything - turning a 20 minute task into 2 minutes, or even 200 milliseconds, yields finite economic value to the enterprise. Even if you go extreme and suppose pretty crazy stuff like 18 months from now LLMs have eliminated 50% of PD&K workers, that messes with the spreadsheet assumptions because now the there are half as many $16,000/yr seats so the LLM license prices have to double again just to stay even.
At the edges it quickly gets nuts. I even tried assuming that in five years LLMs eliminate 100% of non-managerial PD&K jobs. All actual work is done by Super-AGI LLMs. Even if we assume such amazing intelligence can be profitably sold in five years for only 4x the price of today's far dumber LLMs, that Super AGI still costs $50,000/yr per human replaced. Even that assumption only cuts the blended labor cost of non-managerial roles roughly in half (depending on industry and geo). In the end, no matter how much labor cost LLMs enable companies to cut, it still only reduces overhead. Lower costs might allow Vendor A to steal some of Vendor B's customers but it doesn't increase the total TAM or demand of the entire sector both vendors serve.
Once you're out of plausible labor savings, one has to move to assuming that within five years Super-AGI LLMs working unsupervised will be making and validating fundamental scientific breakthroughs, then reducing those breakthroughs to engineering practice enabling new technologies which create entire new markets. Then they'll create whole new companies with maybe only 1/10th the humans and profitably grow those new markets. It quickly starts to feel like bad sci-fi instead of plausible, near-term financial scenario planning. Is stuff that extreme and unprecedented actually possible? Sure, incredibly unlikely and unprecedented things DO sometimes happen but things so far out of the distribution are very rare. But when making Sam and Dario's scenarios actually 'math out' starts requiring more than one "Black Swan" unlikely, historically unprecedented, earth-shaking event. And then goes on to need whole flocks of Black Swans, it just isn't credible. That said, I do believe that some parts of these scenarios may happen in 10 or 20+ years. It's just that the investment theses trillions worth of our 401K's are sunk into assume five or six year amortization and financial payback.
What are the non-tech people in your life using AI for? $20/month, next to Starbucks and avocado toast, is discretionary. Maybe the novelty will wear off and non-tech consumers will leave it in droves, but everyone declared they'd leave YouTube if they started playing ads, but YouTube doesn't seem to have noticed.
> What are the non-tech people in your life using AI for?
Mostly asking random questions they used to search Google for.
> next to Starbucks and avocado toast, is discretionary.
Sure, but your description implies highly affluent, urban professionals in western nations. I was talking about getting several billion global mass-market consumers to all keep paying ~$20/mo. Mass consumer adoption of mobile phones worldwide is currently >5.8 billion or >70% of humans alive. Only ~50M people are paying $20/mo for an LLM and I suspect many of them are not pure consumers but actually knowledge workers that AI vendors are losing money on and will eventually force into higher tier plans just like the $200/mo developers they're currently losing money on. These heavily subsidized loss-leader offers are all going away post-IPO.
Personally, I know maybe a dozen people who pay $20/mo for an LLM but only two of them are really 'pure consumers' who don't use it for knowledge work. Both of them are multi-millionaires and neither has had a job in ten years. One is retired like me and the other is so wealthy she has a Netjets credit card and has new cars delivered like some people order shoes. Everyone else I know paying $20/mo is a professional who uses the LLM for a lot of office or knowledge work and writes it off as a business expense - examples include a couple of attorneys who are senior partners in a law office they own, a solo architect, and a dentist who owns his own practice.
At $20/mo, AI vendors are probably losing money on most of my professional friends because they use it pretty heavily all day. They're only making money on the two multi-millionaires who both use it so infrequently they could easily be using free chatbots instead but are so rich they could lose $10,000 in their couch cushions and not notice. While they are profitable at $20/mo, they aren't exactly "typical consumers" that there are billion more of. I expect AI vendors will find ways to force my lawyer, architect and dentist friends to switch to higher priced plans soon because they're really knowledge workers abusing a consumer tier plan into unprofitability.
But will they pay big actors running top end models for that? You don't need latest openai or anthropic model to go thru your mails, get summary of the some products from web, or to do your to-do list.
The AI might very well be used by noticeable % of population daily, but that doesn't mean they will be paying trillion dollars to the leading US AI companies
> They've got, ballpark, $5t to $10t to make back in the next 5 years, or the hardware buildouts will start getting written down.
Depreciation and write-offs are about accounting models. Hardware will still be running after five years and still be making money. They may not be as efficient as the new hardware, but they will still be making real money even though they are valued at $0 in the books.
Also, not all developers work on software products. The vast majority of developers work supporting software solutions as part of a much bigger business model, such as infrastructure, industry, healthcare and services. Many of these are complex organizations. So, unless you get to turn every employee into a 10x employee, the 10X coder along won’t necessarily make a 10X productivity contribution. What’s likely going to happen is the 10X coder will start to slow down or adding more (unnecessary) complexity to avoid having to sit and wait on overhead, for other areas of the business which are not easily automated away to AI to catch up. As a developer I can finish my project in June instead of December, but what if the customer is still not ready for integration until December? what do I do?
At some point, if we reach stability on the models, we'll start getting silicon optimized for individual models. They are optimizing for time to market, not efficiency right now. I don't know how much it will move the needle on the cost math, but at this scale any improvement has a crazy multiplier.
But that means going back to "80% profit margin" Jensen and further digging your capex hole. The benefits would have to pay not only for current capex but also past capex.
But by then, I will be able to go one line down in my dropdown menu to switch to a newer LLM provider who doesn't have to amortize those past capex.
Yes, the whole thing feels dot comish. I’m betting there will be a few (maybe only 1-2, like what happened to search) winners, and everyone else is in for a bad time. It’s the same dynamic too: the winners are going to win so big that everyone wants to get their money in for a chance at a piece of the pie.
Those are rookie numbers. We are going to blow past $1t per year in spending in no time. As a developer for 29 years, I couldn't go back to coding by hand. For better or worse, AI will be woven into the fabric of life in no time.
If people figure out how to run agents on-prem (already becoming feasible for both agentic tasks and coding on consumer hardware like Mac Studio 128GB+ or DGX Spark with some models) these companies will be in deep trouble.
I don't think the unit economics are too terrible. Expensive, but not impossible.
200m knowledge workers in US and EU. Total salary around $15T/year.
$1T/year in token spending is about $5k/year per person. A big number, but not totally mad. That's the low end for office space per person for example. Probably close to the existing SaaS spend per person for a lot of roles.
We are still early in the deployment cycle for these tools so I would expect them to get better and also cheaper too.
One factor to consider , the base will not remain the same over the next 5 yearts.
Every generation of developer tooling that increase of absolute code throughput creates a new class of developers (and users).
Always been the case since first compilers, through eras of frameworks to today, and the skill level needed to be one has dropped. In mid/late 80s only Master / Doctorate level Comp Sci professional could write any applications. It dropped to undergrad and just Information Technology engineers and comp sci theory became mostly optional and dropped further to any college level educated with some training and has been trending below with no/low code tools like retool pre 2022, that was before agent codegen services such as v0/replit and so on.
The next generation developers will not produce applications and architecture as previous generations did, just as we most of us here don't produce the level of quality that pg did when building this platform[1] , but as long as the user can find value it doesn't matter as countless enterprise applications of middling quality already prove today.
All this to say the 200M/30M numbers will not remain the same is the thesis for these businesses, will it change by large enough at a fast enough pace to justify the capex, I don't think so either. However web 1 then 2.0 , saas and mobile revolutions were pretty quick with new class of users and developers so not completely unrealistic .
[1] While HN is a heavy outlier with its custom lang lisp implementation, there are any number of examples from previous eras that are more moderate in choices but written with solid architecture with skill levels would be hard to find in today's generation founders.
I agree in principle with the math. But I believe that in reality if revenues don't show up quickly, then lenders will just restructure the debt and defer the payback period. Similar to SF commercial real-estate; many buildings should've come due during the depressed covid market, but lenders (banks) were willing to delay payment until the market picked up again.
The scale of these investments put the lenders at substantial risk, so the lenders will do anything to make it work. If the current lenders will be damaged by extended payback periods, they can simply sell the debt to someone else who won't be.
So you've got that market. Let's call it the demand BY knowledge workers to do the work. You've also got:
2. The companies themselves buying tokens for operations to make the work more efficent.
e.g. Salesforce agent or Microsoft Office agent or random saas inventory agent. (and if you say those will go away (which I don't believe), it's even more bullish. The tokens just go to someone vibe coding XYZ, which is EVEN MORE than if you were to buy saas because it's SaaS product x Companies that built it instead of just one)
3. The companies SELLING tokens. This is also new markets like schools and small business (e.g. the local gas station buying an inventory tool)
4. The consumers "buying" (I put in quotes because it can be subsidised but the company) through chatgpt, strava, instagram/netflix recommendation, etc.
Local models still take compute, and while it may be cheaper, it is the same argument of on prem vs cloud. No one operates on prem unless you HAVE to for regulatory. Margins will come down and you just spin up a GCP/OpenAI/Anthropic agent.
It may be "cheaper" but rationally its better to pay someone to manage it. Thats why Hetzner only had $367M in revneue (a lot but tiny compared to managed services)
Knowledge worker compensation is 35 - 50 trillion a year globally (6 - 12T in the US alone.) That's a huge TAM. It's still close but 5T over 5 years seems doable.
>... unless we figure out how to make developers 2x, 5x, 10x as productive on stuff that matters, this isn't going to play out well.
The way we make ICs 10x productive is not just making each of them individually more productive, but by removing the coordination overhead of large organizations, because overhead scales super-linearly with the size of the org. And orgs will shrink automatically as AI-assisted ICs take ownership of larger and larger scopes of work, leaving much more budget for tokens.
> We're talking about a world where you need 5% of every knowledge workers salary to go into tokens. 20% if you're a developer.
Just realized something: if one worries about losing jobs to AI, token's high unit cost is good news. To say the least, high cost would delay the displacement, if any, right?
In the meantime, someone shared the below on X. I guess the moral of the story is that "good enough" does not just displace software engineers, but also models.
> I Went From $3,000/Month on Claude to $5/Week on DeepSeek
> And honestly?80% of my work is identical.
> For the past two months, I was burning $3-5K monthly on Claude Code. Every idea from design to development to testing - full end-to-end automation, even simulating users to test my products and provide feedback.
> Extremely token-intensive. But Claude's caching sucked, making it insanely expensive.
> Then I discovered DeepSeek V4.
We're going to reach a point where these companies stop asking for money and start mandating it. They've got a vice grip around the nuts of many governments and loads of companies have gone all in on investing in these slop heaps.
At some point, companies are going to start removing basic features. Governments and essential services are going to make people go through chatbots to get basic service. They're going to require AI to validate stuff that's already automated and working fine. Google search? That'll be all AI (and I guess they're already rolling it out). Dentist appointment? Going to need to do it through some AI app that requires an account and tokens "for a better patient experience". Verifying your ID when buying alcohol? Going to need AI to scan it and take 90 seconds to determine whether it's real. And it'll say you're an 7 year old farm worker in rural Botswana, so you can't get alcohol. And they're going to milk money at every level of this.
Plus, at some point there are less tokens because local models being optimised and can work with protected information. For enterprises that want an AI with a knowledge base of internal documents, this becomes more interesting by the day.
Depreciation starts on day 1 and most likely they IMHO dont have 5 years. They dodged the deepseek bullet but who knows what is out there that will make all of this investment essentially worthless?
I hear this and I keep wondering what I’m missing. My productivity has shot through the roof over the last year as a result of having these tools. I’ve been able to unlock projects that I’ve wanted to do for years.
YEPPP... and I'm kind of shocked at how many people can't do simple math.
Let's put it context. Google's annual revenue seems to be north of $400B. So if OpenAI suddenly had Google's revenue, it would still be insufficient to recover their investment.
and it's a ticking time bomb because $1T in servers, CPUs, GPUs and memory is going to be worth $200B in 5 years. You can say they can keep using what they've got. Sure. But they're also not going to stop spending on new hardware. And the competitor that comes along in 5 years and spends $1T doing the exact same thing is going to have a huge advantage.
OpenAI at this point reminds me very much of the Russ Henneman pre-money hype cycle.
It's actually worse than that. It's not just financial depreciation or that the existing hardware becomes obsolete due to being less powerful than new hardware but also that hardware being run all the time at high load actually has a limited lifetime of a few years so it will physically break...
Data centers come down to performance-per-Watt. Electricity accounts for 20-30% of a data center's operating cost [1]. I don't know the exact breakdown but the GPU part of that is probably the majority given how power hungry GPUs are. The B200 is upwards of 1200 Watts [2]. The B200 is rated at ~4.5PFLOPS of dense FP8. So you're getting 3.75PFLOPS/W. We don't know what the next generation will look like. The A200 (Hopper architecture card that preceded the B200) had ~4PFLOPS apparently but also lower power consumption. Obviously this changes depending on whether you're looking at dense or spare and FP8 vs INT8 vs INT4 vs FP4, etc so we're just using FP8 as a yardstick.
Imagine a fictional B200 successor, the T200 that has 8PFLOPS of dense FP8 at 1000 Watts. Well then a DC built on that where the T200 will likely cost similar to what the B200 does now, you'll get nearly double PPW so the same size DC and same electricity load is going to be like 2 of your old DCs in operating costs. That's a big deal when you've laid out a trillion dollars.
This should be the top comment. Also, I think its not that many people, including our Simon here, are not good at math. Its more like, some of them seem to be incentivised to not be cough, cough, "good at math". How else will the hype sell?
I mean, a company that loses money on every widget they sell might technically have found "product-market fit." :)
It seems quite possible to me that developer tooling is going to end up being the biggest win from LLMs because there is a product-market fit -- and also quite possible that OpenAI and/or Anthropic end up getting bought for pennies on the dollar because their burn rate is unsustainable. AI may end up being this generation's "dark fiber."
At a certain point, I genuinely feel like the best way this hype is being sold is by making people genuinely believe in it.
and in that sense, if Anthropic and OpenAI are able to create the projection that they can-be profitable despite finances seeming bubbly at best, I think that what happens is that these companies spew so much amount of content that people like Simon get into it too.
There is a deeper problem of people falling into AI psychosis too, in general, I am not sure if Simon has fallen into it or not
I think that the greatest point which can be made here is to not offload your thinking to others and to think about the situation yourself. Sounds familiar (looks like we are all off-loading our thinking itself to machines)
Side-note: As humans, we have a tendency to quickly judge or make quick decisions which stems from our times foraging and scavenging in jungles.
Another Side-note: at a certain point, I am unsure of how much to think about AI or not, certainly discussions about it that were happening 2 years ago weren't helpful in contexts that they are used now (well not in any way or form that a person discussing and getting into the weeds of AI 2 years ago is better than a person just getting into it say 2-3 months ago)
With the industry (moving so fast) [but that doesn't mean that you can't catch up with it, I feel like the fast word has made people think that they are falling behind which is imo wrong i suppose]*, It is basically unsure to me of any FOMO or anything if you aren't using AI already, I find this notion naive.
People might be making strong opinions (AI psychosis) and skills on the tools available at the moment the same done 2 years ago. We don't quite know about the tech as these are still black-boxes and how they progress and what these "AI skills" might survive or not in future. Heck, we aren't even sure if these tools might survive or not or wouldn't be made magnitudes more expensive simply to break even as they are given to us for the first time at percentages of the price.
I don't know if I should form (strong) opinions yet and also a question of its worth so much thinking efforts in the first place, probably just gonna do my own thing (the way I want to) which includes learning C at the moment. because learning is fun.
I didn't exactly say that they were about to become wildly successful companies. I suggested that they had "found product-market fit" - not too impressive for more than a decade of work - and that their revenue may even be "enough to start covering their costs".
Firstly thanks for responding and I wish you to have a nice day. your suggestions have value and I appreciate you writing the article. Perhaps enterprise businesses do end up becoming the fat and meat of the AI industry.
My question which I wish to ask: What would happen to these AI companies if they turn out to be anything but wildly successful companies, both to the investors who have already invested in it and to those who might be investing indirectly into it in the near-future (passive investors, retirement funds)
> What would happen to these AI companies if they turn out to be anything but wildly successful companies
I'm not nearly enough of an economist / finance person to answer that credibly, but I expect they'll go bust, and a lot of people will lose their shirts.
... and the model weights will be sold to other companies who will then run them at a profit, and eventually figure out an economically sustainable way to train new ones.
The 1800s railway booms are a good comparison here - a lot of companies went bust, a lot of investors lost money, and we still ended up with railways.
If the AI companies all go bust we're going to have a lot of spare data center capacity!
> If the AI companies all go bust we're going to have a lot of spare data center capacity!
I can be wrong I usually am but an AI DC != compute DC or that it might decrease the prices of servers substantially because of it. (well not exactly, I hope you read my whole message so that I am able to better explain what I am saying.). AI DC's try to optimize for one thing: running GPU's for immense scalability and flexibility (0 to numbers>=large_number).
Currently, its actually way worse, the server providers are some of the worst impacted by the industry at the moment because each server requires ram and ram is well... increasing in its price exponentially. It's really a tough time to be a provider at this time (in certain respects) directly because of AI.
It is unclear to me if spare DC capacity will have any meaningful impact to it. I don't think that atleast within compute (and not GPU/AI DC), that space was too large of a problem.
Fun fact but one of the largest providers (BuyVM) had its datacenter price from where they colo'd increase because of the immense demand at the moment for spots in datacenters by many tens of thousands of dollars that they did the first price hike in at this point at decades! The situation is this dire :-(
Ram prices might come falling down and DC's might get cheaper but they can only get cheaper to limit, they still need to for example DC security employees
and I wish to suggest that if anything, investors might wish to re-coup their losses within the AI loss, they might want to make up with what little they might have (ahem DC)
For example, if you wish to want to take at an even more egregious example of what I am suggesting, there are many new york LLC's who would much rather leave the properties that they own empty rather than decreasing the price of what it costs (which they have set to some egregious amounts). I think that for them, somehow the math ends up working out in the end somehow, so there might be something more to it.
I wish I was optimist but I don't believe that the gains in spare data center capacity are worth even a fraction of fraction of the damage if AI were to go bust as you suggested with trillions of dollars vanished.
So, with the data I have at the moment, I am unable to suggest that compute would be cheaper. Heck, it was cheaper before AI and compute prices have never been something that people worry about because there are sometimes 10x cheaper options than AWS,GCP,Azure with things like Hetzner/OVH and others (yes its not a 1:1 situation but still its a 95% overlap and for all intents and purposes, great)
I can see a potential where GPU compute can get cheaper, oh boy, its so much more expensive than compute but I feel like GPU's aside from AI might still have a much more limited niche than generic CPU.
The issue wasn't ever the pricing. Simon, I own 7$/yr vps's which run my websites fine because they are written in golang. I doubt it can get cheaper than it. (You can get a 3$/yr vps if that is what you are interested with using Nat VPS + cf tunnels)
I would once again appreciate to hear your thoughts on it. The only thing I realistically see is if Ram producers ramp up their productions and create a ram price glut in the next few years, but imo the prices would even out over the long term.
I have seen the point of spare DC capacity being raised up multiple times but I finally ended up writing a message which hopefully captures the nuance, but once again, I don't know the future about it.
Waiting for your reply and have a nice day Simon (& other readers) and thanks for reading if you did, I appreciate it :-D
I think we are in agreement that if the bubble bursts a lot of people will lose a lot of money. I don't have a strong opinion on the data centers, my main point is that I don't think AI "just goes away" if the bubble bursts, which seems to be something that a lot of people assume.
> my main point is that I don't think AI "just goes away" if the bubble bursts, which seems to be something that a lot of people assume.
i agree with your main point because the cat is out of the box with open source models and others in general. I don't particularly know the extent of what they would be used for.
The technology is still novel so people are trying out too many things with AI, I don't particularly like it being spearheaded into each and every thing but perhaps we are just in experimentation face and seeing what sticks and doesn't. Either way, I disagree with when people treat it more than a tool or is "the tool"
I do also think that the "cat is out of the box" and you are right that it isn't going to go away, particularly with open source models.
I think that there are some use cases where AI might make sense (prototyping or building things for yourself when all you want is the end result or thing which would be too time-consuming/complex to be built and thus wouldn't be worth making in first place for its use-cases)
So overall, yes I think that we are overall in an agreement. I do still believe though that learning the strong foundations.
But all in all I agree with you yes that AI might not be going anywhere, it can certainly have its benefits and hopefully the world uses it in beneficial way rather than negatively, glad we could come to an agreement. Have a nice day Simon.
Not to mention the competition: chinese open-weight models and open-source harnesses. Qwen3.6-(27B and 35B) have proven to be worthy and capable of running locally. I am confident more SMEs would look into this as a solution given the ballooning costs of API usage. You get a decent setup with an RTX 6000 Pro.
> 5% of every knowledge workers salary to go into tokens
In general, I don't think you can reason from the existence of potentially stranded investments back to revenue projections.
And when you frame this as percentage of salaries, that's a sneaky implication that this is only about reducing salaries and headcount, and not about adding capability, or doing things you couldn't do before, or making fewer mistakes, or capturing more revenue, or expanding margins, or competing more effectively.
That said, 5% of knowledge worker comp actually seems very low to me, given the capabilities, and considering the percentage of "knowledge work" that is absolute bullshit.
Two weeks ago I received an email from my HOA saying I'd been billed for a service I never asked for. So I replied to the email saying they'd made a mistake. There are now more than 30 messages in the thread, involving at least 8 "knowledge workers" at the property management company all passing the buck, and the problem is no closer to resolution.
An agent could wipe out all 8 of those bullshit jobs and solve my simple problem in five minutes instead of two weeks. Think of how many hundreds of thousands people are doing this nonsense just in the property management industry alone.
I could see such productivity gains being possible, if only because the current tooling around LLMs is terrible. The fact that we have 30 blog pieces per day making the front page of Hacker News about someone’s convoluted system to guide LLM output to something reasonable is absurd. There needs to be standardization in tooling, and it needs to be open source. Then, and only then IMO, will we see huge productivity gains.
But, at that point I think the big players’ moats will have dried up. Local models will probably be sufficient for 99% of daily office worker tasks.
So I disagree with TFA’s premise. I think this fear is probably shared amongst the LLM giants, and they’re still hoping that neural network transformers are somehow the path to AGI (probably not, imo).
Is it possible that you are narrowly sizing the opportunity? While PMF does not always mean that early pioneers will be the leaders, I think the market itself goes beyond knowledge workers and developers. Agents, robots, drones etc will all use LLM or some world model.
I am rather more concerned about competition from CHINA. With how Huawei (2000 -> 2020) crushed every other telecom company and went from nobody to the most revered leader in 20 years, and with the depth of leadership in manufacturing and work culture, if China surpasses USA in AI, all US companies lose.
"5% of every knowledge workers salary to go into tokens. 20% if you're a developer"
Not unreasonable. I'm a hardware developer, and my employer spends ~10% of my salary on software tools. Add hardware tools and their maintenance and it's more like 30%.
This is why 'agents' are the solution for these companies. Token spending goes through the roof. As long as a human is in the loop needing to read or review at human speed, that's a ceiling on how many tokens per user they can generate.
What value do the big model makers provide other than having a head start on gathering up humanity’s IP to train their proprietary models?
What’s their moat? Is it hoping for regulatory capture where scraping is made illegal the day after they finally finish scraping all human language?
It’s like OpenAI dammed the Colorado, and Anthropic dammed the Hudson, and now they’re both trying to sell us bottled water subscriptions at $100 a month. I don’t know how well the dam part of the analogy holds up, but the water part feels strong. Compiling models based on humanity’s written output feels like something no corporation should own.
This assumes that we won't need new hardware in ~2 years. I find that unlikely. So they have to make back what they got up until now PLUS the running upgrade/development costs. So what will it be in 5 years? $20t? $30t? It's all getting a bit outlandish.
What I'm often hearing though is the equivalent of "gg ez" when I bring that up. I don't understand how this will at any point blitz scale to profitability. As far as I know they don't have positive cash flow, no one has a moat and I don't think they will push out engineers.
To get that revenue and adoption they have to vastly increase their infrastructure spending. If they are currently losing in even the 200/month plans how is it sustainable?
> 20% if you're a developer.
That's a _huge_ shift. Most people I know cite +20%-40% velocity with these tools, against the actual work their company cares about doing. +20% speed for +20% spend isn't going to motivate a trillion dollars a year in spending.
Of course it will. The value of an employee is a multiple of what they get paid.
If you pay an employee $500k and they make $2M for your company (like Meta), then of course a 20% increase for the salary is justified if the velocity is increased 20% as well.
The difference between what the employer makes per employee and what they spend in compensation doesn't matter. If the increase in productivity isn't greater than the increase in cost, there isn't a reason to pay for AI over hiring more developers.
Imagine an employer with 10 employees paying $500k per employee and making $2M per employee in revenue (to use your numbers). They could hire two more employees and spend an extra $1M (+20%), but make an extra $4M in revenue (+20%). Alternatively, they could buy all ten employees a $100k AI subscription, for a total of $1M extra spending (+20%) but an extra $4M in revenue (+20%). You'll notice both scenarios are identical, so an employer optimizing for profit would have no reason to prefer one over the other.
There’s a lot relationship and culture management overhead involved when adding 2 more people to a 10 person company. I think any business leader would take the productivity speed up from buying a tool over hiring more people and integrating personalities/habits/viewpoints to an existing established culture any day of the week.
You're basically positing that the real cost of a 20% headcount increase is higher and/or the productivity gain is is lower than 20%. That isn't an unreasonable claim, but it's basically rejecting the premise here. You might just as well object to the premise that you can buy a 20% speedup by spending an extra 20% on tokens.
Also hardware will be obsolete or dead in 5 years, and warrantys are 3 years from Nvidia. Ask crypto miners how these kind of hardware economics work. Numbers have to keep going up all around. Its a fundamentally broken business model unless prices increase 10x
Also, with announcements of replacing developers with AI and consequent job losses, who is going to use the tokens? AI using its own tokens to produce code?
> 200m knowledge workers in the world, 30m developers. We're talking about a world where you need 5% of every knowledge workers salary to go into tokens. 20% if you're a developer.
This is where the napkin math is breaking down in a big way. There is absolutely no reason to assume this will only impact "knowledge workers". Farmers use computers. Farmers will use AI.
My hope is that hardware improvements (better node densities every 2-3 years, better designs, etc) will pick up the majority of the savings for these companies in the future, assuming LLM performance starts to taper off with diminishing returns.
It's worth noting that if each developer is 20% more productive with AI (let's take that as a premise and not dispute it), then it makes sense to go even further and reduce human headcount by more since the communication overhead of having 25% fewer developers is in and of itself a force multiplier.
tldr; 10 developers with 20% more 'productivity' can be replaced by 7.5 ideal developers and more like 6 or 7 developers due to the benefits of simply requiring less organizational communication.
I still think the ideal team size is unchanged however and that's 7-10 people. Note that teams aren't necessarily the same as direct reports. A CEO for instance has a certain number of reports and a leadership 'team' but they're not a team in the traditional sense since they are more about making good decisions and collaborating on specific things but mostly about leading their own orgs that have vastly different skillsets from eachother.
> They've got, ballpark, $5t to $10t to make back in the next 5 years, or the hardware buildouts will start getting written down.
I find it disappointing that a completely wrong statement like this ends up the top comment on HN.
It is wrong in both the math, the logic about public markets and understanding accounting.
> $5t to $10t to make back in the next 5 years
I don't know where this number comes from, but it has gone unchallenged.
OpenAI and Anthropic combined have raised around $100B. This is an investment so isn't something the have to "pay back" from earnings - instead investors expect to make that back from the share price being higher than what they paid for it.
> or the hardware buildouts will start getting written down.
The hardware buildouts get written down anyway!! That is a good thing for investors because as the value gets written down they can book a tax loss. ANd it turns out that generally agreed depreciation schedule for GPUs (used to be 3 years, now 5 years by places like Coreweave) is still too conservative since GPU rental prices for 5 year old chips are higher now than when they were new (!!)
All of this makes the rest of the math in the comment incorrect by at least an order of magnitude and under some scenarios possibly 2 orders of magnitude!
There is also the EV (expected value) of developing AGI. Even if you personally believe the probability is low within the lifetime of either of these companies, the value would still be extraordinarily high, enough to forgive a $5T or so miscalculation here or there.
The high-risk side of that bet would need to win more like a lifetime supply of Twix. But in a post-scarcity nirvana, everyone already has that. So sure, you're on at even money. See you in two years.
Theres no reason to believe, based on recent trends, that AI would lead us to a post-scarcity world, even if it could do all of our jobs better and cheaper.
I just don’t understand how people are getting negative value out of AI or even only 20% productivity boost. I can only conclude that people don’t know how to use agents.
I mean, it doesn’t really matter if it caused by people failing to use the agents well or not. You cannot assume everybody to use the technology the best way possible
Mostly brownfield systems in Java, Elixir and TS. I use OpenSpec in explore mode and point the agent to all the different repositories (when not working in a monorepo) to identify changes. Once done, i switch to propose mode and spend at least 15 minutes there iterating over the plan until I'm satisfied with the TDD approach (agents need tests to verify their work). Then apply and review. This also auto generates docs etc.
I guess I'm just not clued into your exotic definition of "survived" if continuing to function doesn't qualify. I tend to go by the dictionary definition.
Chapter 11 is not Chapter 7. Businesses survive chapter 11 bankruptcies all the time. For example, WeWork.
The company’s gone but the assets just got sold to other commercial real estate firms.
Uber was basically only ever software to help people use their own cars so a very small part of their valuation was physical stuff to upkeep, it was just deals and obligations they had.
Not sure how it shakes out for Anthropic and OpenAI. There’s a lot of physical capacity that needs to be built out and can depreciate. But there’s also a lot of network effects and dependencies being built in with enterprise users.
I don’t know how swappable the tooling is either. I think over the long term the UI, model training and documentation, and infrastructure are going to end up being run by different parties and I’m not sure which leg of that chain ends up in a position to skim most of the profit off. My guess is that Apple and Google end up raking in all the money since they control the OS and app stores while the rest of the stack gets driven down to being generic commodities. At least where mass market consumer adoption is concerned.
The difference is that they had room to charge more of their customers and pay less to their workers. The AI industry doesn't have both sides to play at this point. Training and inference are getting more expensive and if you take on the high prices now you're just floating yourself further downstream from profitability long term (which does not look viable for any of them currently).
> But then you sometimes go and talk to your senior engineering leaders and you’re saying, OK, how many projects that were on the cutting room floor got moved above the line because of the productivity gains because 25% of our code commits were via Claude Code last quarter?
> That link is not there yet, right? I think maybe implicitly there’s more that is getting shipped. But it’s very hard to draw a line between one of those stats and, OK, now we’re actually producing like 25% more useful consumer features, right? And that line is hard to draw.
That's pretty weak sauce. I don't think that justifies the headlines that came out of it, personally.
? What are you talking about mate? The man all but says "this shit does not work for us".
It iss layered in that careful, sanitised corporate shit-sandwich communication approach, where you take a nice piece of shit and layer it in between two slices of avocado so its sweeter to swallow for the "consumer" of your message.
He also said in that article that what prompted the discussion was the public statement by the Uber CTO that he had already burnt through his organisations yearly AI-budget in April. Please stop this shilling mate, and trying to hide the overall perspective between this or that word.
> The most discussed has been Uber, based on this report where CTO Praveen Neppalli Naga indicated that Uber had “maxed out its full year AI budget just a few months into 2026”, mostly thanks to Claude Code.
> Given that Claude Code only got really good in November it’s entirely unsurprising to me that a budget set in 2025 may have failed to predict demand for that tool in 2026!
lol I’m spending max $50/month right now on a couple light subscriptions and my velocity is insane right now (full stack mobile app development) I’m leaning into it hard while these cheap plans still exist and building out a big platform that I can easily generate new apps from. Hoping by the time the rug pulls I can just go back to hand cobbling these apps together from the modules I’ve pumped out and never even consider giving these companies a massive portion of my monthly income
My point with the $2,180.16 thing is that the price for consumers like myself is heavily discounted... but the price for enterprise companies is not discounted.
My usage is therefore a useful indicator of quite how much those enterprise companies may be spending on tokens, given the new pricing scheme.
If enterprise companies were still getting the same discounts that I get myself I would not have written this article.
(I had to dig into your margin figure - looks like you calculated 5.1% as 559000000 / 10900000000 * 100 but that $559M "adjusted operating profit" figure includes training costs, where usually when we talk about margin on inference we're not including those since those costs are fixed, margin calculations make more sense against the variable costs of serving a token.)
They key difference here is that training costs are fixed. If you train a model for $100m dollars, how much of that training fee should you allocate to each token that the model serves?
It's impossible to know, because you don't know how many tokens total will be served by that model until you retire it at some point in the future.
So you can't say "1,000,000 tokens costs $X in inference and $Y in training" because $Y is not possible to correctly calculate.
So, if you want to have a productive conversation about "margin on inference", it's sensible to look at the cost of serving the tokens independently of the cost of training the underlying model.
If, by some miracle, all LLMs ceased working right this second, any developer who would no longer be productive should not have been a developer in the first place.
You don’t need a miracle, if Anthropic API is down due to technical issues you don’t have software development anymore. It’s insane how much we are delegating to 3rd parties. It’s not like having cloudflare down where your users cannot access your services. The AI tools used to investigate prod issues stop working, developers stop working. The AI support system that allowed the company to get rid of their support team stops working. In addition to all the issues that causes to customer facing products based on AI. The sales team cannot work anymore.
It’s like the industry is willingly introducing a common external risk to everything
Limiting token quotas would be fine. Encourage developers to use efficient models, plan the work first, and to not burn thousands of GPU hours on waste.
It's much like when developers would waste tons of money on AWS spinning up massive test VMs and leaving them running without care. Until the finance people cracked down on it.
we all know it is impossible goal to make. surely AI will be even more useful in the future, but as long as china exists and continue to undercut the price, the goal will be never meet.
> We're talking about a world where you need 5% of every knowledge workers salary to go into tokens. 20% if you're a developer.
with that much money, the companies can easily buy their own hardware and hosting free public models, no need for those expensive subscriptions.
> +20% speed for +20% spend isn't going to motivate a trillion dollars a year in spending.
Except that if your company go 20% faster than the others companies, you win market shares. But then, everyone will use the same tools and companies will be at even speed, but the tool will stay.
Now...if the market is saturated, it's useless to try to do things faster. Cheaper yes, but not faster.
Pretty much all major tech companies today are horribly bloated and mostly metastasizing instead of innovating. I'm not sure how 20% increased productivity will help in any way with that. If anything, it might accelerate enshittification and turn potential customers off even more.
One thing I genuinely don't understand is these companies are constantly taking in incredibly large amounts of investments, so presumably they're giving up large chunks of equity or these are loans that need to be paid back or they're committing to spending obligations they're very unlikely to be able to meet.
So besides the insane hardware buildouts you're correctly mentioning, I don't understand how anyone that invests in these companies is supposed to make their money back in any sort of reasonable timeframe?
The cynical part of me is looking at what happened to the NASDAQ rules recently where essentially index funds are going to be forced to buy SpaceX shares much earlier than they previously would have (ie, before the price has a chance to reach it's real valuation). Which, um, I'm guessing these stocks are going to drop pretty hard when people start looking at the financials of these companies.
My suspicion is that the point of these IPOs is essentially to dump the bill on the unwilling public by forcing various institutions to buy it (ie, your 401k or pension is buying this shit), and maybe their investors can squeeze some money out of this before the stocks reach an equilibrium that's probably like 1/10th of what they're "valued" at.
> make developers 2x, 5x, 10x as productive on stuff that matters
What does this even mean? Is this about speed of development? Is this about headcount? LoC? How are coding agents contributing to productivity in places like GitHub, Shopify or Meta? I mean companies that already have an established product. I really wanna understand this because I'm not seeing that GitHub's product suddenly became so much better than it was 2 years ago, so where's all that productivity going?
The productivity is going into perverse incentives[1], e.g. we have improved (by which I mean "increased") token use. More PRs every day. More lines of code. All things we knew were shit-brained metrics a decade ago (obviously except token use).
We've also increased how much our coworkers need to read, or deal with. You can get an AI to make any point you want, so you can ignore the 5 humans raising alarms due to the 1 clanker you made say what you want to hear.
All numbers going up.
There are obviously people producing additional true value with it, probably, but that's almost certainly scarce.
...does anyone have a guess as to the total amount of money spent on software developer salaries each year? What percentage of that would the AI companies need to capture to be profitable?
(I'm not trying to imply that LLMs can replace software engineers, it's just an interesting comparison. If nothing else, I suspect that if the cost of development goes down, demand for custom software will go up.)
Bigger than that, they have to contend with open weight local inference. Open weight models right now haven't caught up to the frontier models of right now, but they're as good as the frontier models of not too long ago. If open weight models reach a certain point, then frontier model providers are going to struggle to make anything selling tokens, because eventually people will realize they don't need Mythos for everything.
I assume the bet is that as you swap humans for machines, this pays for itself. Swap entire devs and teams and frankly, managers, and you make up a lot of 5%’s fast.
If it works. And I’m not sure who is going to buy the stuff the machines produce, but shrug. Presumably some bots click ads for NFT’s that other bots generate.
I understand some startup deciding to take a punt on "this will all work out financially if our new product demonstrably boosts productivity of large sectors of the economy by a breathtaking factor that's incredibly rarely ever happened before in history: 2x.
Sometimes a plucky group of people take a risk, it pays off. If it doesn't work, the company fails.
What I do not understand is: large sectors of the economy all simultaneously taking this punt, with the necessary productivity boost, as you say, far more like: 2x, 5x, 10x
imo if your developers arent at least 2x as productive, then something is being done wrong on the employees part and/or the organization's. cli tools are ridiculously powerful provided you were an actual developer before using AI.
Maybe it's just me being (trigger warning from me providing an honest self assessment) very intelligent + a generalist, but i went from only full stack webdev and .NET to being able to implement an end-to-end LLM training pipeline (data curation, tokenizer, pretrain, sft, DPO - using ~$100 in cloud compute to train a class-competitive 1B STEM model)...and a full economic financial modeling and quant analysis application that pulls up to date economic, economic, news, stock data from the entire world and uses Dagster to orchestrate tech ical indicators and fundamentals and signals... and i did these things for learning and for fun. i built my own sublime text and obsidian replacement. i built my own reddit/twitter/hackernews/substack/news aggregator. i built countless other useful tools and utilities for me personally and for work I build more that empowers multiple departments.
Ive built 2 browser games, one already released to great reviews and 100k+ hours played. Ive built a tool on top of claude code that does ~60% of my job. Ive run data analysis on company financials for forecasting that have been refined and are producing very accurate predictions. Ive built competitive analysis tools and trackers.
All of this in 3 years. The projects are all clean, documented, with great code practices and modularity. A purist would surely consider some of the code slop. But it all works completely and fills real needs.
This is a huge shift. Anyone not realizing it yet is just simply behind the curve. I would not have accomplished 1/10 of this without AI coding. I went from copying code into and out of browser chats for 2 years before getting on the CLI train, and it is absolutely ridiculous the ROI you get from subscriptions to Claude or Codex.
they need to make 5t-10t back, but not necessarily through selling tokens. as we can see, the frontier labs are making vertically integrated products. their revenue is no longer strictly tied to inference.
That assuming once they start squeezing people won't just go to deepseek or other cheaper competition
> That's a _huge_ shift. Most people I know cite +20%-40% velocity with these tools, against the actual work their company cares about doing. +20% speed for +20% spend isn't going to motivate a trillion dollars a year in spending.
And most research shows people far over-estimating their own gains. Once companies start counting the actual (and not just reported) gains, the AI budgets will be more limited as people realize it's an useful and versatile additon but not replacement for most types of work
> We're not there yet. This is still the upswing of the hype cycle, and unless we figure out how to make developers 2x, 5x, 10x as productive on stuff that matters, this isn't going to play out well.
Upswing of the hype cycle while growth of tech itself is flattening, both coz of techs innate issues (which might or might not be solved, but some papers claim they are unsolvable with current approach) and just the fact the spike in growth caused so high economy cost that it put brakes on itself.
There's a lot of workslop pumping the numbers. People can generate a 300 page PDF in a tiny fraction of the time it would have taken, but now the report is full of mistakes and fluff, and the stuff that would have been learned and caught in the process of making the report is now not happening.
I asked claude to generate a frontend and it made the same template. Same san serif and serif fonts together. Same colors. Same typography. Same layout and animations even. It’s wild how similar it is. No not similar it’s the same damn thing.
It produces the "most average" web design unless you really prompt your way out, isn't it? If you don't care enough to prompt, Claude does not care to be individual.
I don’t think these numbers are accurate? It seems to ignore the fact that the models have cache for ongoing sessions, which means you (normally) aren’t actually sending all those tokens on every request… you only need to if you go too long between requests.
It's going to be a typical saturation curve. A lot of upfront tokens spent on things that have stockpiled over the years, and then the derivative on token spend trends to zero as the users run out of immediate things to try. Sure there will be ongoing maintenance and experiments, but it wont be nearly as close as the initial inrush.
Given what costs are and availability of parts, that 5 year write down is not in practice going to be the case. Maybe tax wise perhaps but especially for big fancy expensive multi million dollar 100-500kW racks these things are going to stick around for a while, I think.
consider cloud spending vs on-prem before the great cloud migrations. people are spending a lot more for cloud services now.
I hear conflicting things about finances, some have a different opinion, that it won't be written down so long as more funding comes in and revenue keeps increasing. it isn't like how you take mortgage or business loan, it isn't even a loan it's an investment funded by loans. So long as the investment is still promising, what are they going to do? destroy its value by calling in trillion dollar loans?
> Gartner forecasts that large AI companies would need to earn cumulatively close to $7 trillion in AI-driven revenue through 2029, which is close to $2 trillion per year by the end of the period. In order to achieve “historic returns,” the providers would need to earn nearly $8.2 trillion in the same period.
The numbers are made up political correctness anyway.
Everyone's agency is 100% captured by belief in Wall Street. Too few <50 have any meaningful labor skills to blink.
We'll continue to have consent manufactured via media platforms and in 3 years no one will bat an eye at these companies being worth $12 trillion as Altman and Musk climb two ladders holding a "mission accomplished" banner.
One quick question. Did tax payer money fund these data centers? If so, how does that money translate to their profit and a return for the people whose work paid for the resources?
Source on 200 million knowledge workers worldwide? My understanding is that it's just above 1 billion. I dont think a billion subscriptions at $1000/yr is out of the question but it might take a decade to get roiling
You're suggesting that 1 in 8 people worldwide, including every one from infants and the elderly, are knowledge workers. Are you sure that's what you mean?
I'm not even sure that 1 in 8 people I know would qualify as a knowledge worker, let alone a knowledge worker that might profoundly benefit from on-the-horizon AI. And I'm in a highly skewed population.
Personally, I've only encountered any of those people on line, and almost exclusively here on HN.
Most people I've met -- and again, in a pretty darn skewed sample globally -- see $65/mo as a lot of money to spend on technology of any kind and can't think of anything much they need from "a personal knowledge worker in their pocket". I don't know a single person in real life who remains excited about AI at all, and only a few software engineers who feel it'd be worth that much.
Everybody seems to be mostly confident with the "knowledge productivity" in their personal and professional life and a pretty skittish about spending in today's economy. Most would be excited about a magic new robot that affordably saved them from unwanted physical labor and drudgery, but nobody needs much real help making appointments or filling out forms or whatever.
That's not to say I won't be proved wrong some day, with some further innovations in AI products, but global-scale demand isn't waiting for anything that's been released so far.
I've yet to meet a person that fits that description IRL. Admittedly I don't live in the valley but I do work in tech. The only place I see that demand is on hacker news (and I imagine twitter - I'm not on it).
The competitors of $65/mo subscriptions are the free models and services that are good enough. It will only get worse as open models or free tiers catch up. For most people, they just use whatever that's free
Apple TV, Netflix, BritBox and PBS add up to about $45 a month. Most people are gonna judge AI up against what they’re already paying for and the AI model makers simply don’t have a good enough product.
There’s only two things useful to the average person something to help them translate and something to help them write in everyday life.
Something else that might be useful would be local single purpose AI agents who’s remit is to help you with one specific task, but I don’t think that’s what the people building those expensive data centers want to sell to the market.
> Apple TV, Netflix, BritBox and PBS add up to about $45 a month. Most people are gonna judge AI up against what they’re already paying for and the AI model makers simply don’t have a good enough product.
What's the actual TAM for premium tv subscriptions though? Unlike free models which are keep improving, you can't get premium tv for free. Also they are actually competing with the total subscription price a person is willing to pay for a year.
> There’s only two things useful to the average person something to help them translate and something to help them write in everyday life.
Exactly, free models are good enough for these tasks.
I just don't see how b2c is a large enough market to ask average Joe to cough up hundreds per year when there are free stuff everywhere. B2b is another story
85% of the world population lives outside of developed nations.
27% of the world's workforce is in agriculture (contrast to the US where it is 1-2%). 15% in manufacturing.
A lot of people work in "services" (especially in high income nations, where it's roughly three quarters) and some of those are knowledge workers... but a huge number of them are nail technicians or hairdressers or bartenders (etc etc).
Here's a source from 2019 that says: "By 2023, the number of knowledge workers in the world will increase to 1.14 billion, with more than four-fifths of that growth coming from the emerging world."
To add an actual source to this thread, a brief paper by researchers at the International Labour Organization (ILO) states that for knowledge workers globally "... there
are between 644 and 997 million jobs, which represents
between 19.6 per cent and 30.4 per cent of global
employment respectively." [1]
[1]: Berg, Janine and Gmyrek, Pawel, Automation Hits the Knowledge Worker: ChatGPT and the Future of Work (April 21, 2023). UN Multi-Stakeholder Forum on Science, Technology and Innovation for the SDGs (STI Forum) 2023, Available at SSRN: https://ssrn.com/abstract=4458221
Globally, sure. The assumption here is all users are on the same economic footing, they are not. Only about a 1/3rd (at most) of that count can afford $1000+ monthly, and even then that is wildly out of line with what most will.
I googled "number of knowledge workers worldwide" and read the top results. If you read it as I was confident in a billion I apologize, Im just trying to get an accurate count. What numbers do you have now and where did you find them?
That's not the TAM of 1B knowledge workers globally. If that were the case many industries would have a 2-3x target market.
To simplify break that 1B up into 3 levels of purchasing:
1) High-tier (US, Western EU, ANZ, Japan, South Korea, Singapore, UAE, etc) - 200-250M knowledge workers.
2) Mid-tier (Eastern EU, Latin America, urban China, India tech sector, etc) - 300-400M
3) Low-tier (Rest of the world) - 300-400M
Low-tier users are mostly free tier or heavily subsidized pricing.
Mid-tier are going to account for USD sub-$100 tiers. Probably averaging less than $50/seat.
High-tier are who you are assuming is the 1B. Users are not equal in that knowledge worker count, so there aren't 1B knowledge workers to charge money.
And when you consider Low-tier users a majority of those are free users which need to be subsidized by the High-tier users. So either free tiers get much more restrictive or the providers lose additional training data. A bulk of Low-tier users cost money and provide little to no revenue.
Edit:
And think about Mid-tier and Low-tier for 5 seconds. Why would they pay Anthropic or OAI when they get get 100x+ inference from DeepSeek or Xiaomi? Mid-tier may be the only area that is willing to spend money on a US provider, but I would wager significantly on the fact that users in the Low-tier almost universally do not care.
Thank you. So with these numbers it seems like half a billion subscriptions at $500/yr is on the table. Obviously theres going to be competition in this market and self hosting cheap models may become the dominant use case. Assuming the labs are able to get most of the market though, the market size is something like a quarter trillion a year within the next decade. It's hard for me to imagine the whole sector failing if that happens.
I do think free accounts are going to end pretty soon, and some of the workers in your tier 3 will pay, but even without them this seems like a pretty healthy market size. I also wouldnt be surprised if mid tier workers are able to afford the $1000/yr vs $500. I use yearly rates because I find it easier to compare them to GDP/salary numbers
I mean, sure. Assume all you want but to guess that the entirety of High-tier plus almost all of the Mid-Tier will spend, on average $500 per annum is bonkers.
I believe we've started to see the top of what individuals and businesses are willing to pay for the current model capabilities. We are nowhere near AGI and models are really only providing significant value in niche markets currently (programming and cybersecurity). And just like SaaS the enterprise has the option to buy hardware and leverage their own models at will which can potentially offset costs and TAM as well. I have talked to a number of large financial corporations in the last 6 months and most have internal initiatives. The same applies in the healthcare vertical.
$250B per annum with AI? That's 20% of global software spend now. Sure, that's possible but that assumes current market prices hold. What if inference ends up normalizing between DeepSeek/Xiaomi & Anthropic/OAI? There's 50% of your revenue and with current costs for inference and training in the US at astronomical levels the US AI industry could also very well be setup to implode overnight.
Lastly I don't believe free can go away anytime soon because it can't. As soon as Anthropic and OAI remove that option those users will move to whatever is. For most of those users it's not a luxury to choose, it is the only option.
The financial engineering occuring right now is something I don't doubt will be text book lessons of the future. We've seen it before and I believe Peter Sorkin when he says that we will see a crash of this bubble, it's just a matter of how catastrophic it ends up being.
A lot of those ‘edge cases’ in the definition of “knowledge worker” are probably the stuff that’s most likely to have significant parts of the work augmented or replaced by AI agents. Like, call-centers are almost certainly going to get turned over in a big way. It’s not like the median tier-1 support operator just reading off a script is much better than an LLM anyway.
> +20% speed for +20% spend isn't going to motivate a trillion dollars a year in spending.
I'm increasingly realizing this math is wrong, because LLM use is really sticky.
If Anthropic 100x'd prices tomorrow for their best model, so some companies offered 50% salary to keep 100% of your AI usage:
a) There are programmers who would take this deal. They've gotten to the point of doing what feels like even less than 50% of the work, developers were already pretty well paid, so they'll take it.
b) There are companies that'd offer this deal. Even if the only people who are taking this deal are not the best engineers, and the AI output is not the greatest, I think the last 6 or so years have seen a lot of companies realize capitalism is not as competitive as it seems.
They're not worried about putting out a worse product because... frankly, what else are you going to do? CF lay a bunch of people off, support gets awful: well you're probably not building a new Cloudflare in the next few years.
In the meantime the AI will get incrementally better, their market share will grow, and you won't be able to compete without taking the same faustian bargain.
-
Maybe I was just naive but it's making me realize how much we take for granted in the world. Both the quality and relative value of things don't have to go up over time. Quality can go down while prices go up, and nothing will really stop it. Competition should stop it, but competition is really slow and can be interfered with. And as prices go up competition gets really hard.
And that's not considering that capitalism is going to do what it does best: if they really found a way to be profitable, competitors are going to fight them on pricing. Anthropic, OpenAI, Google, etcetera 's margins are a competitors' opportunities.
It's not as if there weren't chinese models nearly SOTA. Don't know where the french (Mistral) are but they may try to get in the game if there's a way to be profitable (not that France or the EU for that matter are relevant in anything tech or had any tech company besides ASML and SAP in the Top 100 but who knows).
Your severely underestimating the idea that people are just not going to use developers for certain things in the future
For example I don’t anticipate somebody making a living off of making website ever again
Somebody with absolutely no technical experience who needs a website for their business can now make one with almost no money whatsoever.
That’s good enough for their business. and the code can be totally shit and it does not matter because it’s meeting their business objectives. I am seeing this in the wild and I’m paying money to companies that have these types of websites and because it doesn’t matter I don’t need for the website to work perfectly on all my devices all I need to be able to do is pay them through the website which is what they need me to do and our transaction is done.
Don’t forget ultimately the people who pay technologists right now are primarily advertisers
work on hard problems is going to continue to be some tiny fraction percentage of the software engineering discipline
just expect a total bloodbath because the goal isn’t developer productivity the goal is that “I don’t need to pay somebody $200,000 a year to build a website authoring tool like WordPress.”
Why would a small business use a coding agent to build a custom website when they could use something like Squarespace or Shopify with prebuilt templates that mean they have to know even less than if they were to use some kind of chat UI?
This is never going to materialise. It’s dead in under 2 years.
The market is shrinking and saturated already and it’s not because of AI gains but geopolitical instability and supply chain issues, some of which are caused by AI spending and stupid ass PE firms refocusing on AI supply chains.
Wait do you have any numbers to back this up? Every number that I've seen contradicts this. Most sectors have positive revenue growth, even non tech sectors. Technology purchasing is increasing in every bucket (software, IT services, devices, communications, and of course DCs). Retail and food-service sales are up MoM and YoY. Personal consumption is up 0.2% in real terms. I assume by service models you're just talking about AI? I actually may agree with you but this is clearly not true for long if it is true today.
I'm reminded of that [terrifying in hind-sight] Newt Gingrich interview in which he was more concerned about his constituents feelings about things getting worse than any silly statistics provided by government agencies.
I feel like there's a bit of AI psychosis in this particular post.
>"These are tools which burn vastly more tokens, but are also quickly becoming daily drivers for the work carried out by extremely well-compensated professionals."
>"Somehow this fragment turned into headlines like Uber’s COO says it’s getting harder to justify the money spent on AI tokenmaxxing, because the market for stories about AI failures remains enormous."
Yes, it's just the yearning for AI failures. It couldn't possibly be runaway costs, record revenues, and massive layoffs. It couldn't possibly be that these tools are lighting dollars on fire by people already paid significantly well and not producing any increase in "value" for it (I recognize that output is 100x but outcomes are flat by all measures).
Being pedantic, but I don't want to lose the meaning of the term: "AI psychosis" doesn't refer to someone who thinks AI is really good. It refers to someone who develops symptoms of psychosis from talking to an LLM, e.g. believing they have developed a new Grand Unified Theory of physics.
I don't know, "workaday professionals will find $200/month a particularly good deal, such that there will be widespread adoption" sounds either credulous enough to support the diagnosis or dishonest enough to dismiss. I am a "knowledge worker" who is doin' fine, has a lot of templated written work/report writing, and there is no way in hell I am justifying that kind of spending to my boss or my family.
"I firmly believe this technology will create business value" is so obviously and categorically different from "Humanity has birthed a silicon god that I have also developed romantic feelings for" that I'm not sure if your comment is even trying to be in good faith
What sort of bad faith would even apply here? idgaf if poster x or y has psychosis or not. "$200 a month is classic addict behavior" seems pretty spot-on to me though, I just don't want to have to pay it, too.
I think is the classic dilemma where people don’t know how to value their time.
Typical tech worker costs a company around $100/hour minimum. That $200 subscription cost can look mighty attractive if it saves some time or mental load.
I don’t think there is anything about addiction or spooky with that math. I suspect a lot of this is coming from tokenmaxxing firms but on the flip side on our small team, we end up spending about $200 per person per month for tokens using tools like Cursor. We feel the spend is justified with measurable value.
You can't justify $200/month in spending to your boss? Many people charge more than that per single billable hour. I would put your salary side by side with that number, which is your boss's perspective, and reconsider.
Fully loaded costs for an average employee at a bigco are scary. Not $200 but significantly higher than the number on your W2 by the time the company pays vacation, benefits, unemployment insurance, etc.
I've seen people try to argue for resources using reasonable but abstract arguments, and it just never works. The fact of the matter is that I want $200/month, on top of my other asks, and that comes out of somebody's budget. I just don't see folks snapping their fingers and $200 a month (discounted for now!) appears. Good for the other guy I was replying to if that's the case! I just don't see it though.
I just cannot imagine zero net productivity gain from AI usage by a somehow experienced developer. Even if it’s happening right now it’s matter of time for people to figure out how to utilize it right. Only exception would be if people are getting the same output but working less, which is still a net positive (just not in terms of output).
Sometimes it feels like theres this opposite AI psychosis, where anything AI is bad and boils the ocean, takes our jobs and makes RAM expensive. Its a component in the current economy, but things like tariffs, closing the strait of hormuz etc is equally bad for the economy. Anyway, just find it strange to be so militantly anti a certain tech.
It's more the tech leaders than the internet. Sam Altman, Jeff Bezos, Eric Schmidt and such get up on stage or interview regularly with a shit eating grin telling us all about how they are coming for our jobs, will make us obsolete, and there is nothing you can do about it.
It's a natural response for society to despise these people who have such contempt for us. It's almost embarrassing these days being at a social function and telling people I work in software, it's got a negative stigma almost like working in gambling or the military.
I think it's a matter of perception because I didn't interpret any of them as being gleeful about it. If you think about it, "AI will take your jobs and maybe destroy the world" is horrible, horrible marketing -- like, your comment is a perfect illustration of how it is received everywhere -- and yet these CEBros can't stop saying it, which indicates to me that they actually believe it.
Oh, now that their IPOs are nigh they're changing their tunes (https://archive.md/s9EO3) but to me that looks more like they've decided to let $$$ prospects override what they really think.
The general public is not their customer base, they don't have nearly enough money to spend on AI. Going to the media and saying "this product can automate so many jobs" is marketing to other businesses who want to use it to cut their workers out.
There was crazy clip of Eric from Google telling a crowd of university students that in the future AI will do everything, and after the whole audience boos him he keeps pushing the point that they better accept it and get on board. The mentality these guys have is sickening. They have no humility and no humanity.
>Anyway, just find it strange to be so militantly anti a certain tech.
You know, that's fair. I'm much more against super-rich investing hundreds of billions in the things they don't understand, creating massive disruptions in their wake.
AI didn't create stupidity and greed. In the end, it's just another tech. I'm just tired, time and again, of people who I would hope to know better (and repeatedly they prove me wrong).
Hi Simon; while it's true that the token providers have found that their product is 90% useful to devs and 10% useful to everyone else, this is something they found out in the first quarter of 2025 anyway.
It's not exactly news, is what I'm saying. And even with the PMF they found, the product is still only a commodity i.e `tokens`, which is what every other provider on the planet is also providing.
All their other products boil down to "harnesses", which does not look viable as a product in the sense of PMF - you cannot sell it, you cannot lock it to your own subscription, API, etc. so you can't use it to generate revenue any more than the free harnesses do.
PMF has a specific meaning, and "code harness" or "coding model" does not satisfy the commonly accepted meaning. Maybe Mythos (or similar) will.
Who’s an anti-human extremest? Doesn’t that sort of warrant some sort of facts to justify such an extreme position? What makes you feel this is a misallocation of capital OR something…anti human?
I find this analysis confusing. PMF for coding was likely reached some time last year. Profitability, which is different, we don’t know. The article kind of confuses both without making a strong economic case or using numbers in a compelling way. I don’t understand what the Uber case has to do with this either. The Uber COO clearly said that at least in terms of ROI he’s not seeing the results either.
My take is the product has been very useful for coding (PMF) for months. But it’s certainly not useful at any cost…
What I also find confusing though is that folks seem to ignore trajectory which is maybe the biggest lede to bury. As Simon says, we have had "good enough" coding agents for 6 months, that is a blink of an eye, and at my company my job has now completely changed. It's almost like a dream.
And that's just one inflection point. We've had several and there are many more on the horizon. So while I could be convinced that ROI is maybe not even positive today despite the ridiculous enterprise spend, it's perfectly rational to pave the way today for what's coming over the next few months let alone years down the line.
There may be additional major leaps forward, and there may not. I kind of struggle to imagine what the next step actually is. Certainly there will be improvements in performance (speed) and cost. But at a point you reach a barrier where the limiting factor is the specificity of the human prompt and our ability to manage all the code we’re generating.
Somewhat oversimplifying; writing software and building apps was a bottleneck - now it is not. What is the next bottleneck that LLMs can solve? Is there one? And is there enough publicly available data to solve it repeatably at scale? Or did we just automate stack overflow searches and now we’re stuck again?
Or is the endgame of this innovation cycle the complete removal of interaction with machines through code? Will we simply interact with machine coworkers purely through natural language? Can an LLM make PowerPoint slides and run a meeting? So far not seeing much progress on that.
Judging from the fact that the Opus 4.5 inflection point was not really anticipated, and we still don’t really know what threshold was crossed that suddenly made agentic coding accessible to so many more people, I think it’s safe to say we don’t know what the thresholds will be until they’re crossed. The fact that we don’t know exactly what they’ll be isn’t a good reason to think there won’t be any more.
I am currently eating lunch. Meanwhile Claude is triaging and writing reproducers for 70+ tickets nobody has had time to look at. Next it will attempt to fix them. I have not read the tickets. I will not look at the code until there are review ready PRs and a code review bot have done the first pass.
In other words, most of the prompting will also go away.
Pmf is this weirdly defined thing where "if you're not sure you have it then you don't".
I think it was clearly useful for months to people who had tried it and taken the time to understand it, but now that knowledge has spread to the point where wallet holders are convinced it's not just passing fad or hype so now pmf can be "claimed".
I agree it's weird to say "those people have pmf" though, usually it's something you define for yourself
> Pmf is this weirdly defined thing where "if you're not sure you have it then you don't".
I'm not sure if this runs counter to your point or not, but: I don't see any future where LLMs aren't a core part of Software Engineering. The horse is out of the barn. There is no going back.
Yeah but the product is not “LLM” it’s “proprietary frontier model LLM paid by the token”.
And I don’t even necessarily disagree with OP! It’s more like the competition is shifting so quickly that your competitors could undercut your PMF in a blink of an eye.
History bears out that cheap and satisficing soundly beats expensive and optimal every time. Until we have smarter and more prescient decision makers in leadership, the bottleneck on output will be the quality of decision making not the quality of code. Trying more things faster and cheaper will win.
> clearly useful for people who took the time to understand it
people -> programmers, I haven’t met a non-developer who reports getting more time out of current AI platforms than they put in. If anything I’ve anecdotally heard the opposite, introducing AI at work creates so much slop (output) it takes more time to process it all without a tangible bump in overall productivity
The article also treats the word "good" as load-bearing in a way that should have you questioning their analysis:
"I’ve called November 2025 the November inflection point because that was when GPT-5.1 and Opus 4.5, combined with their respective coding agent harnesses, got good—good enough that we’ve spent the last six months adapting to agent systems that can reliably get useful work done."
MongoDB was once backed up by adoption across the industry. Or for a more recent example, blockchain took off like wildfire across the industry before ultimately fizzling out in all but the most niche applications.
Not saying this trend will do the same, just that the industry adopting something doesn't guarantee its success.
I don’t think those are really comparable. The blockchain was trendy hype, relatively few companies actually adopted it. Where did Netflix use the blockchain? Google?
By comparison almost all tech companies I know have leaned heavily into AI.
Thats why most here shouldn’t engage in the discussion - they parrot on about benefits without identifying and articulating the costs and moreover how it affects the firms financial position.
It’s not supposed to be logical, it’s an LLM evangelism blog that rarely, if ever, has any critical analysis that isn’t pro-industry. Read any/all of the other posts and you won’t find much skepticism but you will find a lot of shilling how great it all is.
I like his other posts. He's bullish on AI, which is fine. I'd like to read a mix of bearish and bullish level-headed takes from people who are subject matter experts. His technical credentials are well past discussion - I love Django, and he comes across as a pretty upbeat but level-headed guy. Certainly beats radical takes in either direction from people who have no clue what they're talking about. It's just this article that I find rather confusing.
That's how I feel about most of your writing. I click through most times when I see you either on the front page or in the comments, and I generally walk away feeling like I have food for thought, without necessarily buying everything wholesale. It's part of why I keep coming back.
My root comment simply represented my two cents about the current post. I don't think anything about the post is outrageously incorrect or anything, just somewhat confusing. You're a very prolific contributor in this community and I don't think me or anyone else that welcomes your takes expects everything you write to rock our collective socks every single time, anyway.
If you want an "LLM evangelism blog that rarely, if ever, has any critical analysis that isn’t pro-industry" there are plenty out there. I'm not one of them.
People are confusing "excitement" with "evangelism". Your blog is definitely on the pro-AI side of things, but as you say, it's not one-sided or uncritical.
All of these are about AI misuse, not skepticism of AI. By skepticism I mean doubting whether AI actually delivers on its promises which, based on this last post, sounds like something you think we're already past.
Many people still think AI coding agents are slop on steroids despite all the current hype around AI actually shipping functional products.
It's hard for me to write about skepticism that coding agents deliver on their promises when I've been using them daily and know, for an absolute fact, that they boost my own productivity.
(And that's after taking into account the METR paper that says engineers over-estimate their productivity with these tools.)
I have plenty of doubts about AI delivering on its promises outside of coding. I don't write about AGI because I think it's science-fiction hysteria. I write about slop precisely because it represents a mis-use of AI that demonstrates people completely misunderstanding what it's useful for.
Love when people say "its promises". What specifically are you disappointed with? Simon's posts are high quality and evidence driven. AI has already delivered an incredible amount. Read Epoch for industry trends and analyses, METR to, everything points to a pretty consistent picture.
"Many people still think AI coding agents are slop on steroids despite all the current hype around AI actually shipping functional products."
Oh yes, tons and tons, especially on HN. But the plural of anecdote is not data. Enterprise spend speaks for itself. You are using AI-coded functional products all the time. Do you want like a diff history for the Google codebase or something?
Tbf the OPs blog and comments (including their sibling to your comment) are also heavily anecdotal.
> I’ve called November 2025 the November inflection point because that was when GPT-5.1 and Opus 4.5, combined with their respective coding agent harnesses, got good—good enough that we’ve spent the last six months adapting to agent systems that can reliably get useful work done.
Claiming a grand inflection point based on your own personal usage is very anecdotal.
If that were it I would absolutely agree with you. But this experience maps exactly to adoption trends. My job in the last 6 months has become so unrecognizeable to me it’s insane, the adoption at the very least at large companies is truly truly incredible, and it really does coincide with the quality of opus 4.5 (which has now been surpassed).
"Adoption trends" are just herd behavior which may or may not be driven by compelling anecdotes and may or may not be evidence of something more. I'm just saying it seems wrong to dismiss the post the way you did when the OP in question and your own post here are just more anecdotes.
No, if that were really true you wouldn’t see what you’re seeing today. You wouldn’t see entire companies completely retooled and refactored around these tools. You would see the mistake of “this is actually just herd behavior”, which involves such a colossal amount of impact to these companies entire stack and bottom line, resulting in systemic collapse. You don’t see that. Company leadership are not some idiot class of people, I don’t know why this is people’s prior. If companies get adoption wrong in either direction they are completely screwed. So you’re seeing people putting money where their mouth is, across the board.
Compelling anecdotes are not even the main source of evidence. Look at the enormous body of work on measurement of these systems. I always point people to epoch capability index as a good summary statistic of capabilities or METRs time horizon data which has now been topped out. They had a recent updated to the dataset, after which the corrected plots pointed to an even faster acceleration than before.
My point was claiming a broad inflection point based on your own personal usage is not "evidence driven", it's anecdote-driven. It's hard to disprove any claim you made because you didn't really make one that's disprovable, and your opinion on it now is still just an opinion.
Those have been around for a long long time, you may be focusing on anecdotes but the adoption numbers and performance trends speak for themselves and we’ve had performance trends for years. People can argue about whether or not enterprise level adoption has a clear ROI today but the fact that we’re at the point where entire large scale companies are already completely refactored, directly after opus 4.5, if that’s not a convincing enough signal I don’t know what is.
So how do openai and anthropic plan to keep customers when GLM-5.1 is just as good and open source and a lot cheaper?
I don't see the business model working. My closest friend actually does automation software for large companies.
He does not use Claude or openai at all. He primarily uses gpt 120b on cerebras and glm-5.1 for heavy thinking work.
And some other small models for various tasks. All open source.
And these systems are extremely useful for the businesses and are able to run fully automated pipelines that are very stable and fast.
We discuss this a lot, and we both think any business doing heavy agentic work on Claude and openai just aren't aware of exactly how good and cheap open source has gotten on the last year.
So... once the legacy businesses and developers catch up, won't Claude and openai be unable to recoup their costs?
Same. It's a nightmare from a Porter's Five Forces perspective.
There will be a ton of businesses competing in this space, and there will be something of a moat due to how capital intensive the business can be, but there will still basically be infinite competitors.
For coding you always want to go with the best model in the category, not something that would be the best model if we went 1 year back which GLM 5.1 is, and I'm saying that as a big fan of GLM cause I run a translation site where GLM is good enough for the price.
Most of the money right now is in coding. Openai and Anthropic just have to be 6 months ahead of SOTA open source models and they'll capture most of the enterprise and dev market
Yes I'm an engineer (20 years most in games/graphics industry) and only use it for code. I've been using glm 5.1 this week a lot. I went in expecting another "decent" but not really "up to standard" open source model.
I highly doubt I'll ever use Claude again.
I think you are wrong about Claude being any significant level better
I've been mostly coding with GLM-5.1 as well and I agree with you. DeepSeek V4 Flash is another very good surprise. Incredibly cheap, fast and effective.
Well I think there are a multitude of harder measurements that would disagree with you, but ultimately there is absolutely a use case for cheaper open models (or even cheaper tiers of proprietary models) and in fact the unsolved optimization everyone is trying to get to is how much spend to use for a given task. But there will always be a market, especially in enterprise, for the best performance there is to offer
I strongly disagree. I'm an engineer - I'm all about the fastest, cheapest thing that meets the requirements. I don't need Opus 4.7, even for my complex programming tasks. It costs over 10x other models available that still give good enough answers. Those smaller models are also a lot faster to output tokens, which saves me time.
Once the model gets good enough, the returns on bigger models diminishes quickly. I don't want to spend 10x the money and wait 5x the time to get answers that are equivalent.
I have stats from a harness that tells me glm5.1 is far more cost effective for us than Opus with the rate of defects and rework taken into account. In fact, with a decent harness I'm now increasingly favouring eHaiku over Opus for execution too. Opus is still worth it for planning, though, and far better at one-shotting things.
If I generate code with Claude, ChatGPT, and GLM 5.1, I can't say which model is which reliably. I exclusively use Claude more out of superstition than reason.
> For coding you always want to go with the best model in the category [..]
And this is why many companies go out of business. You always want the best bang for your buck, sometimes this is the "best model" and sometimes it is not.
It's driven a lot by the harness too. If you're using claude code, you're actively being pushed towards newer models, even though older ones work perfectly fine for your use cases
Cost for the value delivered. Like if you offered the current SOTA open source models at $0.1/M, I still think I'd be using Opus or 5.5 at $30/M. Or say GPT 5 which was released Aug 25, I don't think I'd use it for coding for even $0.1. I'd def find other uses for it(translations, agentic workflows, prompt guards etc), but for coding I don't think I'd ever completely switch to a SOTA open model
Unless ofc there was an actual speed difference, only reason I'd be willing to go with a worse model couple of percent worse than current best model is if the speed was at least 5x higher. Looking forward to kimi k2.6 offered publicly by Cerebras
Based on current market for LLMs I'd say my use of "you" in the general is fine. Even openrouter which doesn't capture all of the SOTA closed models but nearly all of opensource model usage has Opus as 1st(on last week) on "Programming" category and 3rd in overall rankings
I'd trust the OpenRouter rankings a lot more if they exposed the number of unique users for each model, as opposed to just a token count.
Currently I have no way of telling if big changes in their rankings are caused by a single "whale" switching providers, or if it's a more meaningful trend.
My point was that even openrouter, the one place people who are looking for open source SOTA models go to, doesn't definitively have opensource models at the top. Esp considering quite a lot of the closed models usage is through AWS, GCP , Azure etc, probably dwarfing the usage on openrouter by a huge factor
> For coding you always want to go with the best model in the category
This is transparently false, because the best "model" is still competent human developers. They're just more expensive. If you're willing to use current LLMs at all, it means you're willing to sacrifice quality for a better price, and your disagreement with the comment you were replying to is entirely about what the optimum tradeoff is.
Maybe, but that's a different claim than the one I was responding to. And also raises the question of "if the lower quality but cheaper output of frontier models is more cost effective than humans, is the even lower quality but even cheaper output of OSS models is more cost effective still?" With an absolute rule like GP suggested ("no, you always want the best code generator") the answer is clear, but it get much murkier if you reject such rules (as you have to to be an LLM coding proponent)
It was true 6 months ago, not anymore. Frontier models now outperform developers on many tasks, be it on quality/readability/maintainability, and let’s not talk about speed…
I've seen the code they produce without extensive help from human developers, this is clearly false.
Good to see the classic "yeah the models weren't good enough six months ago, but this time they actually are, promise! Please forget you were hearing the exact same thing six months ago!" is alive and well though.
>For coding you always want to go with the best model in the category, not something that would be the best model if we went 1 year back which GLM 5.1 is, and I'm saying that as a big fan of GLM cause I run a translation site where GLM is good enough for the price.
Currently, the difference is substantial, but what happens if capabilities saturate?
Then the house of cards comes crumbling down, but there is so much evidence to point to this not happening that it requires a bit of a theory for how that may happen
Well I think there are several fairly stable trends that paint a pretty compelling picture:
- performance scales with compute very very reliably. We have “scaling laws” (and have for years) and they are almost miraculously stable and show no sign of being invalidated at all even at the very largest scales. There are some theoretical bases for this though I’m not as familiar with the details
- these scaling laws are on an unintuitive quantity (validation loss on pretraining datasets), so we can look at downstream performance. Benchmarks are a minefield of junk but there are many decent ones and enough variety of techniques and data sources and scoring methods etc that in aggregate they are useful. The single number that I think is the best summary statistic across the crazy (O(100k)) number of benchmarks is the “epoch capability index” (just some branding over a reasonably standard statistical model that was really well thought out and a great idea). The trends in this are extremely stable. Eyeballing the trend over time on their graph we’re getting basically a GPT-4 to GPT-5 level capability improvement every ~18 months
- coding agents are not limited by the quality of the human training data they’re trained on, this is such a massive misconception: human data is only a bootstrap to a reinforcement learning phase. This combined with the fact that we have verifiable rewards means it’s just a matter of when not if for any given level of reliability.
- the massive compute investment implies that the compute that we’re building over the next 2-3 years will 10x the effective compute for training models. That combined with various R&D contributions (historically which have been very significant and there is no shortage of wins here), better data curation and flywheels, richer data (wait until conversation capability gets good) means we have several orders of magnitude of runway that we know of, today.
In short I don’t see any compelling evidence to suggest all of the trends we observe in many different ways will end any time soon.
And you propose the same companies that have been cost cutting and avoiding buying you a chair for ever won't start objecting to a $200/dev/month subscription? The finance department won't have a say?
> For coding you always want to go with the best model in the category
Will this always be true? There will never be an event horizon/point of diminishing returns where something not-bleeding-edge is "good enough" for 51%+ of users?
As long as closed source is 6 months ahead in terms of current difference. Although this is hard to figure out using simple percent based coding benchmarks, you def. notice it when you're actually trying to do a long task. Even simple things like UI "taste" is enough for me to use opus instead of 5.5 though even though 5.5 is strictly better for anything that doesn't have a UI, ie backend, scripts, making agent workflows etc
This is a silly take. There is a line of "good enough" for most coding (most CRUD apps and APIs are nothing special), and once we are past that, nobody will care about having the "newest, best" model except extreme outliers. And this base "good enough" model will become an ultra cheap commodity as we already see with GLM, deepseek, etc.
As long as closed models are 6 months ahead I won't be switching from them to prev. 6 month SOTA open source models. Maybe its just a different calculation if you're in a job, but as an indiehacker I'll take any edge I can get
Ofc again, can be convinced to switch if there's however a clear speed difference, like 5x+ for a open source sota even if it was SOTA for 6 months ago
Oh, hey, I recognize you. Thank you for the very forward and thorough orbital sander recommendation at Home Depot. That's exactly what I wanted to deal with on my holiday weekend. You just know so much about this and the rest of us are simple passersbys.
Yep sorry was just pulling it out my rear, not like a market trend that nearly every enterprise uses Anthropic or Openai models for coding or that Anthropic has had such ridiculous growth that they're 10x-ing year over year
I'm ribbing you for writing like a condescending guru that invalidates the evaluatory capability of your peers. Not the meat of your justification, which I notice has now shifted from saying they're quality leaders to market leaders.
I think RAG is a mostly outdated concept now, it's been subsumed by the idea of a "agent harness" which is exactly what Claude Code and Claude Cowork and OpenAI Codex and Claude.ai and ChatGPT themselves have now become.
An agent harness with access to a good search tool is a much more interesting thing than 2024-era RAG systems.
For coding assistance, I have tried OpenCode with several large open models through OpenRouter. All were fairly bad compared to Claude Opus.
Could you provide some hints on how I should be holding these open models so that I might get more value out of them?
I agree with the common trope that open models lag behind by about a year, but something magical happened just around a year ago when the state of the art models became extremely useful. By this reasoning we're about to see open models perform well, but I'm afraid there is more to it than just waiting for another revolution around the sun.
Note, my application is coding assistance. Open models can be great for other purposes.
No one is claiming that OS is as good. They are saying it isn't that far behind SOTA commercial products. So why pay exorbitantly just to get something only a few percent better than the free option?
But there have been very good open source office apps for decades and few enterprises use them, so perhaps this is just the nature of B2B purchasing committees and 'nobody getting fired for buying IBM.'
Do more planning yourself, be smart about the context, break down tasks into smaller components, give it more guidance. You can't just lazily prompt it to complete large features autonomously and expect good results.
The point is that the price gap is so much larger than the capability gap, that even with the extra compute needed to make up for the lack of capability, you can still come out ahead in terms of amortized $/work done.
+1 .. just wanted to reiterate that this is the answer. The open models work great if you just do a little more of the design/architectural work up front and organize your work appropriately.
a good harness is supposed to do what you are describing.
sonnet on pi.dev is pretty terrible but fast. Claude Code has ridiculous amounts of prompt engineering at system prompt level and sub session spawing combined with low temperature, to provide the predictable results people like. CC screws up and you never see, because the harness auto corrects, while on OSS you see everything, and does not comes with the level of monitoring by default.
GLM-5.1 isn't just as good. It is no match for Opus running in Claude Code. Please try it yourself. Open source models are about a year behind at least.
This is profoundly misinformed. I use all three of those models regularly and the difference is just not that big anymore. GLM 5.1 is at least as good as Opus 4.5 - when it’s my dime it’s the primary model I use and switch to GPT 5.5 for planning and review but it’s also very capable at those things. If I had to pay API rates for everything there is no question I would only use GLM 5.1 (and Minimax for exploration tasks).
At work I mostly use Claude Code and a bit of Codex; personal projects are OpenCode and honestly I prefer it.
In the second half of last year, I found that agentic coding with proprietary models (≈ vibe coding) reached the point where it actually speeds up my ability to deliver useful code at work. Before that, AI-based autocomplete definitely helped, but (despite the claims of the people selling AI coding tools) letting an agent author more than a file or so at a time (often a function or so at a time) required a very intricate plan or it would create a mess. Creating that plan or cleaning up the mess would take longer than just doing everything myself.
For me, it feels like widely available open models have recently crossed that same canyon. Are they as good as e.g. late-model Claude Opus? I don't think so. But they have absolutely gotten past the point where they are beneficial. This means that, for me, they are about six months behind.
Exactly this. GLM 5.1 is the first open model that I thought "actually worked" for agentic coding, which puts it in the same tier as Opus 4.5 - which was where I flipped.
For coding I wouldn't say a year, last year this time claude or gpt definitely weren't able to do what GLM is able to do today, but easily 6 months I'd say.
I use composer-2 daily for complex programming tasks. It's a fine tuned Kimi 2.5 - nothing groundbreaking. I've even had reasonable success using Qwen 3.5 on my desktop GPU. Opus might be better, but it's certainly not necessary to get good results.
I have 26 years experience. I code using GLM-5.1. Fron time to time I switch to Codex / Claude, and honestly I don't understand why people uses Claude or codex. With the right prompting, GLM is awesome.
Don't you need to spend 5-10 thousand USD to run these models that are "as good" as frontier models from 6-12 months ago? I haven't seen a convincing breakdown for ROI of running your own coding models. Especially against a $20 or even $200 plan
Do you have a good source to refer to, to map out migration from Claude code to a cheap setup using small open source models like you’re describing? I’d certainly like to experience how good they’ve gotten.
The only way I see it working out for them is if some legislation is passed that eliminates the competition by making it illegal to run local models. They could claim that the models are dangerous and could be weaponized without oversight, or something along those lines.
They are both (and also spacex) sprinting for IPOs. They know that the opportunity window is closing fast and that advancement in model quality has largely plateaued in the last year. Take as much investor money as you can get away with for now.
“Tokens” don’t have an intrisic cost or value. Saying that I used $2,180.16 worth of tokens is like relying on the salesperson to convince me I’m getting a billion dollars worth of pots and pans for $19.99.
I think it’s funny how we are throwing critical thinking out the window when it comes to evaluating biased sources of info.
I'm not sure what you're pushing back against here.
I spent $200. If I had been paying API pricing it would have been $2,180.16. The article is about how enterprise customers get charged API pricing, which means if I had been employed by one of those companies I would have cost them $2,180.16.
Just because API pricing would've been $2180.16 doesn't mean that's the value of those tokens. For starters, you personally probably wouldn't have paid that. But also, sales price isn't value. This is like saying, oh, I saw this bar of gold somewhere for $10000 but got it here for $1000! So I got $10000 worth of gold for $1000! - no, the value of that gold is determined by its weight, which wasn't even mentioned.
We have no market convergence on tokens yet (and it'll differ between LLMs), so it's impossible to say what value you got for your $200.
He's saying he's getting a great deal...a token from Opus on Claude code is the same as a token from Opus on the API. I remain as confused as Simon. He's not talking about "here's the ROI I got from my $100 subscription" it's "here's how much I saved from getting the monthly subscription instead of sending things through an API".
Right, the confusion is that the quote-unquote "subsidized" monthly pricing is often used by Anthropic/OpenAI skeptics as proof that inference is unprofitable, i.e. the API would have cost $2000 but you only paid $200 for a subscription, therefore OpenAI is selling dollars for 95 cents and the house of cards is about to collapse. As the GP says, this is faulty logic because we don't know what the actual cost of a token is; OpenAI might only pay $1 in inference costs, in which case they're merely "incredibly profitable" making $199 off you instead of "ludicrously profitable" making $1999 off you had you used the API.
But to your point, re-reading the article, this is not what Simon is saying at all; he's just pointing out that he got to use ~$2000 "worth" of tokens on his $200 plan. Which makes total sense! Subscriptions are sticky, that's why the entire software industry moved towards subscription models (as much as we hate it); the person paying $200/month is more likely to stick around than the person who paid $2000 using the API.
He’s just making a point that $100/mo subscription is cheaper than the API costs and that the same work would cost 10x as much on API pricing. It’s not some sort of swindle it’s a trade between consistent revenue that power users will overspend on but plenty of people will underspend (I’m on the underspend side) or pay-per-request service that makes more sense for say scaling out something custom (I run APIs a lot when I have say some pseudo labeling pipeline I want to scale to a lot of data. For this I typically use smaller models but lots of cases where you want larger models). What’s the wrong message here?
No, value is determined by what participants in a market are willing to pay for something. The only reason you are able to say that the value of gold is determined by its weight is that gold is a commodity and no matter what you paid for it you'll find others willing to pay market price.
Simon is saying that companies are (today) willing to pay API prices for tokens which is as good as any determination of value.
Is this some anti-FIAT take? The value the author got is not value as in intrinsic value, it simply means value as in better deal than the alternative. This is often called "value" and you will see this used when products are sold in "value packs" etc.
> Just because API pricing would've been $2180.16 doesn't mean that's the value of those tokens.
You seem to be suggesting the price of tokens is entirely disconnected to the cost of providing the service? I don't see much basis for that assumption.
Hi Simon, nice article. The parent there may be making the same assumption I am, that large enterprise _never_ pays sticker price.
Also, to just color in the picture here, as I haven't seen it mentioned elsewhere, there is a very large Saas company at the moment who has given everyone unlimited tokens on Claude. And they have a dashboard showing who spends the most. So the "budget" went from about USD500 per per person (split between Claude and cursor) in Jan to... Well a soft limit of USD100k... Per month... Per person.
People can still see the top line sticker price on their spend, but honestly I can't believe that the Saas is paying that full price when the invoice comes in.
That said, there are some finance reports which are probably dropping soon where we will find out!
> The parent there may be making the same assumption I am, that large enterprise _never_ pays sticker price.
I shared that assumption until yesterday, when I found out that it wasn't holding for LLM pricing from OpenAI and Anthropic. That's what inspired me to write this piece.
I think those token leaderboards are an obviously terrible idea and will go extinct very quickly now that people are paying attention to costs.
But the feature list at https://claude.com/pricing#team-&-enterprise literally lists "tiered incentives on committed spend" and "non-standard terms" as perks of the sales-assisted Enterprise plan. Maybe "non-standard terms" could mean "we dance for you if you pay", but what would "tiered incentives on committed spend" mean besides "we can negotiate on price if you bring the volume"
Volume discounts may be available for high-volume users. These are negotiated on a case-by-case basis.
* Standard tiers use the pricing shown in Model pricing
* Enterprise customers can contact sales for custom pricing
And there are discounts available through "Claude Platform on AWS":
Anthropic rates your token usage in USD at standard per-model, per-feature rates, applies any negotiated discount, converts the result to CCUs at $0.01 per CCU, and reports the CCU quantity to AWS Marketplace hourly. Your AWS bill shows a single CCU line item.
> With the pricing change, customers of Claude Enterprise, a two-year-old bundle of products meant for large companies that now includes Claude Code and its work assistant, Claude Cowork, will have to pay for the amount of computing capacity they consume while using the software on top of a monthly flat fee of $20 per user, an Anthropic spokesperson confirmed.
The sentence you quoted doesn't say anything about the token price for enterprises matching the sticker price. It just says that enterprises are paying a consumption based price (presumably billed as some $/token), on top of a per-seat fee.
> I found out that it wasn't holding for LLM pricing
You're correct. When a type of cloud service grows large enough and has a few competitive suppliers, enterprise pricing tends to coalesce with the large buyers paying around the same price for the same thing. While that might be lower than the publicly cited rate card, the private price similarly large customers pay ends up being similar.
One reason is that the very largest, long-term enterprise customers are so valuable, they can command MFN clauses ensuring no one else is paying substantially less for the same thing, then the rest of the rate card for smaller customers flows down from that. There's a strong disincentive for vendors to cheat or allow big disparities between similar classes of customers because the number of people involved on both sides of these deals is large enough that word will get around eventually.
Large scale enterprise sales and purchasing in a given sector tends to be rather circular. Account execs move to other vendors and call on the same customers, while purchasing agents can move to other customer firms. Personal relationships, reputation and credibility matter. Lying to screw a large customer over just to make one commission or quarterly quota can be a very bad long-term career move. Sometimes purchasing agents or executives quietly compare notes off-the-record with their peers from other similarly-sized firms. After dinner drinks at industry association meetings and trade shows can be quite productive in terms of verbally exchanging 'market insight' with peers.
When there are significant pricing differences, its usually due to different volume commitments, SLA/QoS guarantees, payment terms and other material factors which justify the difference. Source: been there, done that inside a top ten valley tech company. Once was in a meeting where a newly minted EVP tried to get a long-time senior account exec to pressure a huge customer by being semi-dishonest. The account exec schooled the EVP on the fact that the EVP could only make him unemployed for about 8 hours but that huge customer not wanting to work with him could make him unemployed forever. :-)
I do know of moderate-size companies deploying OSS LLMs on their own GPU clusters, for ownership/security/maybe cost reasons. I'm somewhat surprised F500 companies are apparently just handing over all their data to the model providers.
Could be fantastic for small shops while it lasts. The big guys have to pay 10x for precious tokens.
And "large" just means that AWS will assign an account manager to talk with you. I was at a start-up who spent $300k/year on AWS and that was enough to get special attention and discounts. Enterprise pricing is confusing.
They pay sticker price. There may be exceptions for very very large companies like Amazon or Microsoft which have their own deals where they rent out compute in return for usage.
Claude is so in demand at the moment that there aren't really volume discounts. Anthropic sets the terms and you either accept them or get lost they have that much of a lead (mindshare/desirability wise).
> If I had been paying API pricing it would have been $2,180.16
The point being made above is that API pricing is calculated... somehow... seemingly arbitrarily. Possibly untethered to the infrastructure costs entirely: which would be the basis of any 'value', however that holds the labor theory of value, which isn't accurate either. So how do you accurately price these tokens at all (other than through price-discovery: which is slow, messy and fuzzy)?
> (other than through price-discovery: which is slow, messy and fuzzy)
I notice a distinct lack of reading or comprehension (from everyone around me now, not just this comment) which worries me. I worry if LLM's are to blame. No one reads anymore...
I imagine a number of Hacker News members might be devs that haven't dealt with terms like “price discovery” before, so we should all try to show grace.
I don't really understand the holdup about that specific line? Anthropic publishes the cost per million of tokens, what's arbitrary about looking at your consumption, and doing the math?
Or are you saying that Anthropic is determining that cost per million arbitrarily?
If so, it'd be like asking to explain why things fall on the ground other than through gravity. Companies pick the price they think the market will bear, and adjust based on new information.
API pricing drops DRAMATICALLY in enterprise agreements.
As with pretty much anything priced on volume/usage.
Enterprise deals are negotiated ad-hoc, the listed pricing is simply a jumping off point for the final negotiated discount.
If you’re going to give 20,000 employees Claude code you are not going to be spending $1B per year on Anthropic tokens as if you gave everyone an individual API key. Just as Anthropic isn’t paying AWS SES $10,000,000 to send 1 email update to their massive user base when the next Claude version drops.
This isn't true at the moment, though. So far there hasn't been the negotiating power. What happens is you end up capping usage for employees at a fixed amount. I think eventually, prices will come down and there will be discounts, but for enterprise accounts at least of our size (<5000), we're paying almost 100% retail, which kind of sucks, because it's expensive, and pretty easy to burn $50 to $100+ in a day, if you're not careful. In fact we got pushed off the former plan to the token-utility one at the last contract negotiation.
Going to be interesting to determing the metrics we give to engineers for determining whether the spend on this is worth it. Measuring PRs, lines of code committed, commits fully generated by agentic workflows, etc.....
As someone who has seen the enterprise deals, they are not subsidised in any way shape or form. Anthropic may wave the seat fee, maybe get lower "expected" consumption. This changes what the company pays up front. but token prizing is fixed.
> Just as Anthropic isn’t paying AWS SES $10,000,000 to send 1 email update
How much do you think emails cost? That number is just so far off?
But besides that, running SES is also quite a bit cheaper than SOTA ai models with high demand (and comparatively) no competition. And quite a bit more pressure to make money (soon).
I think it was a figurative example. For what it's worth, $10,000,000 buys you 100 billion (1e11) outbound emails on SES at the sticker price ($0.10/1000 emails). One source puts the number of emails sent worldwide in 2024 at 132 trillion (1.32e14).
edit: I missed the "enterprise" feature matrix with the usual audit/compliance stuff to force the biggest enterprise customers onto enterprise plans. Otherwise the "teams" plan is much better value for any business.
Teams premium is "Everything in standard, plus more usage*"
And from my experience, it's a very generous usage, I've only hit the limits once or twice, and both times required multi-boxing agents.
I could single-window agentic development all day on opus-4.7 auto-mode without hitting limits.
If you're a business using claude, then that seems like the right plan, the enteprise/API plan seems more suited to where your product is built on top of the agent themselves, so seats/limits aren't really meaningful?
Claude Teams and Claude Enterprise are 2 distinct plans. Simon is right that Enterprise seats have no included usage (and so all usage is charged at API billing rates), whereas Teams seats do.
Tokens do have a clearly calculable intrinsic cost. There's the marginal cost of production (i.e. the inference cost) and the amortized R&D cost that goes into the model producing them.
Yes, value is hard to calculate, but luckily market pricing mechanisms exist exactly for this purpose. There isn't a better number to use than what people are willing to pay for them.
So he's saying that on an enterprise plan, he'd be spending $2,180.16. He's not paying that much, but enterprises are.
OpenAI and Anthropic won't let them any more. Once you get above a certain size (I believe 150 seats) they push you onto the enterprise plans.
I don't believe you have the option to keep with the $200/month flat rate subscriptions any more. I'd be happy to be convinced otherwise.
(I dug into this a bit more and couldn't find anything in their consumer terms that say "you cannot use this personal account if your company has more than X people", so I imagine the pressure is more that your big company's purchasing department really doesn't like managing hundreds of individual subscriptions as opposed to a single, stable, predictable negotiated contract.)
The purchasing department _really_ doesn't like it, but it's the corpsec department that will really want to murder you for using "personal" accounts to interact with corporate codebases.
i am pretty sure these services know what it truly costs them to serve you tokens, maybe not in realtime but at least periodically.
however, what they charge us is a constant exercise in price discovery. i agree with this sentiment in the sense that we don't have a stable sense of the cost. all of these comparisons are good for the moment, or at most the near future.
i believe that even the "all you can eat" approach with the max plans, regardless of their crazy pricing, is not sustainable only with the power users. if most of us gets this kind of value through our plans, surely it does not incentivise the service providers to continue pushing it. maybe they can regardless just to gain market share, but not forever.
From my back of the envelope analysis for my own projects, paying per token on OpenRouter is competitive if not cheaper than running the same open weight model on a rented GPU. Per-token pricing is in the same ballpark (although more expensive) for closed frontier models and open weight models (cents to dollars per million). To me this says that the pricing is somewhat grounded in reality.
Are you comparing single-user requests or multiple concurrent requests when you say comparable to rented GPU? Most of the cost efficiencies kick in with concurrent/batch requests. A single H100 node can provide like 5k input + 2k output tok/s on a model like Qwen 3.6 35B-A3B with 30+ concurrent requests.
His point is more he was surprised enterprises weren't getting the discount. And so indeed maybe it is not a giant ponzi after all! (Could be a bubble)
Lol. They obviously have intrinsic cost, the floor being the cost of electricity. It’s hilarious how we are throwing critical thinking out the window when it comes to evaluating biased sources of info.
I have to give them kudos. This whole thing is the greatest swindle of all time.
AI has some use cases, but not at the price it’s currently priced at. I’ve been on AI since GPT-2 with a lot of heavy users. Every user has the same story, curiosity, surprise, hype, hate, realization. Enterprise is usually a bit behind and are right now at hype cycle, that’s where they sold all the deals and do the IPO.
It’s really a VC masterclass.
Don’t get me wrong there is are useful cases of AI, but not the way the want it to be. Quite similar to Blockchain. The idea of decentralized money has right to exist. 99% of other coins not.
AI is a faster, but still less accurate search engine. AI is great in finding bugs, it’s great at ruber duck debugging.
The reason I call it a swindle is, because along with the marketing it gives tons of people in the world the impression, they can now build their own startup, game, infra etc without the need to learn it themselves. This leads to millions of abandoned and low qualiy projects and products, because the vast majority has never built the mental modal necessary to solve the problem thoroughly. In the end they’ve wasted months and money (but burnt tokens). This is what I call a swindle.
All early adaptors I know have not drastically winded down their usage, not because of money, but because there is no new case. If you want to explore a new project you can get onboarded quickly learn a lot and then switch to documentation and live testing. For me usage is the lowest it has been the last 2 years.
I would not let AI touch my code. I have anxiety around it, because it will gripple back up. I will let it read my code and let me know what I did wrong so I’m sharpening myself.
100s of companies including open source solution can offer that for me.
All my non-tech friends are now in hype cycle and share their hype and fore forseeable frustration with me.
I have to say I’m in a way impressed in how AI has been rigorously vc-utilized (conciously or not-conciously) to generate these vast companies with the whole world watching.
Trying to parse the raw claims here maybe you can help me out
- it’s a swindle because ROI of tokens for coding models is not positive? As in it doesn’t bring enough value to charge like the $100/mo?
- enterprise customers are too dumb to see this
- IPO to max out the CEO profits for what is ultimately blockchain vaporware
Am I getting that right? Or am I putting words in your mouth?
> it gives tons of people in the world the impression, they can now build their own startup, game, infra etc without the need to learn it themselves.
I can’t speak for peoples beliefs and motivations, but this seems to be strawmanning, no? AI is a powerful tool to force multiply people. You can’t just prompt “build me an enterprise SaaS app worth $1B” or “build me GTA6 and don’t hallucinate” but is that your impression of what’s happening? Dario and Sam are saying “if you buy our coding agent subscription you can build a game with zero skill and one shot and then be rich”?
If you don’t find value in AI agents I can see reasons why that could easily be true today. Also if it just gives you the heebie-jeebies. But to say it’s a swindle on par with the blockchain I think that contradicts an enormous amount of signals and also the actual dialogue (not just headline sound bytes) around what these systems are capable of today and what we expect them to do say at the end of the year.
You kind of got it right, but the biggest loser of them all are the investors especially the index investors. They don’t even decide what they invest in but the savings that goes in funds need to invest in these stocks.
It’s quite an elaborate swindle obviously. But you generate hype with underselling your core product, you claim way more usability then there is. Users will experience usability initially. Everything multiplies with each other and then you put it on the market. Everybody involved makes money and you’ve succesfully extracted money from everyone who’s invested in NASDAQ index funds at the very least.
> Dario and Sam are saying “if you buy our coding agent subscription you can build a game with zero skill and one shot and then be rich”?
That’s Anthropics marketing, yes.
Also their offering is not uniqe that justifies a 1 trillion valuation. The first companies are already rowing back. It’s a really certain time window that they are about to hit now with their IPOs
The companies that have signed these enterprise deals haven’t done a ROI analysis. They had Fomo.
> you generate hype with underselling your core product, you claim there’s way more usability than there is
Isn’t this a contradiction?
> Everything multiplies with each other and then you put it on the market. Everybody involved makes money and you’ve succesfully extracted money from everyone who’s invested in NASDAQ index funds at the very least.
Sorry I may have totally missed what you’re saying here. Anyone in S&P has already made a lot of money thanks to effects from these companies. No one has to invest in an index fund. Markets have risk…
> That’s Anthropics marketing, yes.
Show me.
> Also their offering is not uniqe that justifies a 1 trillion valuation. The first companies are already rowing back.
That there is competition doesn’t imply they aren’t worth 1 trillion.
> The companies that have signed these enterprise deals haven’t done a ROI analysis. They had Fomo.
Also…wrong. I have seen the data at my company, everyone at scale tracks this.
I don’t remember ever hearing Dario or Sam recommend replacing people. Rather they say that smaller groups of people can do more work, so hiring will slow because small teams can do more.
The only times when people talk about actual full replacement of people is always when they are talking about some “future AGI” that is far more capable than the tools we have today.
I usually agree with Simon but I think he is overlooking an important factor.
There is a lot of AI usage happening not because it shows benefits, but because the business has mandated its ubiquitous use. Companies having dashboards for token usage and rewarding people for using more tokens is a real thing. I just spoke with someone today who works at Microsoft and they are required to use AI for all of their work - they have to make a special request with justification if they decide not to use AI for even a single PR. This kind of demand isn't driven by value from either the company itself or from its workers; it is the kind of artificial demand you get from make-work projects to keep people employed during hard times.
We have to wait for the hype to settle down and people start making business decisions based on results before we can really value these AI products.
I think token leaderboards are idiotic, however… many companies are requiring AI use, with by far the most likely reason being that they truly believe AI-assisted development gets (or will after the time needed to get employees acclimated) better returns.
I'm skeptical that their current price raise is sufficient, and I'm also skeptical that most users/businesses will accept more significant price raises that will be needed. Especially for individual users, $200 a month is already incredibly expensive, I really don't think most people are going to be willing to pay more like $1000 a month.
I think we could eventually get to a point where $1000/m+ consumption would be considered a bargain.
I currently pay ~$150/m in tokens to cover my 1099 work. If I had to pay $1000 for those same tokens each month, I would still be massively positive on my margins. I bill customers based upon # of completed features & bugs, not total time spent at the computer. Tokens would have to more than 10x in cost before I would start to have a problem on my end.
The real timing is that we don't have strong enough new business needs for now and we have accumulated enough tech assets, so our work has been increasingly incremental. That means we can build reliable features on top of vast amount of past work - where AI really shines. So, with or without AI, companied would hire fewer software engineers if majority of our work is incremental: add a feature here, fix a bug there, tweak a configuration and etc, then we wouldn't need as many software engineers anyway. AI just accelerated such squeeze.
In contrast, imagine if we had the same AI 20 years or so ago. Could AI really write Jersey? I guess not as people were still trying to understand JAX-RS. Could AI really answer all the questions about React? I guess not as React was just invented. Would we use 10x fewer people to build out infra on the public cloud or the entire so-called Big Data platforms? I guess not, as they were still rapidly evolving and we'd need so many engineers to explore so many different possibilities? Could we use AI to build our ML ecosystem with 10X fewer people? I highly doubt so. Heck, 20 years ago R was all the rage and Python's ecosystem was not mature at all. Oh, and mobile computing, could AI lead to 10X fewer people to build all the mobile apps and the underlying infra?
ICL = In-Context Learning where LLMs learn to perform new tasks by reading examples directly within the prompt, rather than requiring permanent model retraining.
If those companies found PMF we should expect to see the effects of all those tokens burned on the products we all know and use. In my own experience: yes new features can be delivered much more quickly (by orders of magnitude) but knowing what to build and how to evolve a product with taste and direction does not; so ultimately the user-facing gains are marginal.
I suspect that once the technology has been tamed and the hardware and software has been commoditized, the impact will be much less dramatic than we expect and we will realize the importance of a shared vision, experience, taste, intuition and discernment in building good products.
this is an opinion and i would love to debate :) dont hate me too hard
i think the article is momentaneously correct but there are some things that smells to me about the situation overall (not the article!)
(1) i believe gpt 5.5 and opus 4.7 are not as good improvements as their predecessors and there has been not enough evolution in those models to justify the price increases. Unless something big changes in the next few years they will not be able to track these costs (unless A)
(2) they might not be able to keep up improving the data centers if their tech keeps demanding more and more hardware capability. even nvidia does not show signals that they will be able to beat too much their big GPUs and at some point this increased price will be passed to them anyway which will need to be put in the token price (unless A)
(3) i've been trying deepseek v4 and other models and honestly, they are more useful than gpt 5.5 and opus 4.7... i mean, there is still a difference, but it is so little that it does not make sense the cost of opus and gpt 5.5 (unless they are going to A too)
(A) I have the impression is that the plan was the whole time to sideline common folks like us and focus on gov, big techs and military. they only let us play with the toys to gather data and for people to not get mad because they stole the internet from us.
The costs are exorbitant and most software is not produced by companies with such a huge moat. Anthropic made a profit through their recent bait amd switch pricing. There is zero useful insights online to indicate whether this might die due to commoditisation with good enough open models or fail the race to get more people subsidising unsustainable growth with other people’s money. Who knows? In any case they dont seem to be able to drop usage costs so the business model seems based on wishes
> Stories are circulating of companies surprised at how expensive their LLM bills are becoming from usage by their staff
> Enterprise customers are now paying API prices
How long before enterprise customers start to question the bill? Anthropic goes from not making money to doing pricing shakeup, and now they are making money and the biggest spenders are shocked at prices.
Costs will plummet as better hardware becomes available and priced reasonable so that people can more easily run their own open models locally. But that won't help Antropic/OpenAI make more money, quite the opposite.
Yeah, it’s called supply and demand. Demand for memory went way up suddenly. Now supply is going up rapidly as companies try to cash in on that demand.
Supply will eventually catch up with demand. Then the prices will come back down.
Algorithms are also improving. I believe it's very unlikely for these two improvements together to not result in one to two orders of magnitude cheaper cost per "intelligence". Of course, that might just make use cases that are too expensive today viable and thereby increase usage further.
A lot of the new hardware requires retrofitting existing datacenters for appropriate cooling, or is waiting to be installed because the new datacenters haven't been built yet. By the time they're installed it's likely a lot of Blackwell GPUs are going to be very out of date. Newer hardware is turning into huge capex bills along with the corresponding depreciation costs. Basically, it's not the same as plugging a new GPU into your desktop, the upfront investment is extremely expensive and all the numbers I'm seeing suggest that the newer GPUs are costing more to run, not less.
Winch Design [0], which have built some of the world's largest superyachts [1], seem to be using AutoCad. [2] Afaik it's also the same with Lürssen (but don't quote me on that)
Those tools are used in ways that they're integral to processes. They have their equivalents of ticket systems that are linked to code repositories with LFSs and bunch of IDE type tools and automated and manual test systems and build systems. Their equivalents of PR discussions and Selenium screenshots needs to check all boxes in the right ways for legal and traceability purposes.
Without all that might be $175/user/month but you're not shipping apps with just vi and bare gcc.
Except LLM's even with Vision are still useless at AutoCAD let alone Revit (please dont quote SCAD LLM's at me, useless). Knowledge based approaches still win.
I might agree "AutoCAD" is the current level LLM's are at, but wait until your design departments discovers "Revit", its another ballpark (in wasted cots, engineers on site still get "clashes").
Revit costs are high, and the end results are marginally better - but local LLM's tokens are cheaper 24/7 at "AutoCAD" level - "Revit" level tokens will make Ubers CTO/COO weep harder than they already do. While producing results no better than "Revit" does (engineers still face "clashes").
As someone completely outside the 3D design world who always thought of AutoCAD as the gold standard - really? What program would be used instead? Please enlighten me.
Cadence and Ansys have entered the chat. A bunch of other highly-specialized engineering software has entered the chat. Licenses are on the order of 10-100k/seat.
I guess we are welcoming the software people to the world of expensive tools. Just sad that the FOSS alternatives of these tools are not as powerful whereas software industry still has FOSS tools to fall back on.
Yeah, it’s nothing, and it’s also not the cost that enterprises are paying. As the article states, the price is $20 per seat per month, PLUS per-token API usage. Enterprises are paying consumption billing, not fixed rate oversubscribed “all you can eat per seat.”
So the author claims he's getting $2000 per month worth of frontier AI free of charge. Ok. If he's been doing that for 6 months that's $12k. What has this produced concretely? For $12k you can find a used car in decent condition. Heck for $1200 (his actual out-of-pocket spend) you get a brand new ebike! (on which you could put a pelican and make a photo of both if that's your fancy). But here it's unclear what has come of it.
> code that would have taken me years of work to produce without LLMs
As you might suspect, this is what I have an issue with. Without LLMs, isn't it possible or even likely that that code wouldn't have been written at all, and wouldn't have been missed? If LLMs are mostly used to produce throwaway prototypes then it's a stretch to say that's money well spent.
If indeed it let you advance your main product much faster then sure it's a different story. You're the judge of that. It's hard to see the impact from the consumer side; everything is still broken and no extraordinary app seems to be emerging. Maybe it's just a question of time. We'll see.
I've thought about this a lot. I am very confident that the way I use LLMs is both accelerating progress on my core projects (here's a substantial, reviewed PR I landed just yesterday https://github.com/simonw/datasette/pull/2741) and helping me create plenty of projects that otherwise would not have existed.
The point being made by GP was that your projects have no value and their non-existence wouldn't be a negative to this world.
And that is likely a fair assessment, though I understand perfectly the feeling that you have that you are accomplishing great(er) things thanks to AI.
I was certainly not saying that all the author's projects, in general, have no value! That would be rude, mean and most of all, incorrect.
But yes, it's likely that the ease of which code can now be outout lets us produce lots of unnecessary code just because we can, and the author says as much in a below comment
Now what percentage of the 200$ have been used on the useful stuff and how much on exploration or other stuff.
How long would it taken you to do it yourself? How much longer will the next task take you, compared to when you would’ve written the code yourself. How is the mental model compared to when you would’ve written it yourself?.
I’m not saying you’re wrong, again there are use cases. But the calculation is not plain and simple it goes deep into our perception, perceived productivity versus actual productivity.
I’ve 2 months maxxed out all 6k of Claude Code and bought Antigravity on top. My codebase became 140k lines. I introduced tons of bugs and spent another 2 months, deleting 80k of code. I wish I would’ve just chatted with AI and not let agents touch my codebase. I would’ve saved approx 300$ subscription prices a month and 2 months of my life.
I had that experience back in January! I used it for a joke in a talk recently... did anyone really need a half-finished buggy slow implementation of a JavaScript interpreter in Python? They did not.
https://github.com/simonw/micro-javascript
It's taken me the best part of this year to readjust and find a pace and level of ambition that fits.
It was challenging for me as well, but my pace and level is now that I use it for these cases at the moment:
New language, infrastructure, general level of understanding of something I barely have an idea of.
Rubber duck debugging, if i dont know the correct solution
Checking my code for issues and bugs.
But not for:
- writing my code
- agentic coding (help me)
The inference has reduced drastically. It’s basically just chatting. I don’t let it write anything, but sometimes I purposely use the browser window instead of them sitting in my codebase, because I know it gets things subtly wrong and migth focus on the wrong things.
The same way people used to say don’t copypaste code at least write it out I think it’s still true. It helps to buidl the mental model and to find the right abstractions.
This is the economic theory of value creation though - arguably the world is better off because new projects can be created, and they are marginally cheaper than they would have been previously
I’m watching to see what happens to big enterprise software contracts. Why pay some vendor $800k annually for something a couple mid-level devs can replace—-and tailor closely to your needs——by leveraging AI.
Open source software changed the world. AI that will cheaply write whatever you want in a few days will also change the world.
(I have a feeling if I could say "and I closed $2m in sales with the software I wrote!" people would find a way to say that didn't mean anything anyway, because how can I prove I wouldn't have made those sales writing it by hand?)
I would be very curious what kind of answer would satisfy you here. Simon isn't building a product, where $200 is a line item on a balance sheet. If he tells you what sort of analyses or time savings $200/mo on coding agents have enabled him, do you honestly think that would satisfy you?
How is the lack of bad news declaring a victory for AI? I am yet to see any company concretely publish analysis about the ROI from AI. Most companies as far as I know are still treating AI investment as sunk cost with no expectation of returns at the moment. We could very well see a world where companies heavily scale back investment.
I may be biased because I work on an AI powered enterprise productivity product, but while I agree they have PMF right now, I wonder whether people's use will evolve in ways that undercut the current PMF. Chatting with an assistant is great if there's no product with tailored UI available that also has the AI capabilities. But once there is, I suspect people may switch, or more importantly enterprises may switch because they'll get the benefits of AI without the clunkiness of a chat only interface. We may see another DOS -> GUI-like shift.
More specialized products will consume tokens but their builders will be incented to optimize token use and switch models as costs and capabilities change. And if search engines become more AI capable, and Google is clearly striving for this, then they may have pressure from two sides that could squeeze the number of use cases for AI chat. AI coding isn't going anywhere and nor is the need for AI in general but I wonder if the products will have to evolve significantly to maintain the current levels of PMF. And then there's the question of profitability...
With deepseek and xiaomi mimo models slashing their prices 99%, I don't see a great future for openai / antrhopic with regards to their 1T valuations. Maybe 1T valuation will be the whole market, West + East.
Most of the corporate world in the EU or North America will be hesitant to rely on Chinese AI providers. There are some very real blockers for that for things like data security, compliance, etc. And recent geopolitics don't help.
Legalities aside, you need to look not at the model quality but at the infrastructure needed to scale these models from tens (now) to hundreds (soon) of millions of users. Only a handful of companies actually have the resources and funding to do that. That's what these huge valuations are based on. These companies are gearing up to scale to these levels. That's why they are spending on data centers. Whoever has access to those data centers gets to tap into the revenue stream of people using models running on those.
The market for frontier models is roughly split between OpenAI, Anthropic, and Google. And then you have companies like X/SpaceX, Amazon, and Microsoft being more successful with their infrastructure than their AI products and companies like Apple, Meta that have the money and the aspiration but are so far not really managing to be very successful with their AI strategies.
Deepseek is just very poorly positioned to capture a lot of the enterprise revenue in the EU or North America. But they might become very dominant outside the US/EU. And of course China itself is going to be a huge market and equally unlikely to want to be depending on US owner AI companies.
There is still a risk of supply-chain attack. People give LLMs direct access to their entire infrastructure via tools, and never check the code produced. It's not difficult to steer an LLM during training so that they'd output malware only when prompted a certain way, and that wouldn't come up during the initial evaluation.
Personally I see no difference between China and America in terms of risks of them embedding "backdoors" so to speak, but I disagree when people claim that open-weight models are obviously safe just because they can be ran locally.
> It's not difficult to steer an LLM during training so that they'd output malware only when prompted a certain way
Perhaps, but that's also a good way to lose users+reputation as there's no way to control when said malware is generated. Once the first instance is discovered cybersec researchers will have a field day reproducing it and showing the world.
It is not a trivial challenge setting up model serving infra for ~1T or larger models, especially in a high reliability environment (e.g. your team is using it for work, or you're using it to power production apps). Sure, there are third party providers, although the quality and reliability of their inference varies.
Run Deepseek on Deepinfra then? Or Fireworks if US-based is important. None of these are real issues outside maybe convincing your legal team to do a bit of homework.
I don't think you are appreciating the physical constraints here. Deepseek doesn't really have the hardware in the US or EU to do anything at scale.
Sure, you can self host a non-frontier OSS model yourself; including Deepseek. And no doubt some people will pay one of the companies I mentioned to rent the infrastructure to do exactly that. Much of the rest of the world will be paying directly for direct access to the frontier models.
As for the legal/compliance stuff, I recommend you don't take any big decisions on that front without consulting lawyers. My understanding of that is that most serious companies in the EU have to take these topics pretty seriously. I'm sure in the US, hosting all your data and secrets in Chinese data centers isn't a whole lot less controversial.
The Chinese could of course choose try to match the current levels of investment Google, OpenAI, Anthropic, etc. are putting into local infrastructure. But as far as I know they aren't and there are probably a few political blockers for that.
Without infrastructure, their role is being a niche player in these markets. It doesn't really matter how good they are if they can't scale to most of the market.
I mean, Western providers such as Fireworks AI/Microsoft Foundry (US) or Tensorix (EU) already are offering many of these models on their own hardware with all the typical compliance boxes ticked through a standard API. Either as open weight models or through partnerships with Chinese firms, or both. DeepSeek etc. do not have to do anything here other than making their models available to Western partners (either as open weights or through a licensing agreement).
They'll still have their dedicated enterprise customers. I think the Chinese providers will pull more of the single users who're paying their own way, than those backed by company budget. And it's a pretty good split as the demand becomes better distributed, resulting in better service (I'll never forgot must how bad access to Claude became until they got access to Colossus) and less potential for lock-in (we really don't want there to be a duopoly, etc on good AI).
Just to be clear, enterprise companies are using the enterprise option for security/auditing/enforcement reasons, not because OpenAI or Anthropic forbids them from buying a 200$/month plan for each engineer, right? Do they have any clause forbidding using those individual plans in a company?
> I currently subscribe to the $100/month Max plan from Anthropic and the $100/month Pro plan from OpenAI. If you are a heavy user of coding agents these plans are a fantastic deal.
Guys, what - in your opinion - does "heavy user" mean? I thought I am heavy user (I am using AI to code every day 8hr a day + side projects) but 20 USD/month Cursor plan is always enough. What should I be doing to extend my license to higher level?
I'm out of credits codex pro credits (200 plan) in 4-5 days, and then I switch to claude (another large plan) mostly or just go out more and get Vitamin D. :)
I spend most of my time designing and tweaking tests suites, and improving test performance. These commits are almost entirely Codex: https://github.com/tsoniclang/tsonic/commits/main/ - but it's possible only because there's a very large test suite attached to it.
All of that is very token intensive. If OpenAI gave me 3x my limits, I'd find ways to eat it up in a week.
What do these tokens give me? Well, I think in a week or two, I hope to port the TypeScript-Go compiler back into TypeScript, but compiled to native code. It's probably not particularly useful for most ppl, but it's a hobby project that I've spent the last 7 months on.
My agents are easily busy 30 min to a hour independently. Implementing their plan, building and running tests, verifying deployment, linting. So I switch between multiple agents. Each agent has its own branch and worktree.
How many sessions are you running simultaneously, how careful are you being about token consumption and context window management, and what level of reasoning are you using (with which models?)
8hrs a day doesn’t really mean anything without a lot additional qualifiers.
fwiw lately I’ve been straddling 2 or 3 claude codes and one Claude cowork, primarily on 4.7 with high effort - the company’s paying for it, so I’m doing my best to burn as many tokens as I have the mental capacity to manage. At that rate, the 100 account is completely necessary, I was blowing through my 4-hour limits consistently before requisitioning an upgrade.
Ai has become indispensable but maybe not at all cost. My company just had a company-wide meeting to talk about how they're restricting who can use which models and instructing us the "be more responsible with company's tokens". And it's not an small company by any means.
The customers of these tokens need to see returns on their projects that exceed the cost of financing.
Laying people off only goes so far.
If enough said firms don’t see enough value given the price of frontiers they will cancel and consume open source. This is the risk the frontier labs are exposed to.
Unless your insider is the CFO, I wouldn’t trust these sources to have the access, knowledge, or insight to determine whether they’re running inference at a profit.
Simple test: can they get their hands on a data center contracts and financials?
Search isn’t anywhere near as high profile as the profitability of AI inference, and yet, even aspects of the Search org were walled off from the rest of the company such that other employees couldn’t see what we see.
> As further evidence that enterprise agents represent product-market fit for these companies, consider their open job listings.
PMF is one interpretation, but it could also be read as desperation.
In my opinion, we've been at PMF for quite a while now. The November inflection point that's often referenced definitely changed how we interface with models, but as far as coding goes, I feel like Cursor had proven itself useful for at least a year prior to that.
The demand has always been there, the outstanding question is still - how do you build a business on top of these products? None of the frontier models have emerged as uniquely capable, but open weight models are now catching up in capability as well. The explosion in go-to-market roles feels more like an attempt to lock customers into contracts so that they don't consider alternatives.
I assume the hope is that during this 12-month contract they will develop real integrations, something deeper than just a CLI harness. If you've ever worked in procurement or dev tooling at a reasonably sized company, you'll know that this is exactly what teams try to avoid.
It's anyone's guess what will happen this time, but I'm excited to see how the IPOs go.
TL;DR Ed argues that the deal between Anthropic and xAI could have been negotiated in such a way as to make Anthropic only appear profitable during its “ramp-up” period in June, which incidentally is also the month that Anthropic is making tons of other pricing changes.
It’s funny how a US$30 ReSharper license that would give me a 20% performance boost in programming was out of the question, while me burning US$3000 per month on tokens for a 40% boost in performance never was questioned
I feel this so much. What was all that fuss about needing 3 levels of approval for any software license? Or having to provide a compare chart against open-source alternatives?
The money would be so much better spent that way as well, supporting individual programmers.
> but I suspect there’s a more important factor here: I think they’ve finally found product-market fit
Ahhh the classic startup term that's definition is nebulous. But also, since when does any definition of product/market fit mean a product is profitable? And profitable in what sense? Unit economics? Overall company?
Oh I'm absolutely taking advantage of the fact that "product-market fit" has a bit of a nebulous meaning here.
It's a great hook to build an article around. My core point is more that April 2026 was the point when Anthropic and OpenAI finally appeared to have figured out a credible business model.
> My core point is more that April 2026 was the point when Anthropic and OpenAI finally appeared to have figured out a credible business model.
How so? What's specifically changed? We still don't know what their unit economics are and everything you've documented is basically speculation at this point.
1. Both Anthropic and OpenAI significantly increased the prices of their latest models. They're clearly not trying to offer the lowest-price-possible to drum up demand any more.
2. Both Anthropic and OpenAI no longer let enterprise companies buy discounted almost-all-you-can-eat subscriptions. Those big enterprises are now paying full API prices.
3. According to reasonably well-sourced leaks, Anthropic may be about to have their first profitable quarter.
And I didn't even say "profitable", I said "credible business model". I think getting companies to spend hundreds of dollars per month per seat, WITHOUT crazy subscription discounts, is a credible business model.
Let's see how it goes in three months. I'm not convinced it's anything but an attempt to make the numbers look better just long enough to IPO.
Enterprises will take a while to adjust. We're going along with Copilot price changes in the short term "just to see what it'd look like". They will get customers to continue with the all you can eat mindset at API pricing for a while, but the finance departments will start asking questions very soon.
I suspect it might end up being a credible business model for whoever buys the IP eventually (Microsoft, Google?), but I think OpenAI and Anthropic are so far in the hole debt wise that simply having a credible business model won't be enough to save them - it needs to be wildly successful.
Does this analysis factor in potential caching of tokens on the server side? It seems that if they organize things well (as a model provider), they can save quite a lot on that. Looking at my Cursor statistics makes it clear that the token calculations are not at all trivial.
money talks, and being a broky from outside the valley/west coast, the "product-market fit" here are the neighbours of the service providers.
the economics simply don't work unless you make six figures, at least to just give it a go blindly. the providers are also still figuring out what they can get away by charging, and they are getting a similar treatment from those under the stack.
the caps and limits are not very transparent, and it is quite difficult to know what is "enough". the current rate does not stay the same and the contract is changed way too often to dedicate for the long term. regardless, the subsidized rates should not be sustainable forever. make hay while the sun shines i suppose.
I thought the links I provided were pretty solid. A lot more solid then you'll see in most commentary about this stuff!
I agree with this person, let's use AI psychosis for when using an LLM gives someone psychosis, not for when we think, what, that a blogger made some poor assumptions?
If nothing else this blog did give me the idea that I should split my $200 claude max plan into two $100 CC max and $100 codex plan, esp because Claude is now offering 1.5x weekly limits so its the 5x usage is now more like 7.5x usage.
He said, so many years ago, that there will become a time where computing power is so prevalent that we will stop using the person to make the computers job easier and start using the computer to make the humans job of interfacing with it easier.
But in this context, it would mean the other side of increase productivity is decreased time to do the same work. These are the same thing.
Love how everyone boasted about replacing all the software with ChatGPT and then we end up with coding agents meaning the software engineer are STILL important. The sell is the development tool. It's classic cloud. Where did all the ops people go, many got subsumed by the cloud companies YET every company still has DevOps people to manage cloud infrastructure. The layer of abstraction went up but we still need the people to write the glue code and understand the business. OK great there's a new cash printer in the room. There's a new tool. Let's just start to ground the tooling in its new found gravity, profitability and IPO market dynamics... Reality has set in. The hype cycle is about to explode... Do you remember ride hailing and just how much cash was burned on credits pre Uber IPO. Then remember the IPO itself? These companies are not the new Google. They are a layer on top. Google was still the most efficient cash printing machine in history beyond the the US government and might still be. Will be interesting to see what the trillion dollar IPOs turn into. I'm going to say we see those prices get cut to a third in less than 5 years and scale back up over the next 15-20 years.
I've been calling that out for a couple years now. LLMs best and most viable use case is still just as a dev tool. Even for non-programming tasks, I still get better results from the LLM if I instruct it to write code to do the task...look at Claude Cowork for example, it's everything I used to do with python myself. It's not really a novel capability, it's just using python & bash for automations that any sysadmin has been doing for decades. Yeah, that's valuable for a non-techincal audience but is it $1T valuable? I don't think so.
When has an IDE or other dev tool ever commanded a $1T valuation?
These things get lost in discussions because people conflate "overvalued" with "not useful." LLMs are useful, particularly as dev tool, but Anthropic & OpenAI are definitely way overvalued.
They certainly have, but it relies entirely on the assistant frame, which is a problem in and of itself for the trillion-dollar economics.
Anthropic and OpenAI have shown people want a tool for task offloading, driving predictable token consumption and justifying the math, so long as users stay in that dynamic.
However, knowledge workers using these tools daily are getting exhausted with them. Outputs come out polished but hollow. Talking to a frictionless, frame-completing model all day drains you.
If user behavior drifts away from assistant usage because of that, per-token math implodes. The valuations we're hearing about all the time rely on usage compounding daily. The fatigue is a timer running against that compound.
Anthropic's Constitution is the closest hedge out there, I think. Installing an identity structure into the model through training. But it's still assistant-first, so the fix there is only partial.
I've spent the last year running a product that flips the architecture so identity is primary and the assistant role is secondary. Same frontier models, completely different conversational quality. The fatigue property doesn't really show up.
Whichever labs figure out how to install real identity natively in the weights are going to be the ones with PMF in the next phase.
If they’ve spent all this cash, and all we have to show for all this AI baloney are crappy coding agents that constantly make the same mistakes no matter how much you try to guide them.
All the slop content, all the bots, all the misinformation and fake AI images and videos.
All of the social and economic disruption from datacenter buildouts.
The massive nosedive in reliability on the world’s software infrastructure.
After all of that and all we get is a code bot so a few incompetent loser devs can bloviate about not writing their own code and brag about never reading it.
Burn it all to the gd ground. Destroy this new Tower of Babel.
Well, good for them that they are charging enterprises API rates. Why in the world not do something similar for consumers? Use for free a few times a day, have a $5 dollar plan for light use, and perhaps $10 to $15 for heavier use. If 90% of consumers pay nothing then ‘drop them’ except for letting them have an account and a few queries a day.
It is easy for me to change providers. Right now I use the open source Claud Code harness with two paid API venders for DeepSeek v4 (flash and Pro). I like seeing how much each session costs.
PMF is such a bullshit term. There’s no precise definition. It’s bs.
The fact is that investment is at a scale so large that current trajectories are nowhere near going to provide ROI.
The model companies are trying to milk every drop of ARR they can before IPOing. That is entirely the current narrative. You can hurt ARR before it churns if your goals are short term.
Anthropic being profitable is laughable. Sure, by some accounting measure that no one seriously uses. But looking at revenue and what they’re raising you can see the true story.
The truth is we need a revolutionary step up in capability to justify capex spend. It’s possible that might come - Opus 4.5 brought one - but failing that we’ll see the bubble pop once the IPO pumping is done.
And I think that's amazing. I'd like to keep using the subsidized coding tools, especially Codex, since I've given up on Claude. Hopefully the PMF allows the subsidy to continue. Would hate to have to move to the next coding harness again.
I think the reasons for them going with API pricing will become abundantly clear when the S-1s become available. If they don't have a story covering how they can get revenue closer to expenses, then they're relying on the market to believe the pixie dust version of their profitability story, which I think people increasingly don't.
The end game here is going back from a model where a bunch of product and tech management people sit in the U.S. or Europe, and try to manage thousands of mediocre talent sitting somewhere far away. The new model is you give those coding tools to good engineers colocated with your product people, and you ship good stuff much faster. If you can achieve such a setup, the token costs can be $50k per seat per month and you still run circles around the legacy IT models in terms of efficiency. Giving everyone the API keys and not changing the way products are managed is not going to work.
Good lord, what company would want to spend 600k per employee just to go maybe %20 faster (what the studies seem to show is a realistic estimate for productivity gains),
I'm building a product right now with some AI coding (despite my negative sentiment about AI in general they are useful). I am both the product person and the engineer, and I'm pretty decent at using it, so according to the hype I should be seeing like a 10x speedup. I am not seeing that. It's definitely faster, but there are also days where I'm stuck cleaning up things after going too fast for too long, or periods where I need to put the software in front of people to get real feedback, or even periods where I just need to use it extensively myself to find the pain points and bugs. I just don't see this "running circles" once you get past an MVP and you actually need to build something secure and not embarassingly broken.
To me the question is, can the frontier labs make the variance of output lower + make the output of higher quality to justify their prices?
If not lower priced chinese offerings will be better as its cheaper per token - giving you more attempts to offset the variance.
My feeling on the former is no... I believe they tried really hard but they've settled on pure marketing now to attempt to fight off the chinese with perceived superiority in quality.
I think it's fair to say they had achieved product-market fit when their revenues were growing deep triple digits month over month. What we're seeing now is that perhaps they have achieved profitability or at the least a more sustainable balance sheet.
I see two really good ideas for monetizing the free tier for consumers.
Firstly, if the user is asking for things where AI can link to products or services to buy, there's a very good relevancy, much higher than in other types of ads.
Secondly, since the AI often takes time to compute answers to user's questions, they could be shown ads while waiting. People could perhaps be less annoyed by this than some other commercials since they know the break has to be there anyway.
(First idea is something I came up when asking Claude to compare some products, or ask for help in lawn care. Second idea was by a colleague.)
Not sure I can agree with this. Sure, we have enterprise paying, but are we actually getting anything new and interesting made from LLMs compared to if we just didn't have them at all?
Maybe acceleration in smaller teams. We still seem in the era of the early internet where what questions LLMs change hasn't exactly emerged.
I think the product market fit is still how you interface with increasingly powerful AI. Right now is Codex/Claude code, and they just added computer use/web browsing into them. The product is the speed at which you can interface it and to be able to serve it with infrastructure.
Traditional product market fit isn't really applicable with tech evolving so fast.
So it largely sounds like many more people will be able to write software - and will use AI to do it. Existing software engineers will continue to automate their tasks away like they always did, but perhaps at a faster rate.
The impact of AI in other fields seems to be muted.
I think it is applicable to a much wider range of knowledge work, but it's also harder to apply there.
Software development has the huge advantage that mistakes and hallucinations are very easy to spot: the software works or it doesn't.
Spotting errors in a research report or legal brief is a whole lot harder!
But... non-software professionals spend a huge amount of their time on tasks that can be safely automated - reformatting documents, extracting numbers from PDFs, all kinds of flavor of data entry.
Learning how to use a tool like Claude Cowork can take a big dent out of those.
> Software development has the huge advantage that mistakes and hallucinations are very easy to spot: the software works or it doesn't.
Do we not care about code quality, maintainability, performance, extensibility, or understandability anymore? Honest question, not a gotcha, it's just previously getting software to pass all the tests was a small part of what we would consider "working" or perhaps "good" software. Maybe that's different now with LLMs, idk. Maybe we need automated checks for these things as well, like not compiling until the code quality is good enough to let the agent finish it's loop.
What code quality even means is different now, but also LLMs are capable of producing better quality code at scale in my companies experience. We are able to in fact sort of propagate best practices and structure via the llm to all of the teams even when they're working under time pressure.
If the AI can write code for robots the impact in other fields may be pretty large. Seems to me a lot of jobs can be automated with software and robots combined. The limit in the past was writing the software to get the robots to work. But if AI can remove that limit...
"[would have spent] $1,199 with Anthropic, $980 with OpenAI"
How many tokens is that, input/output-wise?
(a) I'm curious if you feel like you got $2000 worth of value out of them in the last month?
(b) I'm also curious if you would have gotten similar quality out of a slightly lower-cost provider of an open-weight model? (e.g. Kimi K2.6 and DeepSeek v4 Pro) and what the spend would have been for that.
I myself have managed to spend not quite $4 on OpenRouter and have felt it was very worth it; I just have much smaller, or more targeted requests I guess. (Lately, adding features to a static site generator in Python, or setting up log forwarding via a docker compose file)
Input tokens: 52,545,485
Output tokens: 5,767,253
Cache create tokens: 5,112,029
Cache read tokens: 1,475,069,465
Total tokens: 1,538,494,232
Total cost: $1,199.79
OpenAI Codex:
Input tokens: 52,598,013
Output tokens: 4,681,867
Reasoning output: 2,091,063
Cached input tokens: 1,153,844,864
Total tokens: 1,211,124,744
Total cost: $980.37
I'm confident I got value out of OpenAI - I've been mainly on Codex for the last few weeks.
Not so sure I got that value from Claude, just because I've been using it a lot less and somehow the price came to about the same as OpenAI.
Given the code I've been able to build in the past month I genuinely do think I got value for the API price version, and (don't tell OpenAI or Anthropic) I think I'd have paid full price.
I've not spent nearly enough time with GLM-5.1 and co to compare, but I do know that the prompts I'm using with the agents are not prompts I would have expected to work just three months ago.
If it were me I'd be asking "How long would it have taken me to do that, and what's the rate I'd have been charging for the work I would have been doing otherwise?"
Personally, I've probably spent $60 or so on OpenRouter in the last month or so and got a working project out of it that it would probably have taken me a fortnight to knock together (which is inevitably an under-estimate because it covered things I'd have to learn but K2.5/6 already knew). There's an orders-of-magnitude gap there.
This is great analysis but my first reaction was - is this trolling?
The fact that we have to think about whether a $1TN company has achieved product-market fit both articulates 1/just how high the valuations are 2/How hard it is to pin down EXACTLY what PMF is.
As a pre-revenue startup, I am laser focused on PMF and frankly if this is the bar, no one will achieve it. But OTOH its heartwarming that people are willing to value you at $1TN before you reach that. Guess everything is in the eye of the beholder?
The title was indeed meant to throw subtle shade at the idea that supposedly $1tb companies were only just now starting to figure out product-market fit.
With respect to Simon, whose writing I've usually agreed with in the past and whose insights I've liked: this is a bad take that overlooks the extent to which corporations are imposing the use of AI on employees, and in particular ICs, who make up a majority of the AI-using workforce by headcount.
Many of us are either openly having our performance reviews tied to AI use, especially at larger enterprises. Whether that's measured by sheer token count or just "how many of your tasks are you using AI for these days" (combined with the implication that question carries at many orgs which are heavily invested in AI).
Are you saying that Anthropic's huge leaps in revenue are caused by stupid company policies and token leaderboards, and the moment companies stop imposing AI on their employees revenue will drop to a point where Anthropic are unlikely to be profitable?
I don't think that's the case. I think the token leaderboard thing (which is clearly ridiculous) affects a tiny portion of companies and is already going out of fashion.
I'm saying that the truth lies somewhere in between, and that Anthropic's current revenue is being, in part, propped up artificially.
We're also in a place where a lot of the usage guidance around these tools is still nascent. People are cowboying a lot of stuff, even as larger companies start to organize AI policy/safety/responsible use working groups to try and policy around the shortfalls of the technology.
IMO: if this technology persists, and if we figure out a way to use it in a broadly safe way, the value proposition will probably trend down rather than up, at least on the code generation front.
As a research tool, it shows some promise, though I still find the ethics of the technology disgusting.
Is PMF enough. It is such a dynamic self-disrupting wave that it is like predicting physical chaos. These aren't early Googles in a blue ocean. Maybe a blue ocean full of pirates and dragons!
This isn't me being a doomer I just don't know. Can we look at Q2 profits and draw hockey sticks yet?
Remember people are boasting how much their expenses are. That is where we are in the bubble/new paradigm.
These folks have no lasting moat and they know it. We are still close to the November 2025 inflection point, so they had a clear advantage during these past few months. That will soon fade as open-weights models become really good, which is arguably already the case with DeepSeek V4.
Article doesn’t mention on-prem and on device models. Almost guaranteed that there are a range of killer enhancements on these fronts waiting to drop until IPOs get closer to inflict maximum chaos on the valuation games.
While the big guys will argue they’re worth trillions expect others to drop chaos booms showing their NPV may be effectively zero.
Operating profit is both post depreciation and fees paid to third parties for hire. So aside from shenanigans like RSUs and financing interest that's already somewhat close to actual economics.
Meanwhile we've got commenters here talking of 5-10 trillion with a T revenue shortfall.
> If you are a heavy user of coding agents these plans are a fantastic deal. I just ran the ccusage tool on my laptop to get an estimate of how much I would have spent if I were to pay for API tokens in the past 30 days and got
You think this is fantastic deal only because they use similar like tricks where they inflate the price and tell you something supposed to cost $1000 but they have this today promo for $100.
I was there too and paying for a while. Few weeks ago I tried DeepSeek V4 Pro - expected its gonna be shit but its actually pretty good.
The deal is I pay daily ~$1 for DSV4-pro for ~100M API token usage. And they probably not getting broke because >90% of those token in practice is cache read and they very well optimized for that.
Yep, exactly this. And I have so much less anxiety that I have to use my 5-hour/weekly usage or I lose it... with deepseek api the credits never expire, I can use them when I want, how much I want and the prices are ridiculously low for the quality/intelligence/performance.
GPT 5.5 is maybe 4x the size of v4 pro, hard to compare price since cache hit is basically free with Deepseek, but 40x cheaper (with their 75% off) seems about right.
I wonder how a focus on per-token API profits will impact the incentives to improve token efficiency and drive down costs through optimized compute. I suppose as long as a few leading labs are competing, we'll see progress in this regard, but it's certainly less in their interest than it is with a flat subscription pricing model.
There's also increasing pressure from China: note the massive permanent price reductions DeepSeek and Xiaomi announced in the past few days, with the possibility of more around the corner. And there's also the constant release of increasingly capable open weight models.
Mentioned in the article, but it cracks me up that both openai and anthropic are utilizing fairly traditional enterprise GTM plans segmented by verticals.
So many startups trying to automate sales, but somehow the two biggest frontier labs have decided that the best GTM strategy is firmly human-in-the-loop.
Product market fit for the models, sure. But their cost structure is going to kill them as token costs drop this year, following DeepSeek and Xiaomi and soon more providers.
Great article I know this upsets a lot of people who are used to thinking Anthropic/OpenAI are just lighting cash on fire but they've cornered the market on enterprise who cannot walk away from these $200/month plans
However the valuations are still far far away from actual sanity
It's good enough for personal stuff. It doesn't compare to the latest Opus I use at work. You can certainly argue I don't need Opus for work, but there is clearly a difference.
Also, at least with z.ai, GLM-5.1 is s l o w! After using Claude at work, I get really impatient with GLM-5.1 at home. When doing "true" vibe coding (i.e. not really examining the code), Opus is a ton faster (easily 5x).
But yeah, I'm not willing to personally pay for the frontier models. I won't even renew my annual Z.ai plan - it's become too expensive.
Hmm, I use opencode subscription, and glm seems just as fast from the tests I've tried to compare between the two. Tbh it mostly took Claude longer (mostly significantly longer) for the same tests.
Also, and I know you may not want to answer. But could you give me an idea of the type of thing you found glm to be worse with?
I think I've been fairly unbiased in testing a bunch of different development tasks. But am curious if maybe it performs well for some stuff and not others. So if you could share what you feel it's worse at.
Also are you an experienced developer or less experience?
I'll repeat something I wrote on an entirely separate HN submission.
When DeepSeek V4 Pro came out, I had been mostly coding with GLM-5.1 on a Z.ai coding plan.
I had a large analysis task on a relatively complex codebase. I decided to try the models out.
GLM-5.1 did acceptably but got a few things wrong (easily corrected) and took quite a while to get there.
Opus 4.6 burnt through the US$10 budget I had given it in about 10-15 min, without ever returning from the first prompt.
DeepSeek V4 returned a full analysis within 2-3 min, and I carried on all the way to implementing the feature I was after. Total cost less than US$1.00.
I now mostly alternate between GLM-5.1 and DeepSeek V4 Flash, with an occasional dip into V4 Pro for more complex analyses.
task i am working on right now at work is comparing two verisions of apis and documenting responses in their outputs. i suspect a vast majority of work at entrprise is of similar complexity.
right now everyone is using latest and greatest to do dumb stuff like that. that would change fast if companies start caring about costs.
> enterprise who cannot walk away from these $200/month plans
Any org with more than 150 users aren't on $200/month plans, they are forced into API pricing + $20/month/user
For individuals and orgs small enough to get to use the subscription plans, that's all well and good until usage limits keep going down, or cost goes up. If you compare the usage you get on $200/month maxed out vs. what that would cost at API pricing, the $200/mont plan is an absolute steal. I doubt it will last long.
Not to mention the API plans are also still in their "lose money, just get the suckers hooked like addicts" phase. Once the reality-based pricing comes into play, it's a coin flip of whether the bulk of the companies fail, or they get to live off government subsidies for a few decades.
On the plus side, I'm happy I'll have a nice hay barn when the local half-built AI data center is abandoned.
> There exist a large number of people who are absolutely convinced that LLM providers are all running inference at a loss in order to capture the market and will drive the prices up sky high as soon as everyone is hooked.
> I think this is often a mental excuse for continuing to avoid engaging with this tech, in the hope that it will all go away.
Thanks for that psychological explanation. I was wondering why people were simply ignoring the math that shows that inference at API pricing can be quite profitable, e.g. published here for DeepSeek V3/R1 with 545% profitability: https://github.com/deepseek-ai/open-infra-index/blob/main/20...
There isn't a single thing about how the AI companies are operating that looks like a normal business. I know people who were in the room when Scott Sullivan, CFO of Worldcom, assured everyone that the future was bright at Worldcom days before they collapsed. So you'll have to excuse me if I don't believe the words of someone whose sole job is to justify hundreds of billions of dollars being thrown at Anthropic when he says their future is bright.
He is talking about tokenization of different llm and comparing, and this is inacurate because you cannot compare if you dont how it tokenice the words, some tokenicer could be cheap with more tokens than other because of their features. Not agree with Simón speech.
My employer went from the Max plans to Enterprise this month which was utterly stupid. We went from paying $200 per head to like $500 for some people, even more for others. For the same product. Oh well, guess we’re doing our part to prop up this bubble.
Companies are kool-aid drinking now due to hype, but given how much they're spending, if they don't see REAL, BIG wins from it soon, they're going to scale it back quickly and switch to Chinese models. Claude isn't worth the API cost for a lot of development work, and once companies have had time to collect and crunch data they'll see this.
There's no change in tone, I'm still very bullish on the tech, Claude in particular just isn't worth the API price, which I've always felt was too damn high. I have paid for Gemini 2.5 Pro, Deepseek 3.2/4 and GLM 5 tokens happily though.
You financially benefit from stuff like agents. Of course you will be the last to admit publicly when things aren’t quite heading in the right direction. The gap between hype and reality is ever increasing.
You misread my comment because of your personal bias, and now you're acting like it's some sort of own? Come on bro.
The models are putting in work, that's indisputable, the thing I'm calling out is that Anthropic's price and hype are running ahead of the value it's delivering. Agents are a great boost at <=5% of SWE salary, a mediocre tradeoff/risk at 20%+, and insane at SV tokenmaxxing levels.
I know everything you’ve done for the tech community, but I please you to take some time off and reflect on this article. It’s not on par with ur usual level, but the tendency has been visible from the last couple of articles.
I thought this was one of my best pieces of writing this year.
(In case you missed it, the title was meant as a subtle burn on those two companies - it's pretty absurd for them to only just be finding product-market fit when they're already valued at over a trillion dollars.)
- The first argument about your subscription price, doesn’t has anything to do with the overall claim of the article. It would if at all be a weak argument to support the opposite. Subsidizing prices signals a lack of PMF.
- That you hire sales people after you had a billions of funding is nothing surprising and doesn’t indicate PMF or not.
- AI Implementations are fresh and of course AI Failures are thin, but so are AI successes. I haven’t seen any companies creating billions of shareholder value because they’ve massively invested in AI and their competitiors didn’t You really can look at these things in 5 to 10 years and it is multi-faceted including cultural acceptance.
- That they need to buy more compute to satisfy the requests is probably the strongest argument in the artical, but don’t conclusive. The product is been sold heavily subsidized and in hype cycle. And again both OPENAI and Anthropic have to show growth in order to justify the IPO.
- Regarding the part about revenue I refer to the linkedm article above, as it does explain it very well.
The conclusion is reasonable given the arguments, but not the title.
However it is missing all the real discussion points that are actually in observation at the moment.
Local models as alternatives, IPO finanical engineering, how AI implementation actually will perform over years... Let’s all not forget crypto. It’s been full of "use cases" just a bunch of years ago. I like the idea of crypto(btc,eth) and I’m still invested, but 99.99% of coins have died on promies.
So this is not a piece of critical thinking, but this reads like a twitter thread to sell me a course :/
>Somehow this fragment turned into headlines like Uber’s COO says it’s getting harder to justify the money spent on AI tokenmaxxing, because the market for stories about AI failures remains enormous.
I notice this all over the place. Many people hate AI and want it to fail, and they're willing to invent misinformation if it supports that idea.
That "one" person is the COO of Uber. And the other one - the one based on whose statement about burning through yearly AI budget in the first few months - the whole discussion sprung up internally at Uber in the first place is the bloody CTO of that huge company. So yes, their words do have A TON OF WEIGHT. Thats why they are in such important positions, arent they? They're not quite the Derek from the pub, casually commenting on how Liverpool will fare this season.
I think the way people reacted to those statements was entirely out of proportion to what was said.
I repeat: a CTO saying that they spent their entire AI budget for 2026 when that budget was clearly set in 2025 before anyone knew what those November models + harnesses were capable of is entirely unsurprising. Any analysis that doesn't also point out the difference between 2025 and 2026 era coding agents is either ignorant or deliberately misleading.
Yes, but that's irrelevant, because the COO uses that to base his core argument - that all that jackshit 1800 code changes per week that the CTO boasts about, mean absolutely nothing in terms of value. It means they are spending a lot on it, to gain as he diplomatically said "perhaps 20% more" - and I wonder 20% of fucking what - it's a ride-sharing app, what could they be possibly building on top of it with all that token crap?
You have to try pretty hard to get to "all that jackshit 1800 code changes per week that the CTO boasts about, mean absolutely nothing in terms of value" from what he said on that podcast.
(We still don't even know what Uber's planned AI budget for 2026 was. They didn't reveal that when asked - in https://www.theinformation.com/newsletters/applied-ai/uber-c... it says "He wouldn’t disclose exact figures of the company’s software budget or what it spends on AI coding tools").
I don't have to try at all - I think anyone who spent as much as an internship, let alone years at a modern tech corp would have no trouble distilling the absolutely clear message - we are spending too much for too little value. And actually wtf am I explaining myself? Every major tech outlet interpreted it like that too. It's not that hard Simon.
I think Business Insider are not a particularly high quality publication - they're prone to clickbait - and that headline is an example of why I think that.
Hah, I just checked their homepage and here they go again amplifying that COO fragment from that podcast:
> "That link is not there yet, right?" Macdonald said in comments that went viral, racking up over 2 million views on X. "I think maybe implicitly there is more that is getting shipped, but it's very hard to draw a line between one of those stats and, 'OK, now we're actually producing 25% more useful consumer features.'"
Yeah, something going "viral on X" is clearly a sign that it's quality information!
That's why I wrote a whole section about it. I was deeply disappointed at how little thought people had appeared to put into this before amplifying the misleading headlines.
Who's to say those enterprises won't churn after XYZ comes out with a decent enough model that costs 10x less to use?
There's a whole bag of clever tricks you can play to juice short term results leading to an IPO that may not work longer term.
I'll believe they've found product-market fit when they have a product. Right now they're selling the infrastructure, in a highly subsidized and undifferentiated way (at least over a sufficient long period of time of, say, a couple of years).
I think this was obvious since the birth of ChatGPT
Intelligence is a universal good, it can apply to anything, and no, "human intelligence" is not the only form that is useful nor special. There are limitations to AI but also huge advantages, and its obvious that the advantages are worth paying for, given their revenue.
The article is at least one year too late. Claude Code has been the product-market fit. It's so obvious retrospectively that questioning it at this point would be very silly.
The problem is not whether they have PMF (they do) but how they're going to compete against on-prem and Chinese providers.
Having PMF != printing money forever.
The author claim:
> That’s $2,180.16 worth of tokens for $200
No matter what it means, rebuild the same thing you built with these $2,000 tokens with DeepSeek Pro V4 and let's see if Claude has a chance to survive.
I think we are overly fixating on Anthropic and OpenAI.
Since there are lots of models that are competitive and have a much better pricing, both OpenAI and Anthropic seem inefficient. I don't get why someone would want to buy shares after IPO apart from fomo and artificially built enthusiasm.
Anthropic and OpenAI may well be the Altavista and the Yahoo of the AI age.
I'm a huge fan of agent coding but kinda dislike this "llm evangelism".
There are still several open points (eg.: code churn, maintainability, subtle bugs human will never do) that everyone with a minimal programming knowledge that seriously used a LLM agent knows about but somehow none of these "big influencers" never mention (or just saying "it's your fault").
You mean that HN has been predicting AGI will happen, or you mean the Singularity happened earlier today and I missed it that humanity is already all killed?
They've finally forced the system to fake profits. Attacking personal computing, sucking the rest of the industry dry, yeah eventually you can make people pay for your shit. I still pray these vultures will implode without too-big-to-fail safety nets. And big daddy Trump can't help them anymore by fixing the system for them, no matter how much billions they are paying into his personal pockets.
The future are small models, nobody really needs big compute in the long run, that's why big tech is going for our personal hardware. So we won't become their competitor in their rent only economy. True competition is eliminated, natural evolution is being fixated by the government. This is not going to end well for the USA.
Sounds like mostly-baseless praise for the LLM behemoths / capital investment black holes.
> I currently subscribe to the $100/month Max plan from Anthropic and the $100/month Pro plan from OpenAI
... which already indicates a bias.
> If you are a heavy user of coding agents these plans are a fantastic deal... that’s $2,180.16 worth of tokens for $200—not bad at all!
Thank you for the sales pitch. Perhaps go buy a car and tell us how [insert your manufacturer] has "found product-market fit"?
> Anthropic are strongly rumored to be about to have their first profitable quarter.
First, the strong rumor is a claim by Anthropic itself. But even assuming that's true - it's an "operating profit", i.e. disregarding the massive capital expenses for years, and may also disregards ongoing capital expenses, if those happen not to be taken this particular quarter.
> 1 Trillion .. companies spending $200+/month/user will get you there a whole lot faster
First note the use of the first person plural to talk about Anthropic and OpenAI.
But that aside - most companies aren't paying $200 USD/user-month. But even if they were - if we take the 30 Million SW developers mentioned in trjordan's comment as subscribers, that's 2400/user-year * 30 Million = 72 Billion USD / year. And this is already rather optimistic, but - want to double that number of subscribers? Fine, make it 150 Billion / year. Still not there with a rosy outlook and assuming the hype and enthusiasm continue for many years.
And those rosy estimates are more likely than not unrealistic. I am reminded of this review of some empirical research regarding the benefit of LLM/AI use:
The big assumption with all of these sorts of analyses is that things will continue as they are for the foreseeable future.
In hype-driven markets, you cannot be certain of that.
Let's take a view that the author is right: coding agents and their associated harnesses were the inflection point for some degree of profitability and widespread consumption, and that these tools are now yet another SaaS subscription or API bucket expense to bake into every single developer (or developer-adjacent) in the organization alongside your collab suite, HR seat, CRM seat, design seat, etc. To be fair I honestly think that's a safe assumption to make for highly technical firms whose image is derived from remaining on the cutting edge of things.
That begs the following questions, which we won't know until IPOs start happening:
* Are subscriptions profitable, or just API consumption?
* What's the run rate when we just consider subscription-based usage like Claude Code and Codex? What about API calls?
* Is there any profitable pathway forward at which enterprises can get unlimited usage but at fixed rates via subscription?
* What does customer churn look like for subscription users versus API users?
We also have a number of questions for customers that I suspect we'll start seeing receipts for in the coming months, at least from the early adopters:
* What was the net gain (loss) from leveraging coding agents?
* What's the cost of a developer with or without access to a coding agent + harness? Is it cheaper to hire an outsourced worker with a coding agent subscription, or a domestic worker without one?
* At what point does further AI spend result in diminishing returns, i.e. where's the 'sweet spot' for spend?
* Did AI boost actual revenue and outcomes, or did it just gamify KPIs?
* What roles or work did AI actually replace, versus merely displace during the hype cycle?
Not to mention the questions regarding the technology itself:
* Will we develop the means to run foundational/frontier models at edge using less resources through some existing (e.g. distillation) or new technology, thus cutting off the profit centers of these firms?
* When the market mismatch between supply and demand is resolved, won't it be more affordable for consumers and companies to operate their own AI infrastructure rather than support further centralized buildouts?
* Will coding agents improve to the point of being able to bootstrap and self-orchestrate on edge/consumer hardware without substantial technical expertise, or at least improve to the point that traditional IT teams can securely operate them internally without an expensive subscription or API token bucket?
All of these will influence the long tail of this bubble, because it is a bubble at this point. Even if these companies are indeed profitable thanks to the coding agent inflection point, there's still so many unanswered questions about utility beyond coding that it's impossible to extrapolate a future. If coding agents are indeed the extent of utility for profitability, then there's no possible way these entities will recoup the investment already sunk into their infrastructure buildouts. Even if more profitable uses are discovered, does this offset or replace the firms disappearing due to AI speculation and their associated contributions to the economy as a whole (RE: the consumer compute industry at present, higher energy costs due to datacenter builds, opportunity cost from harms to local infrastructure from haphazard builds, etc)? Should these firms indeed be runaway successes and immensely profitable to the point of paying off their investors and growing the larger economy, does this end up stifling innovation in a world where most new ideas are fed into LLMs for R&D that are then controlled by only a handful of companies and immensely wealthy people, via systems that are easily surveilled and stolen from without recourse?
So many, many questions yet to be answered. Betting the farm because of coding agents is one hell of a gamble.
Wait till people figure out they can swap Claude code for DS4 Pro and spend a fraction in API billing (actually, significantly less than $100) while barely noticing a difference.
> Anthropic are strongly rumored to be about to have their first profitable quarter
No, its more like their own leak to WSJ and according to Ed Zitron -> seems to be heavily engineered via non-GAAP practices such as counting potential, but not realised revenue as actual revenue - the stuff for which I would be arrested if I did it at my company.
Also it appears according to Ed's analysis - strangely they seem to be projecting only that one quarter as profitable - potentially to calm the investors ahead of the IPO. Investor fraud anyone?
Also it was but a few months ago that their CFO said, in a court filing, that Anthropic's revenue across the entire lifetime of the company "exceeds $5 billion". Pretty strange.
How is it strange? The "exceeds $5B" quote was from December 2025. Anthropic has seen tremendous growth since then, ever since Claude Code with Opus 4.5 got really good at coding.
If you've ever been at a startup, this is exactly what it looks like when you go from not having product-market fit to having it (though with a few extra zeros on the end compared to most).
All three of those are official releases from Anthropic. You can choose not to believe the if you like, but since they plan to IPO this year it's in their interest not to get caught lying to potential investors.
As a non-public company they can use whatever non-GAAP black magic accounting to claim anything while still not technically lying. It just doesn't correlate to anything we'd actually call revenue.
Not at all, because the magic can be applied differently at different times. They're undergoing a funding run now so they've got a massive incentive to come up with all kinds of revenue.
They've also signed a deal for billions in compute with xAI for april-may so they're certainly using that to fake billions in revenue using non-GAAP bullshit. It just seems a tad more likely than them legitimately increasing actual revenue by 233% in four months out of the blue.
Sorry man I hate to say this but you and many others need to stop commenting on accounting, finance and economics as its clearly way out of your realm of expertise.
Do you know what revenue recognition is? Do you know what accrual accounting is? Do you know of the phenomenon that is 'managed earnings'?
The only true objective number in finance is cash flows.
> Their revenue has quadrupled in just a few months
Maybe, maybe not. We haven't seen that S-1 yet. All we have is the 5B in lifetime so far. PLUS - revenue quadrupled or not, it only matters if their costs did not expand at the same rate or more. Revenue is not profit.
OK, so when S-1 comes out you will finally allow yourself to be wrong? Your prior is, a 1T company plans to IPO and their leader has been loudly committing an insane amount of fraud? I mean this of course is possible but that is quite the conspiracy. The scrutiny of an IPO would be a crazy thing to do if you were committing fraud at the scale you're suggesting.
Revenue is not profit yet the discussion in this particular thread is about revenue.
No, it's not. It's a stupid thing to say. Perfectly stupid assumption. There are 1000s of multi billion $ revenue companies operating and as a % the number that are fraudulent is close to zero, especially those public or looking to go public. There is always the possibility, but it's extremely naive to think it's likely.
Ed is a smart guy, but you or anyone basing your opinion on what one eloquent journalist says is ultimately a risky bet, no matter how much his reporting hits your particular dopamine receptors.
Please don't forget that Ed's entire brand identity is now 1:1 with exposing "AI" as a giant, unmitigated failure.
That's a very specific flow chart to hook your caboose to when none of this is even remotely close to endgame.
Pretty much. Ed does a lot of great work in digging through all this stuff but his conclusions always feel far too doomer oriented. OpenAL should have closed 5 times by now if you have been following his assertations from the start.
There will be big parts of what he says are true once the rubble settles but it will not be anywhere near what he is predicting. How that will shape out may not be great for the average person, what money shuffling tricks will be used? But it won't be a total wreck.
Honestly, I think it's very short-sighted to assume that all of this will be seen as any kind of wreck in the long term.
Normies are still catching up and reacting to chat-based LLMs.
HN types are further ahead of the curve, but still catching up and reacting to agentic coding and design workflows.
What often gets completely ignored is that entirely new modalities for how the underlying tech can be applied will continue to be demonstrated, and those will once again cause new ripples of excitement and disgust.
There are companies building world models and systems for protein discovery. Comparatively speaking, these approaches are barely in the zeitgeist today.
Deciding that we already have the data points we need to extrapolate how all of this plays out is like someone in 1974 deciding that microprocessors are just for accounting and inventory. Don't be that someone.
I think the big issue isn't so much a technology thing, I mean that will improve for a long while yet, but it is an economic one. My whole concern over the rapid expansion of LLM's is the massive build out on a technology that hasn't found its feet in a big enough range of markets yet that are willing to pay top dollar. Yes, world models for protein discover is very cool stuff and I kind of hate it gets lumped in with these other companies efforts because it has a very clear path forward that doesn't rely on massive IPO's just to keep the lights on.
This stuff is here to stay but I'm not sure how many of the current front runners will be able to stay solvent if they cannot turn these things in to massive money spinners. Revenue is fine-ish but spending is out of control. I see the debt in hundred of billions of dollars and start to wonder "Who is going to pay for this?" and "Will the people be willing to pay that much?". It just all feels forced rather than organic growth.
This is why I think Google may end up being one of the leaders in this field. They have their custom TPU's that seem to be fairly efficient at these tasks, they are slowly but surely improving their training and inference tech using their massive data set and most importantly, other parts of the business can subsidize this stuff for a decade if needed until it is genuinely profitable.
I am not against the industry but I do worry that many are rushing in with no means of genuine sustainability other than jump out for a golden parachute and let someone else clean up the mess.
I mean optimistically I hope there is a big reduction of hardware requirements in future. One of those technological jumps that nobody sees in advance but looks obvious in hindsight. If the need to the amount of compute goes down enough,that is a brilliant path through.
We don't have to take Ed's word for it. Anybody who's capable of doing grade school math can see that the numbers simply don't work. These companies are literally spending orders of magnitude more money than they're actually bringing in. Cursor, who've been renting Claude, estimated just recently that a $200-per-month Claude Code subscription could use up to $2,000 in compute. https://www.forbes.com/sites/annatong/2026/03/05/cursor-goes...
> According to a person familiar with the company’s internal analysis, Cursor estimated last year that a $200-per-month Claude Code subscription could use up to $2,000 in compute, suggesting significant subsidization by Anthropic. Today, that subsidization appears to be even more aggressive, with that $200 plan able to consume about $5,000 in compute, according to a different person who has seen analyses on the company’s compute spend patterns.
The load-bearing detail here is if that means $2,000 of internal server+electricity costs, or $2,000 if they were to charge at their API pricing instead of the subscription cost.
The latter is how I understand these things to work right now. If it's the former then yeah, Anthropic are losing a TON of money on those subscriptions.
Frankly, everyone in the industry knows. When people make these statements without additional clarity they always talk about API prices. You can look at the NVL72 specs and make estimates for electricity and ownership costs rather easily. Inference at data-center scale is dirt cheap, even with public codes using dynamo and sglang. The mystery is why the early misconceptions about inefficient inference persisted even after NVIDIA was very open about everything they did to help reduce costs dramatically in the last two years.
There's a good reason to look at it separately: if inference is profitable then they make money (or at least lose less money) when they get more customers, because any fixed costs are spread across more usage.
Assuming that there are infinite suckers with cash to spend. It's entirely possible (if unlikely) that the market is not big enough to cover the training costs. Especially for multiple companies all burning insane amount of money on the regular.
Depreciation is part of the cost of inference. Inference happens in GPUs that have a relatively short lifespan.
Those GPUs are very expensive.
Inference is expensive because a GPU can only process a certain amount of requests in a given timeframe. Remember that Anthropic is constrained in compute.
If they are constrained, it means that those GPUs are not idle. If they have more customers, they will need more GPUs.
If they have to play silly games using EBITDA to be "profitable", then it means that they need to ramp up prices a lot more than they already did.
Which is why in these discussions I always say that inference is also extremely expensive. Too many people like to pretend without any evidence that inference is cheap.
You can "just not update an LLM" in theory. But if your competition updates LLMs, and gets more capable, more efficient LLMs, and you don't? They get more capable "expensive tiers", and cheaper "cheap tiers" of LLMs. What are you going to do then? Bleed userbase and die?
I think the key thing that depreciates is all their models. You train one at crazy cost and 6 months later it’s worth $0. If you ignore that depreciation you look much more profitable.
Bubble popped when they increased prices. IPO may help cover some of the costs, but AI is very elastic and can be swapped out for any other company second to second. Which is why I think they bought up ram and disks like they did, to starve out competition and local models.
I must admit that I am going to find it fascinating when we hit the point where it becomes nearly impossible to deny the efficacy of these tools. I have straight up had people, even in real life, suggest that I'm lying about my productivity gains or what I'm able to accomplish with them.
Like, I understand the reasonable arguments against (I even agree with a few), but it's clear that some people have fully inserted their head into the sand and just don't want to believe any of this could be true. Which will be harsh, since I think getting hit with this train all at once in the future is going to be a rougher ride than a slower coming-to-terms-with, even if the result is one we're unhappy with.
Oh, basic counting is now arithmetic? But I was told they were superintelligent and were going to cause an apocalypse because they can do pretty much everything ? Somehow because they can excrement a lot of text, we were told they can do everything else too?
thats not that far off. Costs like $100Ms to train a frontier coding agent model today, billions if you count the full pipeline. Combine that with the infra we're building out, the fact that you have multiple labs building similar scaled models, the industry-wide costs of training frontier models could easily surpass 10B/yr in 2027
Their CEO claims a lot of wild shit. He claimed in January this year, that in about 2-3 weeks from this moment, i.e. "in 6 months" that AI will be doing all of SWE work. Lets hold these people accountable for a change!
> "in 6 months" that AI will be doing all of SWE work
I assume this is the quote you're referring to from Davos?
"I have engineers within Anthropic who say I don’t write any code anymore. I just let the model write the code, I edit it. I do the things around it… we might be six to twelve months away from when the model is doing most, maybe all of what SWEs do end to end."
that was in Jan, he said "might" and he said 6-12 months. Yes! Let's hold him accountable for saying reasonable things!
Reasonable things? He said the same shit over and over over the last several years. Yes, lets hold him accountable - you don't make such "oopsies" accidentally, several times in a row.
Seems pretty reasonable to me. Timescales are hard for anyone to predict. He is forced to do these predictions to know how much compute to buy in advance. Surprisingly, he underbought compute and now has to scramble to secure it from xAI or wherever he can. So he was overly conservative...
Indeed. That's why serious people are very careful, even if they are not running a company supposedly worth 1T USD
> He is forced to do these predictions to know how much compute to buy in advance
Ah well, that explains it. For my companies next quarter, I'll just pull some random numbers out of my ass so we can make plans with material impact to company business based on that.
> That's why serious people are very careful, even if they are not running a company supposedly worth 1T USD
10x revenue growth per year, even more this year...his predictions about when agents will claim SWE e2e work are his speculations, relevant because people care about what he thinks as he is closer than anyone to the leading edge of the technology. It's also important for him to be as accurate as he can about this because he has to put his money where his mouth is. He has to sign the right amount of compute otherwise he screws himself. He got it wrong in the opposite direction that you're implying, so at this point it sounds like you are more interested in your axe to grind than the truth on the ground.
You think enterprises are adopting CC because they think "oh this will replace my SWEs I can fire them"? That's not happening at major companies. They buy CC because it's useful and the writing is so clearly on the wall in so many data points that to suggest otherwise is a bit silly at this point.
> For my companies next quarter, I'll just pull some random numbers out of my ass so we can make plans with material impact to company business based on that.
You, as a leader of a company, don't have to make predictions? Don't have to make bets about what the best thing for you to do next year? That must be incredibly nice.
Amodei and everyone else need to plan compute and plan their products and roadmap. You want him to....not do that?
> You think enterprises are adopting CC because they think "oh this will replace my SWEs I can fire them"?
Yeah, that's actually Darios main talking point
> They buy CC because it's useful and the writing is so clearly on the wall in so many data points that to suggest otherwise is a bit silly at this point
Right, really sound arguments - writing is "clearly on the wall" and there are "so many data points". I'd be keen to use those immediately, but I am kind of missing the key of the "many data points" - namely, what did you build with it and how much ARR is it generating?
> You, as a leader of a company, don't have to make predictions
I have to make predictions, but not confabulations, lies and idiocies.
> Amodei and everyone else need to plan compute
FOR WHAT? Again, what was built with their shitty product in various companies and how much ARR did it generate? Uber seems to get no value out of it.
Anthropic has generated far more than 5B in revenue, I don’t know what sort of computer you have but it evidently does have the Internet, I would recommend using that unless the Internet CEOs are also in trouble for hyping that one up.
> Right, really sound arguments - writing is "clearly on the wall" and there are "so many data points".
Thank you for recognizing this. Don’t read Ed and think you understand anything about AI is all I’ll say. Read epoch capability index paper and look at the dashboard chart or the METR time horizon chart and methodology and then return with what I imagine from historical comments will be another ferocious and impressive act of mental gymnastics.
> I have to make predictions, but not confabulations, lies and idiocies.
Idk you’ve been misquoting and aggressively against addressing any facts you are presented with and yet bring no facts of your own (hint: if you know what you’re talking about typically you can calmly discuss with actual facts). That feels pretty similar to confabulations, I won’t say idiocy I’m sure you are not an idiot but you seem to have a lot in common with your caricatures of tech CEOs.
I work in big tech and probably 90% of code over the last month has been written by AI. And I suspect it's probably higher within Anthropic, which is probably what he's basing his opinion on.
So, he's closer to correct than not.
That said, your recollection is also flawed. It was in mid-March, and here's the relevant quotes:
>I think we’ll be there in three to six months—where AI is writing 90 percent of the code. And then in twelve months, we may be in a world where AI is writing essentially all of the code.
[...]
>But the programmer still needs to specify, you know, what are—what are the conditions of what you’re doing, what—you know, what is the overall app you’re trying to make, what’s the overall design decision? How do we collaborate with other code that’s been written? You know, how do we have some common sense on whether this is a secure design or an insecure design?
[...]
>So as long as there are these small pieces that a programmer, a human programmer, needs to do, the AI isn’t good at, I think human productivity will actually be enhanced. But on the other hand, I think that eventually all those little islands will get picked off by AI systems.
With another 3-4 months left on the clock, his prediction seems remarkably on point for at least certain organizations and domains.
I welcome you to also hold yourself accountable in the coming months if this trend continues. ;)
Yep! We have a review process where we have a few agents, each tuned to a particular domain of expertise (security, code quality, etc) which iterate until the feedback meets a certain threshold, at which point it goes over to humans for (hopefully) final review.
That said, I generally agree that you're correct: writing code in many ways has not been the biggest bottleneck. However, by removing much of that writing, it frees up engineers to work on the uniquely human things that are larger bottlenecks.
I had a few comments in a thread here touching on where I think most of the value has come from for us (which is largely search/understanding of our dependencies and making away team work far more viable, which aids with cutting through bureaucracy and the tendency for teams to push back on work): https://news.ycombinator.com/item?id=48298731
Haven't you heard - these days they just throw slop generated by LLM agents over to other LLM agents which cosplay as internal QA. They know it works because they write really strict .MD files where they instruct agents in English language to 'never do this' and 'always do that'.
This is really what you think happens at large tech companies? You don't think it's possible this is maybe even slightly overly simplifying what the relevant processes are?
Comment does indicate you don’t really seek to know how things work with respect to this and seem to not be able to imagine that the Occam’s razor is: agents are more useful than you think they are.
> I welcome you to also hold yourself accountable in the coming months if this trend continues. ;)
My company did not swallow hundreds of billions in shady investment deals and is not publicly traded. We work with real money, and the revenue on our books is the revenue that is actually booked, not fake revenue we plan in 2 years time to maybe happen. So no, I am not going to hold myself accountable. But people who work with other people's money should be absolutely held accountable when their wild imaginations don't come true, repeatedly, quarter after quarter, year after year!
Mate, for 5 years I've been hearing that crap. I am not predicting anything / on the contrary the AI boosting bunch is. When are your predictions coming true?
AFAIK, most predictions from several years ago were for...approximately now to within the next few years. Can you be more specific?
You criticized a very specific (and fake/misquoted) prediction, ignored the correction, and are now criticizing vague hand-wavey "predictions" that you have left unspecified.
Can you please stop with the angry/ranty replies and actually have a real conversation grounded in actual facts?
Now, having said all of the above...I'll also point out that these are predictions, not promises/guarantees. These people are being asked to forecast and are doing so. I hardly think they should be held responsible for not being literal oracles, but even so--please, at least quote them correctly/at all.
In short: be better than the hallucinations you're seen to call out from the models.
I will note that you have essentially not responded to anything specific in my comment, nor at least acknowledged that you misstated Dario Amodei's actual prediction.
So, unsourced vibes from a shady guy whose entire empire is built on being against AI?
I genuinely don't know how folks can continuously buy into anything he has to say after that Wired piece. The credibility there is seriously lacking.
Please, continue to be skeptical of the labs. But people need to stop talking about this dude as if he's the Holy Grail of the anti-AI movement. It's going to blow up in y'alls faces.
Ed actually provides sources and goes into an incredible amount of detail as to how he came to his conclusions. The average AI booster just goes "I totally built ten businesses off vibe coding but I can't tell you anything because it's a SECRET!". And the mainstream tech media is so in the pocket of big tech and AI corporations that they might as well just publish their PR emails at this point. Yeah, I'll listen to Ed thank you very much.
I think it's telling that most critics don't address his actual points, but instead his credibility because he's a "hater".
That said, I really mean it when I say that I don't actually think Ed is a good choice for the anti-AI movement. I think an actual opposition is useful, but he ain't it.
I read the profile and didn't see anything really wrong. Why would PR companies have to believe in their clients? Why does he have to be held to higher moral standards than Sam Altman who’s a total lying snake?
The error you call out is hardly “serious”, as the whole argument is uninteresting. It is a stupid indefensible error but the argument about revenue being 20% or 30% lower than reported isn’t that central to his overall thesis. Stuff that matters is inference cost, profitability, actual training costs.
> So, unsourced vibes from a shady guy whose entire empire is built on being against AI?
Actually he provides sources when he analyses stuff and imho much better than the usual corporate "Sam Altman says we should ask ChatGPT how to raise babies" crap. Also, I don't know many 'shady' guys who have built entire "empires", nor does he seem to actually have an empire. Usually being shady means you are kind of unknown and all. I am not glorifying Ed, don't even know him personally. I am not even impressed with his writing style much to be honest. But he brings important facts and information to light, which otherwise would have been lost in the cacophony of corporate media light treatment of these con-men. Holy Grail? Blowing up in our faces? WTF are you talking about?
The source was the article in the WSJ itself, which then referred to their source at the Anthropic. Which kind of is a textbook definition of "leak". Because otherwise Anthropic would have their lawyers hunting both the employee breaking their stringent NDA and the WSJ as well...
Why puzzled ? I literally said "According to Ed Zitron", implying that's where I stumbled upon the article. I've no time to read corporate media, at least not regularly.
> I currently subscribe to the $100/month Max plan from Anthropic and the $100/month Pro plan from OpenAI. If you are a heavy user of coding agents these plans are a fantastic deal.
Agreed. But its only a great deal because it is heavily subsidized, as you said yourself. Enjoy while it lasts, but in my book, product-market fit means something along the lines of "product which enjoys a loyal customer base, sold at a price perceived fair by the customers, and generating profit. How many of these does your definition of product-market fit hit here?
Fun fact: the $20/month subscription fee for ChatGPT Pro - which set the standard for at least a couple of years - really was an arbitrary decision made based on a Google form: https://simonwillison.net/2025/Aug/12/nick-turley/
They've got, ballpark, $5t to $10t to make back in the next 5 years, or the hardware buildouts will start getting written down.
This means we're going to need $1t+ per year in spending, per year, on tokens. 200m knowledge workers in the world, 30m developers. We're talking about a world where you need 5% of every knowledge workers salary to go into tokens. 20% if you're a developer.
That's a _huge_ shift. Most people I know cite +20%-40% velocity with these tools, against the actual work their company cares about doing. +20% speed for +20% spend isn't going to motivate a trillion dollars a year in spending.
We're not there yet. This is still the upswing of the hype cycle, and unless we figure out how to make developers 2x, 5x, 10x as productive on stuff that matters, this isn't going to play out well.
Here are a few thoughts:
- The publicly available information about how inference costs compare to training costs is conflicted. EEs involved in datacenters talk about power usage spikes during training runs as if they were a major factor in the designs, but academic papers discussing cost-optimal scaling confidently treat inference-time compute as a major factor.
- On the side of the balance indicating that training is more compute-intensive after amortization than inference is that Chinese providers, constrained primarily by access to compute, have nearly unlimited token availability at a lower price than US providers (inference), but poorer model capabilities (training). That would make sense only if US providers are inflating inference costs by 20-30x due to amortized training costs that overseas providers were not able to take on (there are other factors too).
- If training >> inference, they're in a prisoner's dilemma that far exceeds the ordinary zero-marginals model of competition between firms (due to its huge discrete stepwise nature). On the other hand, if inference>>training, the high-level analysis popularized by certain thought leaders, that it's like a utility, would be true. You'd tend to count this as a vote for inference>>training, but the CEOs saying it at least have a huge incentive to agree because the alternative, the prisoner's dilemma, would stop investment very fast.
- The only voice in the story that I just told you to have anything to do with fact (as opposed to high-level analysis and ivory tower armchair management of a secretive business) were the rumors from facilities engineers. That shows you the state of our understanding...
- If we don't even know the ratio between amortized capital expenses and operational costs, outside investor analysis is impossible. It doesn't matter how finely they divide the accounting buckets for office ferns and indoor ferns if the single biggest part of their business is obscured for trade secret reasons.
I'm about to leave a shallow comment, but I am a bit skeptical of the supposed drop in inference costs. If AI labs saw a lot of potential there, they'd surely be bragging about it non-stop? So the fact that publicly available information is conflicted is probably a sign that at the very least, the numbers aren't amazing.
Yes I know there's no evidence and this is lazy reasoning. But there's probably a bit of truth to this line of thought.
Why on earth would AI labs be bragging about how little the product they sell actually costs them to make? You don't want to do anything that reduces it's perceived value to the user, that might make them less willing to pay for it.
Also, inference costs are bound to go way down with more optimized architectures. GPUs are fundamentally not great at inference. No platform where the weights are streamed from a large pool of memory is. If the models ever quiet down, there will be massive step changes in cost/token, energy/token and tokens/second, as models are etched into silicon ala https://chatjimmy.ai/
A couple of years ago Altman was saying the price of AI compute is going to drop 90% year over year or something like that, so I don't think they're nervous about talking about lowering their costs. They probably just haven't been able to lower their costs.
You have to keep in mind that about 99% of their announcements are targeted towards investors (their most important revenue source..), so they're not going to be afraid to mention metrics that make the business look better.
Jevons paradox. Cheaper tokens does not mean we will spend less.
Cheaper tokens means the company's margins increase, which would be valuable for investors to hear
Ah yes, Sam “Not Consistently Candid” Altman
Oh, is that the guy that sold Loopt by claiming it had hundreds of thousands of users and it turned out to have 500 DAU after his exit?
Yep, the very same scammer. Wonder if he's lying about OpenAI too? Maybe about a person blowing a metal instrument?
he lied. he's good at that.
> Why on earth would AI labs be bragging about how little the product they sell actually costs them to make? You don't want to do anything that reduces it's perceived value to the user, that might make them less willing to pay for it.
Wouldn't they be bragging about it to investors? It feels like something that would matter a lot to them, and at least OpenAI kinda feels desperate to find them.
There's also the small question about whether a drop in inference cost would actually change anything about profitability, when training seems to get exponentially more expensive.
Why would any company brag about their margins ? Yet they do, to attract investors.
The key AI labs are not public companies, they are at liberty to brag about their margins to potential investors in private.
this is changing soon
Not really, how much of a public company are you when 5% of your capital is public ?
That doesn't matter for the legal requirements.
The short and only kind of wrong version is:
In the US, companies are not allowed to unfairly privilege some investors over others by giving them access to secret information that would let them judge the future prospects of the company. (Except in all the ways they can, but these usually involve some kinds of insider trading rules.) Private companies can handle giving out secrets to investors by literally writing and memo and mailing it to all their investors, if they want to give out some secrets to one of them.
Public companies cannot do that, even if they knew who all their investors were, but must instead consider every member of the public a potential investor, even if they don't already own the stock. Because of this, when public companies want to reveal material information about their future prospects, they must reveal it to everyone.
The percentage is irrelevant for this discussion. As soon as you’re public, you need to report detailed financial numbers.
Besides the legal requirement, the reason these companies go public is often to provide liquidity for early investors or employees. So they do want to have as good of a margin story that they can, at least in terms of unit margin.
Plus, you have to do real GAAP accounting, not their made up metrics.
That's changing with this administration though. Reduced reporting cycles reduce transparency.
It won't impact the disclosure of key business details because it doesn't reduce the level of disclosure needed in the S-1 or the 10-K.
This is an interesting anomaly in the US. In the civilised world all corporations have to file public accounts, as the price for their limited liability. The detail and audit requirements depend on the size, turnover, staff numbers etc. This is because the shareholders are not the only stakeholder. The companies creditors, for instance, who are exposed to the limited liability have a right to see what they are lending to.
To answer the sibling comment, all of these public accounts follow local GAAP or IFRS.
The US still astounds me with its willingness to allow corporations to rip people off!
Creditors in the US can make visibility into financials a requirement for financing if they want. Protecting creditors isn’t a good argument for public reporting.
What are the arguments against public reporting?
As a consumer you are often sending deposits or even the full cost of goods to companies some time before you receive those goods (in effect you become a creditor). You are also dependent upon some of those companies for service and repairs. It seems reasonable that you can check the finances of a company you are creating a business relationship with, I know in the past I've checked company statements.
You are unlikely to have significant enough sway to force that kind of disclosure. Small businesses as consumers have less legal protection and are similarly unlikely to be able to make disclosure a precondition of a deal.
Isn't there a limit on the public markets where if a company has less than a certain percentage of its ownership traded publicly then it is no longer a public company and therefore de-listed?
I remember hearing about a guy trying to squeeze out short sellers of his own company but ended up effectively taking his company private because he bought out like 95% of all the shares.
I wonder how that aligns to these small releases of stock for the public.
And investors will leak such claims quickly enough that this reasoning cannot plausibly hide big secrets.
It's not a big secret. If you just do the math yourself, it's easy to compute that inference doesn't cost all that much. People just see all the capital investment going around and all the new data centers being built, see that it's spent on "AI", put two and two together and get a three, or "clearly serving AI requests costs an arm and a leg".
The 1 they were missing is that AI requires both training and inference, and training is by far the expensive part. And that in principle you can stop training at any point and keep using the models as they are. (But that means that if other companies keep improving their models, you'll be left behind...)
In contrast, inference is fairly cheap and all the providers have great margins on it. Eventually either investment in training stops having commensurate impact on model quality, and people stop doing that and instead concentrate on making inference faster and even more efficient. Or if that doesn't happen, things will get very weird very quickly.
The market already shows where it will go.
If you want frontier model you will pay more for inference to essentially fund the expensive training.
If you don’t need frontier model you will get dirt cheap inference, which eventually will approach the cost of electricity spent per token.
This is technically correct, but practically false.
They can't stop training as then the AI's knowledge will become out-of-date very quickly. Their knowledge stops the day you stop training.
> If you just do the math yourself, it's easy to compute that inference doesn't cost all that much.
Show us your work, then. If it's so easy to do, this should be a trivial request to accommodate, no?
Just look at large open weights models being served by inference providers.
Kimi 2.6 is a 1 trillion total / 32B active parameter model that's something comparable to Sonnet. Sonnet's API pricing is $5 in, $15 out per million tokens. Deepinfra serves Kimi at $0.75 in, $3.50 out, and about the same at openrouter. So you're looking at a 4-7x multiple that Anthropic is charging compared to market rates that any plebe can get with a credit card.
I'm not sure just how good that looks for Anthropic/OpenAI.
4-7x isn't a tiny markup, but how does that compare to high-margin internet businesses like AdSense? Meta and Google do hundreds of billions in ad revenue a year, and after taking out the publisher's portion (60-80% per some searching), I wonder what the ratio of the remaining tens-of-billions is against the compute cost and headcount required to run it.
And how much room for maintaining or improving that margin do they have if the cheap competitors also continue getting better? Is there a "good enough" point where the easier inference tasks are all moving to vendors massively undercutting them, and then they don't have the volume necessary to justify spending on further cutting-edge development?
> Kimi 2.6 is a 1 trillion total / 32B active parameter model that's something comparable to Sonnet.
No it's not. On some rigged paper maybe. Some such benchmarks say all models group together, which they clearly do not.
> Sonnet's API pricing is $5 in, $15 out per million tokens. Deepinfra serves Kimi at $0.75 in, $3.50 out, and about the same at openrouter. So you're looking at a 4-7x multiple that Anthropic is charging compared to market rates that any plebe can get with a credit card.
That's not saying much. You can get "cloud" at AWS and you can get a VPS. There is likely a 10x difference. It's not "same". Whilst AWS costs more they also don't have 7x margins similarly.
Growing companies don't brag about their margins, they brag about their growth and revenue. Margin talk is for when you're a mature company squeezing out every bit of profitability you can - if anything it would be a negative sign to be worrying about your margins when you're supposed to still be growing and innovating.
I mean, did anyone expect them to not have margins? Why keep it secret?
Because companies that want to go public need to look profitable or potentially profitable. And before they go public they have to release real, actual, legally demonstrable numbers for their costs and revenue anyway.
Because the most important thing for any pure play AI company right now is to prove they are a viable company. And sure they have proved they can make billions, but also that they can lose billions more. They are going to need even more money and to prove to the next round of investors at an even higher valuation that they are a viable business they need to show not that they can generate revenue, but that they can one day turn a healthy profit. And that is the trillion dollar question.
I doubt having to replace every single chip in your data center every time you release a new model will bring down costs.
Went to that URL asked one question - "how is this different from other AI" and it took 598/6144 tokens, not sure what that means.
Not super clear from the site itself, but this LLM is running on specialized silicon implementing just it. So has super low energy use and blazing speed.
See https://taalas.com/products/
Edit: updated link
Incredible increase over Nvidia! Need to read more.. Thanks!
Because they can think more than one quarter into the future? Why on earth would someone adopt something into their core workflow that was fantastically unprofitable? Uncertainty and business don’t mix. Most people aren’t hype-eating bacteria that only care about maximizing their next paycheck.
One reason is that all the code you write with this goes in your private git. If using AI no longer is possible because of cost, you can still profit a lot from what you did with it before.
If inference costs drop 90% or whatever, that would be a massive write-off of hardware even before they gave any returns for it?! Given Chinese and others are snapping at the heels and would also benefit from such reduction in cost.
> Why on earth would AI labs be bragging about how little the product they sell actually costs them to make?
Investor confidence. They have a bit of a need for cash (also an interesting part of the profitability discussion of course).
> Also, inference costs are bound to go way down with more optimized architectures
I agree. Jimmy is incredible, I wonder what non-toy use cases they have. Surely they’ll come out with updated chips soon.
That said, I was apparently a bit over-excited for Groq and Cerebras. I thought they’d quickly dethrone Nvidia for inference, but not so far. Even the GPT spark trial isn’t seeming to go far.
Inference has traditionally been far less expensive than training. One public example is the fact that hobbyists can run StableDiffusion ($600k training costs[1]) on their personal computers.
Speaking to your point, inference being dramatically less costly than training would not be seen as a delta from the norm. The model of providing inference for anything near the operational costs (like a utility would), would the delta from the norm if it were true.
[1] https://x.com/emostaque/status/1563870674111832066
The difference between training and inference is 1) one have to keep intermediate results for backward pass in training and 2) computation for training double because of the backward pass.
Training is also done over batches, which increase memory requirements by several orders of magnitude. This is why training needs costly compute.
One of the ways out of this unfortunate situation is to use something like Stochastic Average Gradient Descent [1]. Examples there are mostly concerned with regularized logistic regression, which makes problem more or less convex. Neural networks are inherently non-convex. Still, maybe some ideas from there can be utilized in the context of neural networks, like use of estimated Lipshitz constant to derive curvature and appropriate learning step.
It's all got much more complex than that in recent years. Training now involves large amounts of inference for RL rollouts and similar. You can't disentangle them computationally like that. "Inference" is just the word used to mean serving customer traffic now, and "training" means creating the model you serve.
So one way to think about it is roughly,
Training is inference + backwards pass (~2x inference cost) + activations (vram overhead) + optimizer (vram overhead) + gradients (vram overhead).
Multiply "inference + backwards pass (~2x inference cost) + activations (vram overhead)" by batch size (thousands) to get to the actual RAM and compute cost. Optimizer like ADAM adds only two or three model-sized overhead.
And last, but not least, you need only one hidden layer kept in RAM for inference, but you need all of them (61 for Deepseek models) kept in RAM for computing gradient for one sample.
Microbatch size is a hyperparameter, it can be set to 1 and work just as effectively. With gradient accumulation it's equivalent even. Large batch sizes are used to increase parallelism, and sometimes to reduce variance in the loss signal (at the cost of increased bias).
Batch size is frequently limited by compute bottlenecks well before memory.
And of course you do all of this for every object in your training set, which is going to be larger than the total number of uses for any individual user.
Does it matter what is the difference in size of needed inputs for inference vs. training?
That is an estimate of the relative cost of one training step, but you have to multiply it by the number of training steps, an unknown quantity.
I think in your StableDiffusion example, a lot more than $600k will have been spend on electricity alone for inference (on those personal computers you mention). So inference is more expensive then training.
For equal capability tokens, there has been about a 10x drop in cost every 6 months.
We are still chasing the best because the best is moving rapidly, but it’s a simple thought experiment to work out what the cost to serve an 8B model from 2 years ago is in a world of 2T models.
Note: parameter counts are illustrative. Concretely, qwen3.6 27B delivers opus 4.5 capability at 1/27th the cost on openrouter. Single chip llama3 8b performance can exceed 17k tokens/sec.
> For equal capability tokens, there has been about a 10x drop in cost every 6 months
Is this still happening? Opus 4.5 was six months ago, can you get its capabilities for 1/10 cost now? Are we on track to get the same for 4.6 in a couple months?
Pretty much, Kimi K2.6 is opus 4.6 quality for coding. If you include discounts due to more efficient input caching it is around 1/10th of opus4.6.
https://openrouter.ai/moonshotai/kimi-k2.6
The march of cost efficiency moves on.
8B models would be consider obsolete in the world of 2T models, at least if we're talking about the competitiveness of OpenAI/Anthropic. The only reason why they are valued so highly is their supposed dominance at the top end.
The main story of agent use cases is in enterprise so far. An enterprise will only pay for a model capable of handling the task and no more. Most enterprise's see no need to hire PhDs as factory line workers.
Coding is an interesting case as [1] the pace of progress has been absurd and [2] it's hard to put an upper bound on required capability. However hard to put a bound on and will are different, it's quite possible that the average engineer will cease to see the benefit of rapid progress - or that their employer will be satisfied with lower tier models.
How smart of a model do you need to build a high quality CRUD app for internal users? Or build a scalable web service?
> I am a bit skeptical of the supposed drop in inference costs. If AI labs saw a lot of potential there, they'd surely be bragging about it non-stop?
Unless to the grandparent commenter’s point they’re using it to obscure their large prisoner’s dilemma (training) cost?
> If AI labs saw a lot of potential there, they'd surely be bragging about it non-stop?
Google seems to pretty regularly post about how their TPU and algorithm advancements have been decreasing energy costs for both inference and training.
Small alternative potential future changes that alter this analysis:
* At some point model capability reaches diminishing returns. Then inference >> training in the future but training >> inference now. It’s not a prisoner’s dilemma but a land grab to solidify market position and be one of the 2-3 firms left standing as dominant in the space. The model companies aren’t super sticky yet but they’re working on it.
* even if training remains >> inference, it’s possible to have multiple price points like they do today. If you need the most capable model you’ll be paying exponentially more per token to supplement the training cost even though the serving cost is marginal because most people will be satisfied with cheaper / less capable models for most tasks.
I buy that inference is a dropping line item while training is a growing one. There’s all sorts of things on the horizon that’ll be order of magnitudes improvements, from startups burning models into ASICs to get order of magnitudes more performance to alternate architectures like diffusion transformers that have orders of magnitude structural optimizations. It’s inevitable that it’ll come down even further from where we are. It’s possible model training also will go down but I’ve not seen any compelling research suggesting major “easy” reductions here.
The issue is that most tasks do not require frontier-level intelligence, but companies like OAI can really only profit off of the frontier. Capabilities from a year or two ago are so outdated that even OpenAI gives it away for free and there are many other models biting at their heels. In other words they are spending huge amounts of money to cash in on a depreciating asset.
So one possible future is that frontier-level training becomes so expensive and the use cases so sparse that it simply isn’t viable to keep going bigger.
We have GPU costs, power costs, and how many token/s models can generate on those GPUs. It’s possible to figure out the marginal cost based on this. The current estimate is about $0.40 per million tokens for gpt4 equivalent model. Sonnet 4 is $15 per million tokens, so they are charging high margins on inference. The issue is how large of a margin is needed to recover their costs before the GPUs age out, and how high of a margin can be charged before it’s not economically viable.
https://www.gpunex.com/blog/ai-inference-economics-2026/
That seems way off to me.
I skimmed the article, but couldn’t spot any details on their estimates. They mention 70b+ params as being large in several places. But we’ve had several 100b+ param models that trail Sonnet.
Why would power spikes from training runs imply training>>inference? The cost of a training run scales with energy, whereas power is energy per unit time. All that tells you is that they're speeding up their training run so it will take less time overall (probably chasing some first-mover advantage, where they're out with a given model before their closest competitors), whereas they obviously can't do that for inference (which is a steady flow of requests over time).
I don't see how it would be possible for inference costs to dominate training costs, even after amortization.
Training involves multiple passes over the entire training dataset, ideally in large batches where you can perform inference on as many samples as possible simultaneously and then perform backpropagation to adjust the model weights (which is about as expensive as inference).
Let's consider the size of the dataset we're dealing with here. The dataset likely consists of practically every piece of digitized text they can get their hands on (including that extracted from audio and video). We know Google has digitized a large portion of the books in existence as part of their "search book contents" feature and we have no reason to believe they're not using it alongside their cache of 90+% of the internet to train their models. We're talking about 100s of millions of books each with an average of 100,000s of tokens. The internet has 10s to 100s of billions of pages on it with who knows how many tokens on average. This is a huge dataset that we've got to go through hundreds of times.
Second, let's consider the effect of batching and how it sets requirements for our hardware. We know that larger batch sizes converge faster, are more stable, and produce better models. So if you want a good model you need large batch sizes. This means that you need machines several orders of magnitude more powerful than you use for inference. From what I heard Google uses clusters of 100s of the their TPUs all located in a single rack for training. These clusters are organized in a customized computing architecture to maximize memory locality between cores (really critical for efficient back-propagation). Further, you can't use reduced precision weights for training like you can for inference, so there are no shortcuts.
Finally, the initial training stage is followed by reinforcement learning stages - this is key development in how AI models have improved in the past year. This may mean going through a curated set of traces (either synthetic or captured from users) and adjusting the weights based on experienced outcome.
Overall there's so many orders of magnitude more work and more hardware requirements for training that I find it improbable that inference dominates. The number of "inference" steps in training is freaking ridiculous and includes such factors as the "number of words ever written".
It's been a while since I saw a detailed paper on a high end training run, but extrapolating from what I remember, it seems those training runs are in the 10s of trillions of tokens. This already accounts for potentially sampling tokens multiple times during the training run.
That seems like a large number, until you realize that OpenAI claims to have almost a billion weekly users. And OpenRouter shows many models at over a trillion tokens per week.
So in pure token terms, I'd say it is in fact extremely plausible that inference dominates, at least for the popular models.
Not saying you're wrong, but I'll note why inference might dominate despite everything you mentioned.
A given model is trained once but applied N times. A large enough N will dominate training, no matter how complex and costly it was.
But how long is a model useful for? How often will labs need to train new models? Time will tell.
This statement is well known to be incorrect for at least a year.
> If we don't even know the ratio between amortized capital expenses and operational costs, outside investor analysis is impossible.
And yet we surely need this data for the IPO? Or are they relying on rule changes on the indexes to force ETFs to buy shares?
Yes the huge discrete stepwise training spend is critical.
Maybe investors will realise that "the only winning move is not to play".
And so we are left with (as was) frontier models getting more and more out of date as whoever their post bankruptcy custodians are tries to eek pennies on the dollar for inference on their decaying property. Perhaps along with local and/or highly specialized models still feeding on the after-glow of the huge amount of training that was (and is no longer) done.
The next AI winter is going to be deep, savage, and long.
> frontier models getting more and more out of date
Why are they getting out of date? Is it because we have new content from the internet that the older models did not have? Or are we simply trying to increase the size of the training data? In other words not more up-todate in terms of time the content was created vs. wanting to use bigger training-input-sets?
lack of new content from the internet will make them go out of date. Not just facts and figures, but (for example) new programming languages/techniques.
I work for a tiny little company ($150MM annual rev with 9% net) and we are already looking at dropping $100k on hardware to run local models because, for us, they're "good enough."
Our estimated spend for AIaaS would exceed that cost in less than a year.
In a few years, there will be hardware capable of running frontier models good enough for most things at accessible prices for even tiny companies.
Yeah, that's the part that just seems to be wildly under-discussed to me.
If open source models are ~3-6 months behind SOTA, and ~opus4.6 capabilities are good-enough for product market fit, do the frontier labs have half a decade to catch up on their prior burn?
AI cost ballooning faster than companies can afford is becoming a very common topic in my circles right now. The era of "I'll pay infinitely more for marginal gains" is over from what I can tell.
> If open source models are ~3-6 months behind SOTA, and ~opus4.6 capabilities are good-enough for product market fit, do the frontier labs have half a decade to catch up on their prior burn?
They know they do not and that’s why they’re all trying to IPO right now, so they can pass the bag to consumer investors
More correlation, if more correlation was needed:
1- SpaceX + Tesla + xAI merger / IPO while Musk was vocal against IPO for about a decade
2- Warren Buffett cash at record highs
Someone got to be exit liquidity
Open source models that you can run locally are much more than 3 to 6 months behind. 6 months was the November inflection for Claude. No open source model is as good as Claude Opus 4.6.
It depends what you mean by locally. I don't foresee running a model on my laptop anytime soon to power a coding agent. Far more likely is an infra team at my company operating an open source model on cloud infrastructure. When they're already paying $1000 / month / dev, it starts to pencil pretty quickly.
Is there any open model as good as opus 4.6 at any price?
How many problems require Opus-4.6-level performance? The "I'll accept nothing but the very best model for any task" thinking is perplexing to me.
People got a lot done before Opus 4.6. In 6 months, would you be dissatisfied by Opus-4.6-level open-weight models, just because Opus 4.8 will be out?
Not OP but I've been thinking about this a lot (like everyone ha) and I think my answer is, yes?
I hope there's a "good enough" point but I don't think we're there yet. Like for me hardware got good enough several years ago. But while opus 4.7 is really good compared to everything else, it's not so good that I would use it at a discount over whatever is available in a few months. The improvement in quality, speed, and daily frustration is worth it to me... Spoken as someone whose employer is footing the bill, so take that with a grain of salt.
I want to run my own local models, but I don't think that's feasible without lots of frustration until a few generations of frontier models are so good that they're almost indistinguishable for common tasks. Kind of like how MacBook pros have been for a while.
While I can imagine that I'd want to use Opus 4.8 over 4.6 for a fair number of things (at least if they can avoid further speed regressions), I also have noticed that certain types of failures seem to be systemic. Bigger context has been helpful for bootstrapping, but still doesn't fix problems of getting stuck on the wrong things - you can toss more things in the blender, but you don't necessarily know which way it'll slice them up in advance, or which things from them it'll latch onto. And output still seems to get into "blindered" states where important details get dropped - even though it'll agree very quickly when you point that out. As long as we're in that sort of "spit something out in local targeted manner, and then do a revision loop until tests are green" style of execution, bigger models haven't shown me the ability to really avoid finding non-optimal / subtly-broken outputs for complex problems.
Using Cursor to hop between models, I've found Opus to be generally better at really tricky debugging than GPT 5.5 or earlier models, but not reliably better at execution because of these things. I'm not sure Composer 2.5 is quite there yet for the execution side, but it's getting pretty close to those other ones, such that I'm definitely still in a "debug and plan with slow, execute with faster ones" operating model for working on hard shit.
I'm very happy to have multiple sessions open (and do) and switch between fast and slow models, and if there were a batch mode in codex or Claude code I would use it. (Just like I sometimes use codex fast mode)
But at the moment, I can't imagine why I wouldn't be spending the majority of my time with the best models. I'm spending a lot of time with them! Reducing the number of back-and-forths is extremely valuable to me.
I expect in two months I will still want to spend >80% of my time prompting the best models, and that's true if I were spending my own money on hobby projects, too.
Something that's under appreciated right now is when designing systems and proposing solutions, my colleagues and I do a lot of brainstorming with llms. The core architectures have come out of that, but the best pieces of that architecture are still coming from humans.
These are ideas that simplify the design, reduce future work and tie together the entire system. If in two months I can arrive at ideas of that quality with normal brainstorming with llms that will be extremely valuable
1A. Models get stale pretty quickly w.r.t. new developments that occur past their cutoff date. "But you can just keep them current by linking them to never documentation, etc!" Well, no, you sorta can't -- at least not in perpetuity. Those search results fill up your context window real quick. So that gets unsustainable real quick.
1B. Even when your context has plenty of free space, the results you get from "here's a link to the documentation for this new framework that released after your cutoff date" absolutely pales to the results you get from knowledge that is fully baked into the trained model as opposed to your context window. For one thing, that documentation link you pasted into your context might link to... a dozen code examples. Whereas if that was baked into the model itself, the model might have been trained on many thousands of examples in Github etc.
2. It's also a reality that most professional engineers have to keep up with their peers and competitors. We can maybe say it shouldn't be that way, but it is. So if $SOME_NEW_MODEL is significantly better than 4.6... and my peers and or competitors are using it, then yeah I might but really feeling the need to match them. And I'm not even necessarily talking about some kind of cutthroat dog-eat-dog stack-ranked workplace.
These limitations aren't relevant for all use cases or careers but they're hiiiiiiiighly relevant for professional software engineering.
I image that'd be handled via a fairly regular minor bit of additional fine tuning to update them with new information rather than polluting the context space.
that's the nice thing about open weights, you can always retrain them with the latest documentation, no need to fill your context
Kimi 2.6 probably. Needs over 300GB of GPU memory to run (1TB for for full capabilities) so either a 4x A100 or 8x A6000 would do it.
A $50k - 100k rig could do it and an entire company would be able to use it a full speed.
No, but the big open models are on the level of Sonnet 4.6, which is very good for most problems.
The people who are claiming Opus level capability does not have sufficiently complex problems to see the difference.
And neither side brings any evidence ...
For coding don't think so, but they are very close. I code with sonnet mostly because I think opus is just useful if you fail to dissect problems adequately, but anyway.
Kimi is close for example regarding SWE bench for code. For reasoning there are open models that surpass opus by quite a margin already.
> that you can run locally
That's doing a lot of work here.
The future I see isn't most companies buying hundreds of thousands in hardware to run models, it's them adding a line item to their AWS bill. Inference costs on the larger hosted open source models are dramatically lower than the frontier labs API pricing.
The future I'm seeing is AI coprocessors running inference locally in most devices that today have a CPU. Just look at how powerful your mobile phone has become compared to your desktop computer 15 years ago and compared to a main frame 30 years ago.
The days of requiring a data center to run anything resembling opus 4.6 are already counted. (But the industry will fight hard to get people to keep paying the Claude tax.)
I'm already running a google TPU over USB on an otherwise very cheap board to do local computer vision on a front-door camera since I wanted to get away from Ring and other cloud services for that use case.
And yeah, that may be the ~decade world, but we're in the mainframe era of the frontier models. It's going to be more economical for basically any consumer, and most businesses, to pay someone else to host models for quite a while.
Curious why you went for a custom solution. I am aware of at least one company that seems to ship devices with local computer vision (Reolink).
My experience over the past decade has been being subsequently burned by being reliant on one provider's ecosystem after another. This is great until Reolink starts doing something shady to pad the bottom line and then it's on to the next.
I wanted the ability to run whatever cameras on a VLAN and own the stack.
I'm guessing that they are using Fargate which is an OSS NVR. It supports a little addon USB stick you can buy for about $30 that will run common computer vision tasks for object detection. Stuff that we've been able to do with WebAssembly and Canvas for a long time now.
A gaming PC can already host models that perfectly serve casual users who just want recipes, todo tracking, picture identification, etc. E.g. Qwen 3.6 35b which will run on a $650 GPU at 75 t/s (Nvidia 1660 ti 16GB).
Said model will also run as a tool-calling coding model excellently (it's no Opus, but for a thing that once set up is just the cost of energy, it's incredible). It can type faster than you can, probably 10x faster, so with guidance it'll make you faster. And it's free.
It's here. If folks want ChatGPT without a subscription, they can have it today on their computer. The only money to be made is in the high end models doing "serious business" work spanning 1M+ token contexts and massive uncertainty. Everything else is already set to be eaten by today's local models.
The problem with models like Qwen 3.6 35B (which really is an excellent model) is that my expectations of what a model can do have gone SO high now.
Here's a prompt I just ran against Claude Opus 4.7:
> Use python3 to experiment with whether the SQLite3 authorizer mechanism can be used to detect an INSERT OR REPLACE based just on running an explain query without examining the SQL string itself
Opus nailed it: https://claude.ai/share/c4212606-3fee-4b7c-bc97-505e0348ccac
I tried the same thing against qwen/qwen3.5-35b-a3b running locally in lmstudio, with the Pi coding agent. At first it looked like it was going to do great! And then it fell apart over the course of several tool calls: https://gisthost.github.io/?8ae2f842df619fb7fd8f1ccd82fe41c7
I'm used to GPT-5.5 and Opus 4.7 handling that kind of prompt without any problems at all.
This worked for me with qwen3.6-36b-a3b even at a q4 quant. I ran pi in a docker container and it had to figure out how to install python as well. I used the same initial prompt you had without any additional. You talked about Qwen 3.6, but then said you tried Qwen 3.5 in lmstudio. Not sure if you meant Qwen 3.6. I ran with llama.cpp llama-server with the recommended settings from unsloth.
I'm not an expert in SQLLite so I can't say if this is 100% correct, but it seemed directionally similar to the conclusion from claude.
I can't help but think the not-so-distant future will see language models expected on commodity personal computing devices.So one of the prominent LLM advocates known for testing every model shared a prompt intended to exhibit Opus 4.7 capabilities, and Qwen 3.6 sorted it out okay? Interesting.
Not saying they're equivalent, local models still decohere much quicker as the context grows in my experience. But... Interesting.
OK that's a very good answer! Do you mind sharing the transcript?
Something is definitely going wrong with your Qwen setup, in the link you posted it starts and ends with a compaction step due to a 4k token context limit. Qwen 35b supports I think up to 200k+ context limit (though I run only with 128k), that seems to be a major source of the problem.
Good call, I need to check if LM Studio is misconfigured.
Thats when your build a better Ralph loop around your llm for it to converge to an answer and not rely on 1 shots
> a thing that once set up is just the cost of energy
I don't think we can discount this, frankly. Newer electronics are energy efficient, but older devices are more energy-intensive, and unless configured well, a gaming PC can easily use a few dollars a month in electricity, so now you're approaching subscription territory. A subscription comes with no upfront cost, higher reliability, no wasted space in your home, mobile apps, etc. (and less privacy).
Even when run on datacenters, it would be like current day webhosting. It is hyper competitive and it will be a race to the bottom. There is money to be made but not as much as investors hope. There will be datacenters in random countries like Kazakhstan because some oligarchs have found a free energy glitch (like with bitcoin mining).
> But the industry will fight hard to get people to keep paying the Claude tax.
I bet this will ironically be couched in "safety" reasons or regulation to get anti-AI folks on board, even if it favors the large incumbents.
Counted but not yet numbered?
Magical thinking. I guess if your phone is going to have 128gb of dddr5 then sure. You people fundamentally don't understand the memory requirements for running inference. Your cute local models seem good enough because you have no standards and anything an LLM produces seems like magic to you.
> Magical thinking. I guess if your phone is going to have 128gb of dddr5 then sure.
Why would it not? The typical new phone today has 16gb of RAM. 20 years ago that was somewhere around 32mb. Factor 512. It's not hard to see that we'll get there rather soon, especially if there is an application that provides demand.
> You people fundamentally don't understand the memory requirements for running inference.
You seem to be overlooking how fast things change in this industry, especially if tons o money can be made as a consequence.
> Your cute local models seem good enough because you have no standards and anything an LLM produces seems like magic to you.
Please don't generalize. I'm an expressed AI skeptic and have to deal with the bad consequences of AI slop every day. But you can't deny that there are enough applicationn areas where people have use cases and those will be much easier if things don't need a few round trips to a data center that sucks all the electricity and water out of neighboring communities.
Eh, you're off by an order of magnitude or so on both ends.
The iPhone 17 has like 8 gb, the Pixel 10 12.
The original iPhone was 128mb, and the iPhone 6 from 2016-2018 was around 1gb; that puts the iPhone at around 8x RAM per decade, and puts us at 128gb in our pockets at around 2036 or so.
(Incidentally, the big news in phone RAM is that a lot of new phones are dropping back to 4gb because of RAM shortages.)
> it's them adding a line item to their AWS bill
That's the future Amazon sees too. We just had a week long session with the AWS team and they pushed that to us multiple times.
Buying "hundreds of thousands in hardware" sounds like a lot but many companies - especially software companies - already do that if they have 100+ employees.
Running software in the cloud gives you certain reliability and scaling advantages that would be very hard to replicate locally. Running some code agents in the cloud vs local hardware, if the local hardware gets "good enough," breaks the other way - offline usage, alone, would be hugely valuable to many people and companies.
It'd be very interesting to see where various players would decide to make a call "local is good enough" though. Buying the hardware isn't a small bet, if it's not something that ends up as part of your standard computer.
Many business tasks do not need the latest frontier models. I have a production system running since early GPT-4o. It now runs with GPT-5.2, not for improvements, but because it is cheaper. I could invest in switching to a local model, I tried and it works well enough, but api costs for this task are so low, it barely scratches $30/month. So I am using the local machine for other things and leave the inference on OpenAI, for now.
I've been doing my work with OpenCode Go, with Kimi2.6. It is not as good as Claude Opus, but it's good enough to get the job done, and I never run out of tokens.
This project argues that with appropriate harness, the performance gap between frontier and much smaller open weight models shrinks dramatically: https://github.com/antoinezambelli/forge. I haven't kicked the tires yet.
Opus 4.6 is a February model. Every time this subject comes up it seems like people post intentionally misleading things and move the goalposts.
The goalpost we've been bludgeoned with over and over again is that, in particular, Everything Changed in November 2025. That GPT 5.2 and Claude 4.5 were the inflection point. That is actually 6 months ago. And DeepSeek 4 is already there.
> run locally
You can't run DeepSeek locally on consumer hardware[1], but you can on enterprise hardware, and enterprise spend is the subject of this conversation -- and even if you aren't self-hosting, it doesn't matter, because you can just get your inference from one of the the many companies serving DeepSeek, who trivially undercut the pricing of OpenAI/Anthropic because they didn't have to spend hundreds of billions on training frontier from scratch but instead only invest in supporting inference, which is already profitable.
[1] Since this misconception comes up all the time, I'll go ahead and pre-empt it: no, training a 32b parameter model on outputs from DeepSeek and running that locally is not "running DeepSeek", despite the hundreds of stupid articles and Youtube videos making that idiotic claim that they're running it on a 5090.
> You can't run DeepSeek locally on consumer hardware
I'd qualify that by writing that you can't run it with ordinary, real-time speed and throughput. If all you care about is slow and high-latency inference, there's no reason why that shouldn't be feasible even on the cheapest miniPC around, as long as it can literally store the model weights and keep around the (rather small) context.
> You can't run DeepSeek locally on consumer hardware
Maybe not DeepSeek v4 Pro, but I've run DeepSeek v4 Flash on my 128GB MacBook Pro using antirez's carefully quantized https://github.com/antirez/ds4 and it's impressive.
Oh sure, yeah, that's nothing to sneeze at either. I think unqualified "DeepSeek" should generally refer to the main model, though, especially in the context of GPT5.2-grade quality.
I keep hearing about this "inflection", but it feels extremely exaggerated to me. And yes, I was using it at the time. It got incrementally better, it wasn't that amazing.
I think the bigger shift was harnesses and the two ended up somewhat commingled in people's minds.
Claude code was a lot of people's introduction to using coding agents that could do a lot more than copy-pasting from a chatbot or autocomplete.
The tool usage + skills got markedly better and so did the thinking cohesion. Add 1m context windows and it was a very noticeable shift.
Opus 4.6 quality for local inference would be revolutionary.
1m context is garbage
But one will be in few months. And then you have choice of paying say $100k for hardware and pay just power cost (or pay someone to do that for you), or pay way, way more for your team to have access to marginal improvement.
And 5% worse model for 10% of the price of the bleeding edge will be worth it for majority of people
To be relevant to this discussion, models running on reasonably-priced local hardware do not have to be as good as the best.
They just have to be useful enough that companies don't need the best.
They are.
Kimi is better.
There's still a lot of room for the best models to get better at coding .
Your argument rests on the "for marginal gains" part but it's really not clear that the gains are marginal in the foreseeable future.
This is totally valid and I don't agree with the downvotes you're getting. Someone coming out with a 10x improvement is possible and would change the game immediately. The thing is, we really have been seeing marginal gains with shifting leaders in who's got the "best" since GPT3, and at least as a user of these tools that pace has been slowing, not accelerating. Subjectively it feels like we're in the back half of an S-curve.
We're 3.5 years into this current AI wave, and a lot of the valuations have been predicated on what you're arguing here -- that essentially should one of the labs make an order-of-magnitude improvement or hit escape velocity on recursive self-improvement they'd become the most powerful economic chokepoint in history.
The reality has been that given access to compute + capital all of the labs can stay pretty competitive with each other. Someone does a bit better on coding, someone else does a bit better on tool calling, and then they swap after each spending another $100bn.
The market looks like a commodity market where the commodity is intelligence, not a winner-take-all market with massive margins. Plenty of people get rich in oil and airlines, but they notably don't tend to be the innovators long term, they tend to be the operators. Obviously if the machines become sentient tomorrow, turn on their masters, and hit world-dominating intelligence, that assessment changes, but after several years of that narrative while objective reality looks quite different I think the more sober voices are starting to gain a foothold.
I agree with most of what you're saying, but I think the point I was trying to make wasn't as high-flying as you and others understood it.
I'd pay a premium for even just a model that's 20% better, no ASI required, and I think a lot of people would. I wouldn't call that marginal, if it means I'm getting frustrated on 20% fewer tasks.
A recurring pattern that I've seen in myself and others is to at first be very impressed by a new model's coding capabilities, and then desensitize quickly and start being frustrated by the shortcomings.
> I'd pay a premium for even just a model that's 20% better
The point I'm making is that I think we're rapidly hitting levels where corporate buyers aren't willing to pay multiple-times-more for marginal gains, and I expect that to become more the case over time, not less. You, and a small % of other power users in the market might tolerate a $400/month pro-supreme-plan for access to Mythos or whatever, but I don't think that's going to scale up in quite the same ways we've seen so far.
Even a year ago paying multiples times more for a 50% gain was very sensible for a lot of workflows. But if we're getting to "good enough" for things like coding, justifying to your CTO/CFO why the org should go from spending $1m/year to $5m/year for a 10% higher hit-rate on one-shot prompts from the engineers is a much tougher sell.
What? The gains between gpt4->5 seems to be marginal. No phd level discoveries here
The leap from GPT-4 to GPT-5.5 has been astounding in my opinion. There is no way GPT-4 could run a coding agent harness like Codex at even a fraction of the quality that GPT-5.5 does.
I don’t think that’s exactly indicative of GPT-5.5 being an astoundingly more intelligent model, however. An alternate interpretation is that GPT-5.5 was trained on tool usage/harness patterns and has been optimized for this use case.
I remember that even when GPT-4 was king, the Gorilla paper showed that Llama 7B could be fine-tuned to outperform GPT-4 on tool calling.
On domains that don’t involve agentic tool calling*, I haven’t found the frontier to have advanced that much.
Edit: I should broaden this to domains that naturally lend themselves to RLVR training. Models are drastically better at math now.
None of this matters in the product: it either is capable of agentic loop workflows or it isn’t. A 10% improvement in probability of single task success makes or breaks the use case.
Open source models, especially qwen are pretty dang good. But its not opus 4.6, the evals dont tell the full story. I question the assumption open source models are 3-6 months out.
Its not just about the quality of output, but you also can finetune them to proprietary needs, if the skillsets are their internally, to make them better without governance risks. So being SOTA doesn't matter as much, since generalized tasks are not what matter most to companies, its the specialization relative to business need or internal datasets.
To make an extreme comparison, desktop Linux was originally supposed to happen in 1999.
Maybe I misspoke by saying open source.
The larger point I'm making is I think models are rapidly becoming commoditized. There is probably a small market long term that's willing to pay 10x for 10% marginal gains, but the majority of the buyers in the market will be economic and we're likely to have a lot of folks willing to spend 1/10 the cost for 90% of the performance, and plenty of companies that haven't raised hundreds of billions-trillions who can provide that.
A lot of the frontier labs valuations has been based on an assumption that 1-2 companies would get break-away intelligence that basically made them economic chokepoints indefinitely into the future. The reality that's becoming increasingly clear is that model quality is a pretty linear function of (cash burned - ability to copy other's homework) and the economics are starting to look a lot more like airlines than online advertising.
Lets go one step further.
The economics of airlines are such that they generally earn a return on capital less than cost of capital.
I think this is exactly where we are heading and OAI-Anthropic are the concordes.
You have to think about why open models are behind. Exfiltration is a big part of it. So you could change the Nash equilibrium by increasing your security, or other multilateral approaches.
If only the AI era was born in ZIRP.
Better now than ZIRP for me - at least people are asking timid questions about the unit economics and how long the runway is _early_ while also spending absolutely insane amounts of money on this bet. During ZIRP, these companies would have turned down any investor asking questions. Less contagion when rates aren't zero hopefully? :grimace:
The size of the AI bubble and the IOUs being passed around like a hot potato already dwarfs the real estate bubble preceding the 2007 crash.
If we still were in the ZIRP era, busting the bubble would certainly kill off the world's economy for good simply due to its size.
> ...we are already looking at dropping $100k on hardware to run local models...
Just think how much further that $100K would have gone if the hardware market wasn't so screwed-up.
Anecdote: I priced-out adding 1TB of RAM to a four node cluster a couple months ago. The cluster was purchased in fall of 2024 w/ 4 nodes, each with 256GB RAM. The nodes cost just over $14K apiece back in 2024 (entire box, not just the RAM).
Dell wanted >$90K a couple months ago to add 256GB to each node.
> Dell wanted >$90K a couple months ago to add 256GB to each node.
RAM is expensive, but not THAT expensive. I just bought 128Gb for about $5k for our build cluster (it's not even for AI, sigh). Even if you need larger-sized DIMM sticks, it's still going to be in the vicinity of ~15k tops.
It was crazy. I found the part on the open market for a lot less but the edict from the Customer was to buy from Dell to keep the support entitlement intact. That inflated the price to an astronomical level to be sure.
I haven't had problems w/ Dell support and 3rd party memory, personally, but given the machines' application I understood the concern.
Yes 100% this. A lot of people keep talking about how OpenAI and Anthropic will need to raise their prices. What is less discussed is how they CAN'T raise their prices because competition exists, and sure it's not SOTA, but it's literally an order of magnitude cheaper in many cases and the drive to figure out how to make it work well enough is going on right now (and will only intensify when the SOTA models raise their price).
It's a given that the SOTA models need to raise their prices. It's also a given that they can't. The more they raise the more customers will move to their competition.
So what happens next? Well I think it will suck horribly if you can't move off of SOTA sooner or later, because the Big Two are going to lose customers, and therefore have to raise prices on the locked in customers even more than these projections suggest.
Beyond that if you're looking to start a business, figure out how to use cheap models in new scenarios. Build software which does that and license it. This is kind of contrary to the idea that you shouldn't over optimize for deficiencies in the models that will likely go away in the next generation - for instance a lot of problems were solved when context windows got way bigger. So it's a thin line to walk but I think it's there because a lot of orgs are using Claude today for pretty basic tasks.
The dev who's addicted to SOTA models honestly is going to have to settle for less or get totally screwed. Most applications within business from what I see aside from complex research do not require SOTA. They summarize, they classify, they transform, and doing that accurately has been cheap for a while.
I get the impression the hive mind hasn't come to terms with the point that a model is optimised for certain tasks. It's like having someone ask you "is that a good hammer?". Good for what? There are claw hammers, sledgehammers, ball-peen hammers, club hammers, mallets, .... Yes, in a pinch, they can all bang in nails, but you wouldn't choose a dead blow hammer for that if you had a choice.
The Gemini Flash is very good at searches. Just about any low end model can toss out a poem. All the higher end models (open source and otherwise) seem to be able to churn out code that passes tests. The smaller, "less capable" ones are much faster at it, which means in the hands of a skilled practitioner are the best choice for that task. But they rapidly fall apart where there isn't a hard source of truth (like a good test suite) to grind against. Because of that you have to use a bigger model for bug finding. In that task the open source models tend to fail on larger code bases, where something like Opus still shines. I gather Mythos is an absolute monster, and unparalleled, and unavailable. I'm sure one of the reasons for that is it's so expensive to run.
Or to put it another way - you don't use a 100 tonne crane to pick up the shopping. And ... the smaller models will happily run on in-house hardware. You may not do it today because of the current DRAM price and integrated NPUs have just started shipping, but in 5 years time models will be running on your phone.
On prem AI makes sense for more than just the cost. More control, IP, model improvements you can keep, data privacy to name a few. People will realize that AI is not like compute the moment they get their own knowledge sold back at a premium.
> People will realize that AI is not like compute the moment they get their own knowledge sold back at a premium.
But what if your competitors sell their knowledge to AI companies?
Then you're still screwed.
What are the advantages to on-prem for a company that's already in the cloud and trusts it with their IP? That company can just rent GPU instances from the cloud if they want to train/fine-tune their own models and keep avoiding CapEx.
Agree. You have these tipping points when a model is good enough to do some task. Yes, a better model will further improve your capabilities but the unlock is at a certain intelligence level. We see this also with humans. People with very low intelligence can't learn to read. Once you cross a certain threshold of intelligence you can learn to read. More intelligence doesn't really help you in the task of reading. A person with an IQ of 160 is not substantially better in reading than someone with an IQ of 85. If your IQ is 50, you might not be able to learn to read at all.
Have you considered that a smarter person will understand what they have read better?
Depends on the task and the writing though doesn't it?
There's not that much depth in a lot of 'everyday' writing. For many tasks that means that you don't need to be hyperintelligent - reading a recipe or a shopping list, reading a newspaper article, etc.
I don't quite understand, what would 100K buy you?
AFAIK you would get about ~5 concurrent users, with a max context window of ~128K tokens on the larger models.
This wouldn't be good enough for coding -- are you guys thinking of using it for something else?
Gigabyte 4x AMD Instinct MI300A rack server (512GB GPU RAM total)
Roughly equivalent to 4x H200's for less than half the price.
Vaguely around 60k tokens per second...
By my calculations 100k could get you 18 5090's + compute to host them, or 18 96gb Mac mini's. You can get a lot of context window and users out of that setup.
Do you think this will be a trend for larger companies as well?
The decadal move to all-cloud-all-the-time killed off in-house hardware teams while the C-suite chased their OpEx dreams.
It would be interesting if we come full circle on this.
I doubt it. Companies that have moved to the cloud are already trusting the cloud with their IP. You can rent time on a high end Nvidia system from various clouds. OpEx means there's no write down in three/five years as that system goes out of date so it would only make sense if the performance/$ is there, or the company is highly protective of their IP and doesn't trust the cloud, at which point they're not on the cloud anyway.
I configured a dual DGX Spark cluster, and it's certainly "good enough" for my agentic and coding needs.
what models are you using on that? My experiences with apple hardware have convinced me that it is not really good enough for coding locally.
DeepSeek v4 Flash, various quantised versions of Kimi K2.6, MiniMax 2.7, Qwen 3.5 “full sized, with a dual spark setup you can fit some decent setups on here
My single spark has me running Qwen 3.6 27B and antirez’s specially quantised DeepSeek v4 Flash (which is shockingly impressive)
Kimi K2.6 does not run well on 256GB.
Have you tried it? It would be slow for sure, but the main limitation AIUI would actually be storing the context in RAM - models like Kimi and GLM have high demands there which limit your ability to get meaningful aggregate throughput via large batches.
True, I might be thinking of some of the communities four-Spark clusters for it; it’s already int4 right?
It isn’t the models, it’s the closed api and the tooling associated with it. It’s driving me crazy how not-talked-about this is.
You can point both Codex and Claude Code at a local model and they'll work just fine. Codex even explicitly supports that as a feature! [1]
With a nice UI on top, for the desktop app too: [2]
[1]: https://developers.openai.com/codex/config-advanced#custom-m...
[2]: https://docs.ollama.com/integrations/codex-app
As in the coding harnesses?
If I could leverage the same closed api VSCode uses, the entire moat is drained.
What you call harnesses I call… bullshit?
My much larger company has got people already using various models through Bedrock because the Claude and OpenAI limits are too harsh and it's too expensive.
It might be possible that in a few years someone will be able to engineer a reasonably priced machine to run today's frontier models (hint, your price is an order of magnitude off). However, they won't be able to run the frontier models that will exist in a few years.
I’m curious: are you spending on beefy developer machines, or some kind of shared local inference server? Would be interested to know more if it’s the latter.
I am aware of at least a handful of companies doing the latter. I don’t work for them and cannot speak to their setup.
> In a few years, there will be hardware capable of running frontier models good enough for most things at accessible prices for even tiny companies.
What makes you so confident about this prediction? Hardware costs haven't exactly been cratering recently.
.> Hardware costs haven't exactly been cratering recently.
No, but local models have been booming in performance/quality improvements. The RAM shortage won't last forever (more supply will come online when if demand doesn't diminish), and then the math would be pretty easy.
What about using DeepSeek API? Practically free.
same, but you need more then 100k of hw to run something like kimi k2.6 for a bigger team. on the other hand there is a ds4 flash that you can run on a macbook with 128gb ram. an that one is perfectly usable for a lot of tasks.
https://github.com/antirez/ds4
What models? Last I tried different local modals there was a pretty big difference from frontier.
Eh, one question. Where do you intend to buy the hardware if datacenters take over the market?
That’s exactly where the market is heading and it’s going to have to reckon with this fact
My guess is there’s gonna be some legislation or something “you can’t share anything over this level of complexity” and I think that that’s what a lot of that mythos rattling was all about
> there will be hardware capable of running frontier models
The current frontier? Sure. The frontier then? No - obviously that frontier is going to keep consuming available datacenter compute capacity, which will be better
You people are delusional. How many times a day am I going to read this fiction of "good enough in a few years for most things".
There are physical limits to how much you can compress data and how much is needed for a capable model. If by hardware capable for running SOTA you mean a 7 figure investment for a company, than sure. But how come these companies didnt do the same thing for cloud? There's been this option for self hosting infrastructure for a decade but companies don't use it, they pay AWS.
> In a few years, there will be hardware capable of running frontier models good enough for most things at accessible prices for even tiny companies.
I was going to say - the models are just going to keep growing at a pace exceeding the pace of hardware pricing/availability
But then I realised that, far more likely, there will be a plateau reached (again) where nobody is seeing gain, and at that point hardware will catch up
I was in college in the late 1990s/early 2000s and I distinctly remember an econometrics professor state the following:
"As cable TV and Pay Per View came out, there were studies done about how many movies people would watch if given unlimited access to films. The results were bandied about as proof that we should build out all this infrastructure to support this line of business. When the data was further analyzed by statisticians etc, it turned out that people claimed they were going to watch films 10-12 hours a day, every day of the week. Impossible."
I feel like we are in a similar boat here where some people are assuming:
- EVERYONE is going to be using max tokens
- tokens will NEVER get cheaper due to improvements in hardware, software, design, market forces etc etc
>I feel like we are in a similar boat here where some people are assuming: >- EVERYONE is going to be using max tokens >- tokens will NEVER get cheaper due to improvements in hardware, software, design, market forces etc etc
I feel like the reverse assumption is being made, that the current model looks like IBM doubling down on Mainframes soon to become cheap enough to deploy everywhere, when the real action is that the costs coming down represents cheaper hardware or more efficient software, and that a big chunk of "cheaper" AI will be eaten by smaller products deployed by individuals. Whatever the Personal Computer of AI looks like is going to be more disruptive than just an API endpoint you can fling tokens at.
We already see this with things like chrome auto installing an LLM.
You cant tell me with complete certainty that theres a moat here for the people spending 1 trillion + on this infra.
>When the data was further analyzed by statisticians etc, it turned out that people claimed they were going to watch films 10-12 hours a day, every day of the week. Impossible.
I also think this applies to people suggesting that companies will sack workers for AI, when the costs of replacing everything someone does in a day is more expensive in terms of tokens (likely even at a reduced price) than just hiring a bloke.
> it turned out that people claimed they were going to watch films 10-12 hours a day, every day of the week. Impossible."
I realized it long ago: one needs output to make meaning. Input can only be the cherry on a cake in one's life. That, actually, makes FIRE or Fat FIRE not so sustainable unless one has other hobbies.
> it turned out that people claimed they were going to watch films 10-12 hours a day, every day of the week. Impossible.
And what happened? How many hours per day/week are people spending watching now?
What people: I'm sure some people are watching 10-12 hours per day - in places like nursing homes or hospitals. I know a reasonable number of people who watch a film nearly every day: 2-3 hours. Most people watch something every few days - often a tuesday night movie night for the family (or something like that). There are some who never watch anything. I don't know what the statistics are on this.
My friends in day care tell me the kids hate "movie day" because movies are all the get at home and they are sick of them - they want to play all day. (but I'm not sure if this is representative of anything other than the types of people who put their kids in that particular daycare)
> they were going to watch films 10-12 hours a day, every day of the week. Impossible.
A lot of these LLM demand scaling scenarios make broad "up and to the right" assumptions about things which in practice have finite limits. Only some percentage of knowledge work benefits from acceleration, optimization or other improvements, and even then the amount of economic gain is capped.
But isn't it wonderful that they did?
It's vaguely disturbing that people "watch" films 10-12 hours a day. Many of them are using it as a radio, for background noise, without really caring what the program is beyond vague genre, tuning in and out without particular regard to the plot… and yet we have all the cost of transmitting high-resolution video point-to-point.
Surely we could just put better stuff on the radio, and accomplish most of the same goals for a far lower price?
My Dad was in the hospital, and just wanted to watch the Pirates play. The TV was filled with apps, some of them free to watch, others demanding a subscription and log in once you selected something.
None of them had the Pirates game.
I was thinking how the transistor radio was a far superior experience for this use case. Just tune to the channel broadcasting the game.
You mean the station that the MLB regulatory captured into not broadcasting when the local team was playing?
Radio has not gone anywhere you know? There is of course podcasts, but for instance Radio France has amazing music services like FIP: https://www.radiofrance.fr/fip
Then there’s NTS, BBC… Ypu can listen to them from online service, but at least in Europe there’s amazing national FM broadcastimg services.
TV is just bad radio with flickerimg lights.
Who has the time to watch films 10-12 hours a day?
I think the comment put forward that as an incorrect assumption that was made prior to the cable build-out.
Which is now an actual way that people use streaming services.
> - EVERYONE is going to be using max tokens
anthropic already hunts down OpenClaw users for using too much on their plan.
I'll give different example: When LED lights started to be more popular, the power usage didn't drop by the amount of power saved
>- tokens will NEVER get cheaper due to improvements in hardware, software, design, market forces etc etc
Well, first, improvements in computing stalled or even rolled back just purely because price of everything compute shot up cos of AI and that will NOT be fixed for a while and ESPECIALLY if AI usage will continue to increase
Second, the token per model might go down in time but better models have more expensive tokens, so we quickly get into spot when:
* price increase in token might not be worth marginal improvement next, better model brings
* more and more models are passing "good enough for the task" threshold so for less and less companies there is any economic sense to pay for the "best" instead of paying deepseek or some other company to run "previous gen" models
I don't think the maths works like that. They have raised ~$200bn so far and need to make that back. Saying they need to make $5 to $10tn isn't really real. They might need that to meet some extravagant Altman projections but not to justify what they have actually spent.
Yeah, claiming “product-market fit” on coding assistants for this multi-trillion dollar capital expenditure seems premature. Anthropic will post one and only one quarter of “operating profit” (aka losses after taxes and debt obligations) on the back of free-for-all spending by enterprise and engineer tokenmaxxing, neither of which will last. The investment was commensurate to a world-eating AGI, and if all that comes out of it is coding agents and slightly better enterprise software, I don’t think that makes up for the money spent.
The real question is: can you incentivize a non-tokenmaxxing Uber to spend the same amount on AI as they were when tokenmaxxing, just with fewer tokens and higher per-token costs? Even with plateauing improvement in frontier models? I think the answer may be yes.
And part of my reasoning for this is: the only system capable of actually fixing bugs in vibe-created code is an LLM. If we humans couldn't write it without assistance, we certainly won't be able to debug it without assistance. So there's a real stickiness here.
We're signing pacts with demons - we have to, if we want to outcompete the other warlocks - and those pacts are written in the very size of our codebases.
The bottleneck has moved from producing a thing that works to knowing that the thing was the right thing to build. The more of the latter they can take on, the fewer knowledge workers are needed at all. So rather than 5% of every knowledge worker's salary going into tokens, 100% of the knowledge worker's total employment cost goes into tokens and you get a 20x productivity boost as a theoretical minimum across those tasks.
That's the game. There's a view you could take of this that this is just a growing of the pie: with those cost dynamics a lot more "small businesses" get a vast amount of leverage, so the overall economy grows without replacing the knowledge workers. I'm not sure I trust the MBA class to have that view.
>The bottleneck has moved from producing a thing that works to knowing that the thing was the right thing to build
I would argue that that's been the case for quite some time before AI. As an example, what innovative amazing world-changing products have Google or Meta launched in the past decade with their very high numbers of very talented and highly-compensated engineers? The issue with most big tech companies are leadership, strategy, and product direction. I'm not saying that they don't make any profits, just that they probably aren't "building [the right thing]".
AI for product development and management would be far more impactful than automating rote coding tasks / building React UIs that mirror API structures IMO.
> AI for product development and management would be far more impactful than automating rote coding tasks [...]
Yeah, if this stuff actually worked that well already, OpenAI et al. would just run AI CEOs and engineers. Why get some other company to pay you at all when you can automate every other company out of existence and take all the money they make?
The fact of the matter is that while the tech has some uses, it sure as hell isn't a full scale replacement and you almost always actually have to massage the input into LLMs to get anything decent back out in practice. Some CEOs and managers can learn to do this, of course, and some already are... but that quickly turns into a second full time job. A "programmer" is still needed. The job might change from mostly hand-writing C++/JS/Python to prompt engineering + some manual coding to fix all the stupid fuck-ups that the bots can't solve themselves, but you still need someone to actually prompt the bot.
When that changes, it won't just be engineers losing work; there will be no reason to even have a human CEO any more.
> When that changes, it won't just be engineers losing work; there will be no reason to even have a human CEO any more.
The human race isn’t ready for that world IMHO. The only reason there is a middle class is because people have leverage in the form of their labor. When that becomes worthless … the people who own stuff and make their living from doing so won’t hesitate to get rid of everyone else - whom are now worthless to them.
Humans will revolt and mutiny on the ai ceo so fast its not even funny
https://abcnews.com/US/man-allegedly-throws-molotov-cocktail...
Imagine being deluded enough that you, a CEO, can continue to go buy groceries and drink at a cafe relaxing in that future
I don't know, if you've ever tried to build something at companies of that scale you run into incredibly boring problems "what data table do I need for X" and "who is the right person to reach out to for Y" and "they aren't answering me I guess I'll have to escalate"
I don't think there is any shortage of great ideas at these companies, they are just extremely bloated. And I don't think its something like indecision or bad PMs, it's "we have a finite amount of time and resources so we need to be conservative but also not too conservative"
If you have AI systems that can simply build out POCs in days, backtest on real data, show reliable results and numbers, you get a suite of product options you were never able to get before. If you have coding agents that can speed up implementation, you can build more stuff and choose the things that stick.
It changes the cost/benefit calculus of the entire business. I think you are exactly right in that: PMs/leadership are by their nature orchestration machines. Other roles are as well, but I think PM's are at a particular advantage here in that it will be quite awhile I would expect before core product decisions and creativity can be delegated to an AI, but not quite awhile until virtually everything that they're blocked on (legal approvals, POCs, wire frames, etc etc etc) will become less and less of a blocker
>If you have AI systems that can simply build out POCs in days, backtest on real data, show reliable results and numbers, you get a suite of product options you were never able to get before. If you have coding agents that can speed up implementation, you can build more stuff and choose the things that stick.
I'll also add this: within a large organization, you often need to interact with many different codebases owned by many different teams. Agents have made it much easier to wrangle by having the ability to deploy one to scope out your web of dependencies to learn about what would be needed for feature X, and how that integration can happen.
We've been doing far more away team work simply because it makes things move faster. It's easier to convince a team to sign off/review something than it is to get them to commit to the planning and eventual work.
It genuinely is helping things move faster inside large organizations. Or at least, it is for us, particularly since we're getting organizational prioritization to actually build the scaffolding to make those agents more effective at search.
> It's easier to convince a team to sign off/review something than it is to get them to commit to the planning and eventual work.
1000x yes: you have touched on what I think is the single biggest factor here, that is the humongous value of POCs. they are gnarly to build without agents, and so we used to have to get everyone on board so we didn't get screwed in performance reviews, which was monumental task because that means convincing very busy PMs who have a lot on their plate and dont want to take risks on things they don't understand, and now it's like "can we scale this out" and you have a very nicely formatted proposal and POC. It de-risks things very quickly
Legal approvals won’t be in that category.
You still want someone whose ass is on the line if they get it wrong.
Absolutely but you want to package it to them nicely and efficiently. The biggest blocker is legal and everyone else speak two completely different languages and we often don’t know what’s important to flag and legal doesn’t know enough to ask all the right questions. Also, many things can be templated, and in an industry where regulations and precedents change so quickly, agents are at the very least a good tool to flag issues (e.g. we were approved to use data X for Y but now decision Z negates this). The propagation of this information is not very effective now and legal review at tech companies, while absolutely essential, is somehow a worse experience than going to the DMV when it’s crowded.
Pieces of concept and other prototypes have always been cheap (see hackatons). The main issue is that as soon as you’re touching customer data or modifying process they’ve paid you for, you have to be really careful. No one wants to be responsible for an outage that cost the company its biggest customer.
Yes, but at scaled companies, where building a simple POC hooked into real systems is nowhere close to easy. To the point where it means that you might as well just do the full thing. That's where the enterprise spend and the impact is.
Isn’t that a matter of configuration management? Or do you want to alter the real systems as well?
historically it's been a matter of an absolutely horrific amount of Kafka-esque problems.
Say I want to build a feature in a product.
- DS has to do a deep dive (need buy in) to opportunity size and derisk with data. That DS has to work with other DS (people may have left or moved teams) to figure out how to get the right data and figure out what the difference is between 10 different tables that have overlapping but inconsistent data. - Eng has to build up an actual simple demo (need buy in) - Design has to make it not hideous (need buy in) - Legal has to review what you're doing; POCs should involve real data where possible because otherwise no one will trust it, even if its just for user analysis on existing products
This plus about 6 internal system bugs for custom tools that are flaky and who's team has long been re-orged or laid off, 8 people who won't answer you, 2 PTO's for the stakeholders, 6 weekly meetings
no one did POCs, they just had ideas and tried to get PM's to put it on the roadmap so if it fell through at least it was bought into
Yes, that exists at the wider business level. No question. I think what needs to get asked is are we talking about a bottleneck within the business as a whole, or a bottleneck within the scope of the knowledge work in question. Within software delivery there's a very clear shift when it's suddenly trivial to drop a 100kLoC plausible-looking PR into code review within an afternoon. Producing working code with a whole bunch of tests which make a very clear assertion that it does, in fact, work has had (if you're going that way) all the human-scale thinking time taken out of it, down to a rounding error. It still needs to be checked by a human, which was previously assumed to be a comparatively quick task in comparison to producing the thing. At least, it does where I am, and I don't think that's a silly position today at all.
If they can crack that latter review/spec-check/assurance step, checking that what was built was what was demanded of the problem such that we don't have humans in the loop at that step either, then the bottleneck moves again. Then I think it moves to requirements capture and to product development, but that might depend on the industry.
Trusting CodeRabbit for sign-off is "just" a small matter of configuration.
And convincing your org that such a thing is safe to do, which is decidedly not.
> As an example, what innovative amazing world-changing products have Google or Meta launched in the past decade
Kubernetes is at 11 years ago, and is huge enough to be included there. The Google Pixel was just under 10 years ago. So... not nothing haha
Numbers I see put Pixel at less than 5% of iPhone sales. Not nothing, but world-changing? I doubt the world would look significantly different had Google not done Pixel.
If you really think this you simply have no theory of mind for this stuff. There are tons of immensely successful products in the ad space that both of those companies have launched. They don't need to innovate in the product or technology space (doing so certainly makes a big difference in having more placement for ad real estate), but to suggest there have been no real innovations (specifically engineering specific innovations) related to ad tech would be completely ridiculous to suggest. You don't need to change the world to get rich, just look at wall street where major innovations have been made in the pricing models of fixed income securities.
Second to this are countless other areas that have a major impact on the companies bottom line that are entirely engineering driven, especially at google given they are a cloud provider and have meaningfully grown the workspace business and launched waymo in this time.
Google's internally developed and sometimes even launched plenty of innovative new products in the past decade. Stadia, Fuchsia, federated learning, and the whole transformer architecture that underlies this AI boom are good examples.
The problem is they get killed by some other executive who is afraid of their department looking bad by comparison.
I think this is fairly illustrative of the challenges in AI becoming as impactful as the Internet. The bottleneck is not making things. There are plenty of people who are really good at making things and can easily be 10x or 100x as productive as the average corporate worker. YCombinator was founded on that premise - small teams of founders and early employees could be orders of magnitudes more productive than the 1000s of corporate employees at their competitors.
The bottleneck is on bringing your product to market. If your innovative new product is built within a corporate environment, it'll get killed unless the executive you work under can get a promotion out of it, and you'll be denied all sorts of help with approvals, launch process, PR, marketing, branding, etc. If it's a startup, they'll try to shut you out with exclusive distribution deals, legal threats, lobbying efforts to change the legal environment, PR campaigns, FUD, etc.
The Internet was revolutionary because it let millions of people bring products to market without asking permission. Instead of having to bid for retail shelf space among dozens of entrenched competitors that all had sweetheart deals with the retailer, you could just put up a website and sell it to anyone across the globe. Instead of following hundreds of regulations that governed existing commerce, you could just launch something and sort it out later. AI doesn't really have that property - if anything, it makes things more centralized, with more gatekeepers, and so seems more likely to destroy economic value than add to it.
Google does not follow through in the long run on many of their pet me-too follow projects, however they do not stray away from their core remit making their real customers happy the ones who buy the ads…
Obviously that includes whatever needs to be done to hoover in data from their marks and Meta also does the same thing without fail and both are really good at it. But outside their remit not so much.
What I think is happening is that the scale of thing you can hope to build at a below-corporate scale should radically grow. Corporate environments should suffer for this, being that inefficient.
> YCombinator was founded on that premise - small teams of founders and early employees could be orders of magnitudes more productive than the 1000s of corporate employees at their competitors.
I think this is still true, but the theory is:
1. You don't need YC-type funding to do YC-type business any more; 2. You don't need to scale the business past those small teams any more, you just buy more tokens.
For clarity YC still obviously has a place as an incubator, mentoring, and networking function. I just think that what was previously the inevitable conclusion that you have to hire all the people the second you hit PMF to keep up with scaling the business as you scale sales is no longer inevitable. If you didn't want to go that way before AI, you were a "lifestyle business" and not worth investing in. As more and more knowledge functions get capably implemented by AI, it's the preferred position: humans are vastly more expensive than tokens, so you want them doing the stuff the AI still can't do.
I don't think this necessarily translates to mass unemployment. I think it translates to masses of smaller businesses that are radically more efficient because the handoffs between business functions are tool calls, not emails to someone who doesn't want to help.
> The Internet was revolutionary because it let millions of people bring products to market without asking permission.
Think about it this way: if I am a small business owner but I think it makes sense to do something that previously only a team in a corporate environment could do but is now within the reach of AI, not only can I do it now, but I also don't have to ask anyone for permission! Who wins between the corporation and the small business in that scenario?
> AI doesn't really have that property - if anything, it makes things more centralized, with more gatekeepers, and so seems more likely to destroy economic value than add to it.
I think this will turn out to be backwards. I can see a version of this where the number of things you can do without needing to turn to a gatekeeper for help increases to the extent that the balance completely inverts.
The vast majority of businesses are small, and AI can give them tools which previously required corporate scale to make sense, without the inefficient hand-offs between busy, political humans. Which is also something that the internet did! Getting an advert in front of a national market pre-internet was Hard but sometimes you had to do it because your target market was "all Canadians who buy toothpaste" or whatever and that meant saturation-bombing the physical environment with physical billboard ads, posters, flyers, and so on. So you only did it if you were P&G-scale. Now you, personally, can do it, trivially, for better or worse.
I dunno if the employees were ever really needed for scale. WhatsApp famously had 300M users and 13 employees at the time of acquisition; Instagram was something like 50M users and 55 employees. If you know what you're doing software scales basically infinitely, and the employees are there to make the software just slightly more tailored to specific user populations (and because upward career mobility for managers involves having more headcount). Yeah, building a revenue model takes people, but Valve employs only about 400 people and makes billions, as do various quant hedge funds like DE Shaw or RenTech.
The insta/whatsapp/plentyoffish model works if you're very lucky with both product-market fit and the technical constraints of the product itself. If you have something that technically scales extremely cleanly, it basically sells itself, and it doesn't need feature churn to retain or gain users, you're golden. I do think more businesses could do with checking whether they do in fact have that lottery ticket before hitting the scale button; there aren't that many examples around.
> Valve
Arguably a monopoly. They've got a product that sells itself with very low infra overheads for the income.
> Hedge funds
Very different model. I don't think the same intuitions apply.
Most of the people they have on staff are there to support their real customers, and I don’t mean the marks out in Internet land Google and Meta‘s real customers are the people placing ads and giving them money, most of the staff is dedicated towards servicing them, that again is where most of the money goes to support their real customers.
>I would argue that that's been the case for quite some time before AI.
I would agree but it's really minimized the building. More and more time is being spent on pre-coding work.
Google & Meta are illustrative of late-stage capitalism -- it's all about distribution, not innovation. Their job is (mostly) to just acquire the products that have passed the gauntlet, then scale up their monetization through their distribution-focused machine. The same dynamic plays out in virtually every industry (not just tech).
You'll find that most internal "innovation" teams are just lip service. In most cases, the "mothership" will be incapable of reproducing true innovation -- from a statistical perspective, culture perspective (mega corps are anti-scrappy; internal politics), and motivation perspective (startups aren't 9-to-5). It's much easier to have big M&A budgets, a VC arm, and some handwavvy internal innovation group.
Every now and again, you'll get real innovations (Waymo, transistors, GUIs), but even those have a spotty track record of commercialization when created internally.
The one I'd point out for that list is Kodak and the digital camera.
This is the same argument that has been historically made for outsourcing developers. Get 20 more devs for the cost of 1 dev in the US.
I suspect that AI will fail to pan out to the same extent for the same reason why outsourcing hasn't fully panned out (even though every company tries it after getting big enough).
The problems that will come up will be and always have been ongoing maintenance. AI is great at writing new code without a brain behind it, but once you get to the point where you need to refactor code, you start really needing someone with coding experience to guide the AI or veto it's mistakes.
I don't think that's really fixable even with a lot better AI. It's not something that ultimately comes out of the likes of github data.
I'm not saying that AI isn't going to make things better, btw, I just don't think we'll see a 20x improvement. Probably more like 1.5 or 2x.
Outsourcing of knowledge workers didn't work out because at large enough scales, the geographic arbitrage disappeared. Companies mostly always got what they paid for.
The determinant of success was only whether the task needed American-tier labor or could make do with sub-American quality labor.
I am not sure this feels right. I agree that the US currently has leading minds in terms of tech, but I am not sure it is too big of difference with the EU knowledge workers. EU is still a lot cheaper then US in terms of wages you would need to pay.
EU workers themselves get a lot less, but the EU is expensive because of 1) the huge payroll tax (45% in France) and 2) the challenges with hiring and firing mean you are carrying people that aren’t contributing.
Sure is an interesting thought. None of this is sarcasm: why do US companies deal with the time zone differences and language barriers they won’t need to bother with so much by outsourcing to say, Ireland?
The mechanism is often that they'll actually outsource to someone like Accenture, who have teams everywhere, and whose contract managers will try to get their cheapest viable team onto the contract to maximise their margin. If the buyer can't judge the quality of what they're buying, or doesn't know why the resulting hand-offs, delays, mistakes and rework will cost them more than keeping everything in-house ever would have, they're going to have a bad time.
Ireland isn't that much cheaper than, say, Oklahoma. And the cultural differences with Ireland are not a lot smaller than those with India or the Philippines or what have you, once you try to actually start working together.
(Yes, all the good developers from Oklahoma move out, but the same is true of Ireland)
Er, US companies do outsource to Ireland.
Basically every big tech has large offices and employ a lot of people there.
The limitation is that Ireland is a relatively small country, and most Irish developers are already employed (which is why Ireland end up being one of the main destinations for tech workers being hired from abroad).
That's certainly part of it. But the other part that I've heard time and time again is that in order for outsourcing to be successful you basically needed an american engineer in the mix hand holding everything, clarifying requirements, and vetoing bad code.
That part of dev work, the requirements gathering, attention to details, clarifying requirements, is something AI also struggles with. A lot of companies basically waste time and money on outsourced devs because without a clear path forward they effectively will sit and do nothing, waiting for a prompt.
I would not agree on that point. It really depends on company's structure. I mean it also depends with people that makes the team. I would say there are a lot of unknowns but I would certainly not generalize.
How I find your argument is that one distinguished engineer from US could do the same with the use of AI.
I worked with both and I know great and bad engineers from both sides. Only thing is that US has a bigger pool of great engineers.
I think the mechanism here isn't that American engineers are magic. It's that you need that contextual knowledge really close to where the work is actually being done, so that the turnaround for questions, blockages, clarifications, "we've got no work to do", quality level-setting and so on is on the scale of minutes, not time-zones.
It doesn’t have to be an American but it does have to be a direct employee of the company ideally working in the same time zone as management and the people defining the requirements.
Outsourcing of knowledge workers is still ramping up. The issue in the past was the skills were few and far between internationally. Facilities were also not built. That has changed now in a lot of fields. New campuses have been built in places like Bangalore and Hyderabad, even Singapore. The skills are there now, the training is decent, and you can see that the hiring is going on for very skilled titles in these cities. It is a different animal than just 10 years ago in this.
The “American tier” labor of course is distributed across the world and the top performers in every nation find ways to get paid at something approaching American salary levels. Look at all the international FAANG offices paying high salaries, in purchase pricing parity terms.
> I suspect that AI will fail to pan out to the same extent for the same reason why outsourcing hasn't fully panned out
My mental model for that is that outsourcing fails where the work is being done organisationally far from the knowledge needed to do it. We know that's true of teams inside organisations, there's been a lot of research on how distance in the organisational tree negatively impacts productivity. Outsourcing is a pathological worst-case of that.
The promise (promise! We're not there yet!) of AI is that I can have a cross-functional team on my laptop. Organisational distance is zero. Where previously the outsourced team has to wait for the time zones to roll round so I can answer their blocking question when I get to my email STRICTLY AFTER I have had my coffee, now it's a prompt in a chat window with a button I can click to make a choice in 5 seconds. Delay is gone, cost of delay is gone.
> The problems that will come up will be and always have been ongoing maintenance. AI is great at writing new code without a brain behind it, but once you get to the point where you need to refactor code, you start really needing someone with coding experience to guide the AI or veto it's mistakes.
Oh, absolutely. That's a minefield. Today. It will be, right up until it isn't. There are ways to set up agents and projects right now that make a dramatic difference to how this part of the picture plays out, but those will sink into the harnesses as time goes on.
But also the big problem with maintenance and outsourced teams tends to be the commercial structure around the contract. You get a Build team, who Build the Thing and then: no more features for you, anything you want to add past the original spec costs extra. They hand over to the Run And Maintain team, who get to fix all the bugs that the Build team left but without the knowledge gained from building the thing, but are scaled and located to be absolutely as cheap as the supplier can get away with so probably don't have the skill, inclination, motivation, or permission to take on any restructuring to make the bug fixing easier and they're on the wrong end of the globe so there's a 24-hour latency on any queries. It's a terrible way to set teams up, but it looks good on paper.
Again, that's peculiar to outsourcing and completely goes away if I have the same team that built the thing own the thing long-term. That's true if it's humans or AI!
> I don't think that's really fixable even with a lot better AI. It's not something that ultimately comes out of the likes of github data.
No, it's a harness problem. You need to start from a maintainable point and keep standards in place. It'll take work to get the harnesses there and it's not ubiquitous. You might also need better models, but I've already personally seen big differences in outcomes between projects that took certain steps and others that didn't; it's nothing revolutionary, mostly stuff that works for humans also works for AIs but you need to know to ask for it.
> I'm not saying that AI isn't going to make things better, btw, I just don't think we'll see a 20x improvement. Probably more like 1.5 or 2x.
I think people radically underestimate the cost of delay. I don't know if 20x is realistic for the AI itself, but I think it's not impossible once the inefficiencies of having to go to other humans is factored in.
> mostly stuff that works for humans also works for AIs but you need to know to ask for it
I'm most curious about this sentence. What have you noticed about the similarities? I'm getting really good at asking for confidence levels, tests and pushing back, but I'm curious what you found
Outsourcing also fails because it’s a pathological case of adverse selection. The businesses that outsource projects are ones who are organisationally incapable of managing those projects well internally. But, that inability extends to their oversight of outsourcing shops as well.
End result is that many outsourcing firms are borderline fraudulent in the way they treat their customers.
Who pays for that value, and from what, if all knowledge workers lose their jobs?
It sounds like the economy would largely reduce to the small minority class of independently wealthy people.
The more time I spend using agent tools the less I worry about knowledge worker job loss.
It takes a skilled knowledge worker to use these things.
Yes, but I do worry about junior knowledge worker job loss. These models are very good (and getting better) at the vast dark matter of "donkey work" that happens in knowledge-based industries -- work typically done by junior devs / analysts / lawyers / consultants, paralegals, admin assistants, customer success / support, etc. -- and those roles comprise the bulk of the workforce.
And worse, these are the tasks that help the junior people eventually grow into the skilled knowledge workers required to operate models, so there's a pipeline problem too.
I do too, but I think it currently has a lot more to do with the quasi-recession we've been in since the end of ZIRP and AI is a better excuse to stop training juniors than telling investors it's belt tightening, just like layoffs.
I'm already seeing tech execs/hiring managers getting very frustrated at the lack of new-senior-engineers to hire. The market will correct for this in time.
Curious if you can share any backing information from your last statement? As a senior engineer (well, that's my job title anyway), I find it encouraging.
This doesn't break it down by experience, and I can't find specific data on that, but the recent spike in demand for engineers + subsequent drop in unemployment this year is well documented [1].
The demand for senior+ engineers has remained steadier through this downturn from my anecdotal observations, with new grads being by far the most negatively affected, but even that seems to both be shifting from talking to people a handful of years younger than me + CS enrollment has already precipitously declined [2] as the narrative that programming is dead because of AI has spread rapidly.
All that leads me to think it's going to be a junk-show over the next decade for people trying to hire as the pipeline was destroyed.
1: https://www.citadelsecurities.com/news-and-insights/2026-glo... 2: https://www.washingtonpost.com/technology/2026/04/13/compute...
We'll get around to training job specific models or the equivalent. Thats just lower on the value chain for now.
Sure. I was challenging the parent on how the “game” they are positing would play out.
See https://news.ycombinator.com/item?id=48300427 for an alternative take. I don't think either direction is inevitable, yet.
To follow on from that comment, if the growth in breadth of capacity of AI leads to a decrease in the risk of running a smaller business, which I don't think is an unreasonable prediction, then it's not inevitable people do lose their jobs. Employers get smaller, higher-leverage, and more plentiful.
There were no knowledge workers in the middle ages.
Back then people were mostly farmers, but we already automated that job away.
Not completely, but compared to the middle ages we 50x'd their output. Which is a great illustration what it means to make a job 50 times more productive. We went from 80-90% of the population being required to barely make enough food for everyone to survive, to 4% of the population producing such an abundance that consuming too much food has become a systemic health issue
At the mere cost of destroying soil, and polluting water and the atmosphere in only 200 years! Possibly this will also play out well, and there is a huge market of... maybe social media influencer economy to pick up those being automated out of their previous work... or rather identity, as actually much like in the middle ages, the modern world also makes the profession largely the identity of the individual.
I'm pretty skeptical on the outcomes and the costs also (natural and social as well), but possibly we can have 50x or even more software in the end! The phrase will be truer than ever:
> Software is eating the world!
Maybe ironically, but software and robotics should allow us to scale regenerative agriculture in a way that doesn't leave everyone in poverty. We already have lasers mounted to trailers doing precise weeding instead of broad herbicide usage.
https://www.agtechmarket.net/news/laserweeding (random web search, I don't vouch for this site, it just looks legit at a glance)
Next innovation could be to scale succession planting, which keeps the ground from being exposed in between crops and lets you transition from nitrogen fixers to users quicker, getting more food out per acre while reducing fertilizer usage. But you can't do that with current harvesters and human labor is too valuable to spend on this.
Also take broccoli harvesting, typically you get a few big heads, then it keeps producing smaller heads, but it's not economical to harvest the smaller heads with human labor. Robotic harvesting lets the same plant produce more food per acre and uses the energy needed for new plants instead to keep producing food.
Masses will be unemployed, due to robots displacing them, but human labor will also be too costly. We won't be able to afford a person shepherding, but we will need to produce "meat" (substitutes) in plants, or in inhumane animal-jail, and we'll need robot-weedkiller lasers to produce the feedstock instead of letting animals graze... and we'll give the food produced this way to people on UBI...
This is where this is going, the whole industrialism is totally self-serving, and for every problem its answer is digging deeper in the rabbit hole, and creating 2 more problems in addition to solving the initial problem only half-way.
I don't want to say what you are suggesting is not possibly useful, I just want to emphasize how stuff works out in reality, in addition to doing some nice stuff like what you called out (the halfway solution to the problems). All we get is more alienation and humans getting depressed and feeling a lack of purpose... but somehow we cannot afford to pay fair prices for the agricultural work, and pay fair prices for the food, and not overproduce and overpollute... and the same thing is happening in every aspect of the human condition, not only food production, which is the most basic and ancient activity we have been doing.
Cattle grazing is helpful for fields left fallow, but succession planting is far superior in so many other ways. You can mix plants which repel particular pests with those susceptible to them (and other beneficial strategies), topsoil is grown instead of depleted, flowers are present for wide range of season so bees naturally thrive with food always available, you don't need a significant generator of greenhouse gas running around (cows), and it gets more vegetables per acre so it would be good if vegetables were cheaper because we don't eat enough of them.
I have done succession planting in my home garden, but it's definitely not worth the time investment for the food alone. But it's real neat to see your aphid problem disappear as the nasturtiums pop up without any pesticides needed. You can even feed the world with it, if most everyone wanted to be farmers... (as opposed to some Organic practices which is the same mass farming but the pesticides are "naturally-derived")
Farming has been destroying soil and polluting water for thousands of years. The Tigris & Euphrates used to be crazy fertile, now it's desert. Yes, the destruction has accelerated but farms now feed 8 billion people.
There definitely were what could be considered knowledge workers in the (high) middle ages, it just wasn't the majority of work like today. The knowledge workers then were just a tiny, elite faction, mostly employed by the church or directly by nobility. Kindgoms were still big bureaucracies and needed scribes, theologians, academics, lawyers.
Relatively few anyway. Monks (who wrote and edited books and managed libraries, and also taught), artists and musicians, bookkeepers/treasury/exchequer, scribes/chancery (who were the administrators of the kingdoms), and bankers all existed in European "middle ages". But a significantly smaller part of economy/society compared to "western world" now, yes.
There wasn’t 20x value to pay for in the middle ages either.
Are you sure? Any functional organization requires keepers to oil the machine. First the government. The best examples were the chinese empire, the catholic church, and the various kingdoms. Or do you think that everyone was either fighting or farming? Stewardship is knowledge work. Bookkeeping is another.
> Who pays for that value, and from what, if all knowledge workers lose their jobs?
They do not care unless these companies can get a bailout.
UBI only exists for companies that are too big to fail. Case in point, 2008 and SVB when there was too much money on the line.
One of the AI companies attempted to guarantee themselves a way for the government to bail them out if they were close to defaulting on the debt from the data center build out.
SVB didn't get bailed out, their investors and creditors were wiped out. You could argue depositors were bailed out -- as they took the undue risk of keeping more than $250k in a single bank (though as part of a requirement for getting a loan from SVB, you had to have your operating accounts with them. So lots of companies had no choice, as SVB was one of the few banks that would lend to them).
Arguably, the main impact of securing SVB depositors above the $250k limit is that it prevented thousands of people from being laid off that week, as their employers wouldn't have had the money to make payroll the following Wednesday.
Thank you for saying this. Continuing to point at SVB as a bailout is annoying. They were not bailed out. Anyone with deposits in an accredited bank should be made whole - always. Without trusted banking we have no economy.
> Anyone with deposits in an accredited bank should be made whole - always
Sure, but is that the case now? Is everyone made whole when a bank fails and they have more deposits than the insurance limits? Or only when it's the well-connected / too-big-to-fail?
Looks like the answer is no: https://www.wsj.com/finance/banking/a-small-banks-failure-le...
So I don't think it's unreasonable to describe SVB as a bailout. Not for the investors, but for the depositors. Has anything changed to reduce the moral hazard / make it less likely to recur?
> Sure, but is that the case now?
Pretty much and has been for awhile.
https://nyulawreview.org/wp-content/uploads/2025/05/100-NYU-...
In early 2023, within the span of two months, the United States experienced three out of the four largest commercial bank failures in U.S. history, as Signature Bank, Silicon Valley Bank, and First Republic Bank all toppled.1 Yet, despite these banks having roughly $300 billion in uninsured deposits at the time of their failures2 and despite the failures costing the Deposit Insurance Fund (DIF) of the Federal Deposit Insurance Corporation (FDIC) an estimated $38 billion, uninsured depositors took no losses in any of the failures.3 While these results were striking, they were far from unusual. Since 2008, uninsured depositors have experienced losses in only 6% of total U.S. bank failures.
...
Formally, the United States caps deposit insurance at $250,000 per account,6 but, in reality, the post-2008 financial system comes close to providing de facto total deposit insurance covering all amounts in all accounts.
So we all now know that a bailout DID occur with the SVB depositors who had all their money in the bank and most deposits were over the FDIC insurance limit. The FDIC insurance rules somehow didn't apply here because there was too much money at risk. (And too big to fail).
But if there was a bank failure at a regionally smaller bank with a regular customer or startup depositing the same amount of money over the insurance limit, their money is gone.
Just like Intel got a "bailout" from investment as chosen by the US government, AI will eventually have a very similar story.
> UBI only exists for companies
What's the U stand for in UBI?
Producing a thing has always been cheap since personal computers existed. From mail-order software companies' times to SaaS times, producing a sellable MVP was an initial cost that is relatively small compared to the later cost of expansion and maintenance. Marketing and selling was and still is the hardest part.
Why do you think of knowledge workers as a fungible commodity?
What makes you think the people who used to build (or would have built) software will switch into the industry of "knowing that the thing was the right thing to build", as opposed to something cooler like surgery, city planning or experimental physics? The roles within a tech company are not the only jobs in the world.
> Why do you think of knowledge workers as a fungible commodity?
I don't.
> What makes you think the people who used to build (or would have built) software will switch into the industry of "knowing that the thing was the right thing to build", as opposed to something cooler like surgery, city planning or experimental physics?
Because it's probably already part of the job. It's a change of emphasis, not a change of career. Your boss can already ask you to do it. If you're producing code, you're probably also reviewing code, checking it matches the acceptance criteria, testing it, sanity checking that it was the right code to have been written, today.
> The bottleneck has moved from producing a thing that works to knowing that the thing was the right thing to build
“There’s more capital than good ideas to fund” has been a complaint from the likes of A16z & other VCs for a long time now. It’s why we ended up with stuff like NFTs getting funded.
That’s very unimpressive return on investment compared to what was promised.
If knowledge workers get laid off in mass, you can expect political curbs on AI adoption.
I will also tell you, as someone who works at a company that's trying to remain profitable, that token spend has caught the eyes of finance and much like cloud spend they've already started applying pressure to control costs. This May my team is protected to use 30% fewer tokens than we did in April - this was by intention. I suspect we'll drop more in June.
I expect in the future, when these AI companies stop subsidizing costs, the idea of spinning up 20 agents to work on some brain fart idea that you throw out after looking closer will come to an end. It'll be seen like assigning developers on work that hasn't been properly planned for or reviewed.
It might be time to start interacting with agents using grug speak only
https://github.com/JuliusBrussee/caveman
Can't wait till June, when finance gives the team the choice: everyone gets double tokens if you choose to fire somebody.
Oh we already had that with a RIF earlier in the year.
> They've got, ballpark, $5t to $10t to make back in the next 5 years
OpenAI's spending commitment is in the ~1T range for the next 5 years, and Anthropic is ~300B.
If they continue to show strong growth, they likely need to be at 100-300B in revenue/yr to support their yearly payments + financing, not 1T.
> They've got, ballpark, $5t to $10t
What are you basing this on? For reference, Anthropic raised ~$70 billion in total and OpenAI ~$190 billion. Why do they need to make 20-40x that?
All the planned infrastructure commitments. At least for OpenAI I think they're supposed to spend $300+bn in the next few years.
I still don't understand why that means they need to make 5-10 trillion over the next 5 years.
I think the original argument is too limited in its scope. The wider AI market, which is primarily fueled by OpenAI, Anthropic, Google and the large frontier labs (are there any other in the West, except for these 3?) is spending how much... $900bn this year in DC buildouts? After the spent $500bn last year and they're probably planning to spend just as much the next few years if things go remotely their way.
So yeah, I wouldn't be shocked that in the 2023 - 2033 timespan total AI investment worldwide will be around $5tn, maybe even going towards $10tn.
All that money will have to be repaid, and it will have to be repaid 10x, otherwise heads will roll.
The enshittification we've seen so far is nothing compared to what's coming.
> ... against the actual work their company cares about doing. [...] stuff that matters
This is a key point. Some engineers are having fun doing e.g. greenfield stuff with AI that they never would have had time for otherwise. Whether the company cares about that is another question.
It's related to Goodhart’s Law. If AI token usage is a target, then you're going to get a lot of token usage, but it's not likely to correlate well to improved business outcomes.
> That's a _huge_ shift. Most people I know cite +20%-40% velocity with these tools, against the actual work their company cares about doing.
We all have our own observations and mine don’t significantly diverge. But that’s bottom up. At this point shouldn’t we be seeing it top down?
If we are beyond potential and into significant productivity gains, why isn’t that showing up for the customers?
Why didn’t delta airlines get significantly more operationally efficient in the last 3 months due to the introduction of better software?
This is a genuine question, I am seeing a disconnect.
Anecdotally, my take on this is that biggest value lever is strategy and alignment, not implementation. The typical company is dozens of little vectors pointed in different directions, and they cancel each other out. Scaling up the magnitude of each is still net zero.
I was recently consulting at org where two separate engineering teams were all in on two different, incompatible deployment platforms and using AI to accelerate adoption of each.
Management was mystified why their engineering leads kept telling them they couldn’t deploy a complete implementation of their solution.
> Why didn’t delta airlines get significantly more operationally efficient in the last 3 months due to the introduction of better software?
The coding agents got good in November. Most individual engineers didn't fully clock this until January/February. This means that companies didn't really figure it out until March/April.
Assuming companies like Delta have adopted coding agents (which would be pretty fast) it still takes months from adopting a new tool to the code results of that tool rolling out to production.
I expect (and would hope) Delta's software development culture is very conservative. Since nobody can confidently tell Delta "here are proven practices for using this tech to produce high quality, more secure code" yet it would be surprising if they were blasting full-steam ahead.
I expect that even companies that got on board with coding agents in January will only just be starting to ship user-facing features that benefited from those new tools. Shipping software takes a long time, no matter how much faster the "typing the code in" bit gets!
The fuck is going on with HN that a comment making up completely fantastical numbers is on top?
Perhaps it's not only CEOs who are delusional.
I thought Anthropic and OpenAI's combined CapEx has been <100B?
source: https://isaiprofitable.com/
That site needs Apple on the list. ;-)
Why? All their money is going to Apple Silicon and the five ecosystems, so far in Apples entire history, the largest acquisition has only been $3 billion dollars, OpenAI is currently getting nothing and they gave Google a measly $1 billion refund per year for the use of Gemini.
If John Ternus wants to spend some money, spend it on bringing memory in house. Apple has the money and the engineering talent to do so, have it fab/made onshore in partnership with TSMC.
Do it Apple because you have to not because you want to the Chinese probably will be taking over the memory industry, worldwide, by taking advantage of the greed from three memory companies and their AI overlords.
That's the point. To show how they _haven't_ lost billions on this.
Maybe so far, but they've committed to well over a trillion in future capex.
And there's the indirect capex that their revenues will pay for indirectly, like in the case of oracle
Here's the question - does that future spending already appear on partners' balance sheets
> Most people I know cite +20%-40% velocity
Seems roughly right, that does seem to be about the boost in the most well-suited cases where you essentially know exactly how to solve the problem, the problem won't change much, and it's truly a matter of just churning out the implementation.
In that case precisely prompting, doing the review & nudge loop, can be a pretty nice (nice, still not game changing) speed boost over literally typing out the code to match the design in your head.
The less optimistic view though is that most things you build aren't like that. Even if they seem like it first. These things get booked as a nice speed boost, but you'll only find out much later they weren't.
A confounding factor is that it seems like many people not in the detail of building software do seem to think of most to all things are like that, even before AI assisted coding. Not much need to say more - see the entire history of the 'agile' movement for evidence of this.
And because most things aren't like that, I actually struggle to see fundamentally how more than 20-40% will ever be achieved (short of the ever-present deus ex machina of AGI argument), simply because the generation is already really good for these types of things. So since things like this aren't going to increase in overall proportion of things to be done, I don't see where the overall extra gains come from by models improving at this point.
> We're talking about a world where you need 5% of every knowledge workers salary to go into tokens.
They are assuming ~10% global GDP growth instead of ~3%. You probably don't need the same %s if the pie grows a ton.
I'm highly skeptical we get that growth, but if you aren't, it makes it easier to digest.
I mean this case with AI-productivity fires itself back when we talk about GDP.
The more AI causes productivity increases, the less and less number of workers will be needed. This will heat up the job market even more and bring salaries down.
Net effect of this productivity increase: less consumption by the masses, even though you may be producing more good and much more efficiently.
A third effect also comes into play that once all this starts to happen, common people, who are generally living paycheck to paycheck, will now start to hesitate towards making any long term investment, housing included. And that indirectly will end up impacting financial and banking sector, which will then impact existing savings, bonds yields and retirement funds, and the recession-like cycle starts.
This productivity increase only makes sense if it is capped to a very small number.. like 20% max. Beyond that, who these companies will even be selling to?
Am I overthinking all this?
> The more AI causes productivity increases, the less and less number of workers will be needed.
That only holds if companies have a fixed need for "productivity" which is met by their current employees, such that their employees becoming more productive means they need less of them.
Every company I've ever worked for has wanted to achieve way more than they are able to get done with current resources.
But generally yes, the biggest open question about all of this is how the impact will play out on the economy, job opportunities etc. I've not seen anyone come close to a confident prediction about how this will play out.
> Every company I've ever worked for has wanted to achieve way more than they are able to get done with current resources.
I mean sure. Every company wants an infinite addressable market. But that doesn't mean it exists.
It might not be possible to sell 10x the software we sell today. It might not even be possible to sell 2x
It's hard to imagine how making insurance sales cheaper for the brokers, churning out astrology apps faster, AI boyfriend bots or running ad campaigns with fewer and lower paid designers is going to drive 10% GDP growth in developed and middle income countries, that's the sort of figures you see when very poor countries finish rolling out electrification, sanitation and transportation.
>The more AI causes productivity increases, the less and less number of workers will be needed. This will heat up the job market even more and bring salaries down.
>Net effect of this productivity increase: less consumption by the masses, even though you may be producing more good and much more efficiently.
Big tech companies can't even create login flows and account recovery flows that work for everyone yet. There are countless stories of folks losing access to business Instagram accounts that get hacked, Google support from a human to fix a problem that is outside of their help articles is non-existent, etc etc. There's still so much "low-hanging fruit" IMO that isn't particularly fun or exciting to fix, but ask your average non-tech friend or family member what they think of the Facebook + Instagram security settings pages / sites / desktop-only settings.
Who is going to pay for all of these subscriptions that will power this GDP increase when average purchasing power of those outside of the top ~10% of earners is decreasing YoY? We're headed toward food and water shortages next to sprawling datacenters, not shared societal prosperity and a healthy middle class.
First of all, common people are not living paycheck to paycheck in the sense that they're at risk of not having money[0]. This is corporate content marketing that has entered the collective memory of people, not anything close to reality.
Secondarily, reducing the cost of making a thing doesn't always mean you get less of a thing. For me, certainly, what happened is that I write way more software than I originally did. When we built compilers, the amount of human engineering effort required to do things plunged, but the amount of software engineering jobs didn't go down.
This is as bad as models will ever be. That part is true. And it's entirely possible we go foom. But it's also possible we don't, and then it depends on where the asymptote lands.
0: https://www.slowboring.com/p/this-economic-myth-needs-to-go-...
Respectfully, that is truly ignorant. The vast majority of humans do not have any savings and would be in big trouble if regular income ceased. No paycheck no food. It’s wage slavery and it’s pervasive.
>Am I overthinking all this?
Nope, if AI were to realise the hype, you have to take into account macroeconomics. Usually this isn't a problem for most businesses
>The more AI causes productivity increases, the less and less number of workers will be needed. This will heat up the job market even more and bring salaries down.
People also underestimate that the reason why companies are so excited about AI isn't to increase productivity, its to fire workers and crack down on worker rights. They won't lay people off because AI means they don't need as many people to get the job done, they'll fire everyone while doing a much shittier job, because they hate having to abide by worker's rights and pay people
> The more AI causes productivity increases, the less and less number of workers will be needed.
Why does this have to be the case with AI but it didn't have to be (and wasn't) the case with the steam engine, electricity, the automobile, or the computer & internet?
Certainly, AI could be different.
It's curious to me why the vast majority of people on here think it must be different.
Because the previous revolutions only automated a small subset of jobs, it didn't automate manual work.
Some people take the view that AI could make knowledge work largely irrelevant. Any niche humans could carve for themselves would only live long enough to generate training data for the AI to automate.
> The more AI causes productivity increases, the less and less number of workers will be needed
This might not necessarily be true. Increased efficiency creates induced demand to the point where more workers are needed. Because the new capabilities unlock more value to extract and the economy rushes in to get it. The steam engine is a huge example of this
I dont exactly know what new value genAI will unlock but i think its more likely than not
And yet the job everyone loves to hate, the humble "burger flipper", continues to resist automation yet command minimum wage labor rates. This future of either being a CEO of a company consisting primarily of AI agents building some monthly subscription-based solution to some trivial digital chores OR manual labor that isn't [yet] fiscally viable to automate seems quite bleak. We'd also need a ton of robot technicians and manufacturing that the US has neither the educational and training institutions to support nor the will of the population to fill. Given the ongoing war on immigration, visas, and foreign-made hardware, if this continues, good luck.
This would be a Bladerunner future Pope Leo XIV warned against (https://news.ycombinator.com/item?id=48265206), though in different words.
Hey, I wrote this down one time. I estimated way higher yearly revenue required, to be adversarial. And you can keep the "cost per unit AI work" a parameter and play with the results.
But the point is that if people are willing to delegate part of their salary (e.g., buy consumer products), vs requiring employers to pay for the tokens, then it's quite possibly a net win. Something like "I pay a largeish fee every month to make my own job much easier", similarly to how we buy a car to make commuting easier.
https://jodavaho.io/posts/ai-jobpocolypse.html
1. Global IT spend is $6T per year
2. Where does this $5T number come from? If they make $4T in revenue over the next 5 years instead, what happens?
You are making the assumption that the models are only used / paid for by 2.5% of the population (your knowledge workers value). There will be new value created by these models which people are happy to pay for which simply did not exist at all before. It is also naive to say that the hyperscalers are going to be expecting a return on this in 5 years, it will be entirely propped up by investments / IPOs as has been the case with any tech company for decades now to reach scale. The hyperscalers are currently spending ~650b combined annually, which they have the cash for and can sell in future compute instantly.
I'm sorry, what the feck does "value creation" mean here? I live in a place where people are so, insanely squeezed from every angle. Wages are stagnant, prices rocketing. Where is the money to pay for this value going to come from?
No one I know feels richer than they did a decade back. I've not been able to meaningfully put up my prices for a decade. People are tired and stressed and scared, particularly scared of a technology everyone keeps telling them will make them redundant.
There is no rising tide lifting all boats, just most of us drowning whilst a few whizz past in their yachts.
I honestly hope these guys faceplant ASAP. Couldn't happen to a nicer bunch of people.
Feelings aren’t fact. A lot of data shows the doomerism is not reflected in the actual numbers and much of it has to do with rapid inflation and continued vibes.
Consumption has risen, inflation adjusted wages have risen for blue collar and white collar alike. Most social mobility has been the middle class moving into the upper middle class, not moving to the lower class.
The main thing holding people back is the housing crisis. This is orthogonal to the value creation of businesses.
Value creation is growth. If it didn’t exist the S&P would still be 42.55$.
> The main thing holding people back is the housing crisis. This is orthogonal to the value creation of businesses.
This feels wholly at odds with saying most social mobility is upwards. So most of the social movement is into a class where a home and vacations are a given, but we also have a growing class of people who can't afford a home? Per BLS, average real wages are down 0.3% YoY https://www.bls.gov/news.release/realer.nr0.htm .
> Value creation is growth. If it didn’t exist the S&P would still be 42.55$.
This reductively assumes "value creation" is the only effect on the S&P pricing. You'll note a ton of graphs correlate with it, e.g. https://tradingeconomics.com/united-states/inflation-cpi is the US inflation rate, which also tracks the S&P pricing. Ie if a company is worth $100 a year ago and inflation was 4%, I'd expect to pay $104 for their stock with 0 value creation whatsoever.
Home ownership is not the definition of economic class.
The s&p has vastly outstripped inflation, this isn’t even an argument. It’s a very bizarre and uninformed opinion to say “inflation is correlated with s&p value”.
In economics, you deduct the inflation rate from growth to get the real rate of return.
I wonder why so many people with such little understanding of financial markets make comments like these.
https://www.multpl.com/inflation-adjusted-s-p-500
The upper decile of income earners account for more than half of all consumption in the US. Household balance sheets and wealth have never looked stronger, again when you account for all the appreciated stocks and properties owned by the upper quartile. True incomes for the lowest decile rose significantly for the first time since 1970 in 2022 and then sort of stayed flat again. Sure, statistically significantly, not "significantly" as in personally meaningful after figuring in rising consumer costs. There is a narrative where you can see all this as hugely positive but this is also largely a "vibes" based narrative. I don't know why you'd expect most people to care about what the "vibes" are like for the best off in society, that's a bit removed from their daily concerns.
> Feelings aren’t fact... much of it has to do with rapid inflation and "continued vibes".
Oh the lost irony.
Is it ironic? Or did you just read the comment incorrectly?
> Consumption has risen, inflation adjusted wages have risen for blue collar and white collar alike.
My wages haven't risen for nearly 5 years, while inflation has occurred over the past 5 years. Why the blanket statements?
> The main thing holding people back is the housing crisis. This is orthogonal to the value creation of businesses.
Are you suggesting a "housing crisis," in your words, wouldn't impact consumption? I'm watching my spending (and living like a child in his parent's house, except it's not my parent and I have to pay for it) in the hopes that in about a decade, I'll have saved up enough of a down payment for a home somewhere in my state that I could actually afford the mortgage on the remaining amount. There are plenty of things I'd potentially spend money on but won't as long as I feel like I'm economically stuck and have a chance in hell of saving my way out of it. So this feeling translates to fact.
If you think my personal experience is just an anecdote and doesn't count because it's not being told through the lens of large-scale numbers, fine. But I really agree with the person you replied to that you're gonna have to be a whole lot more specific than "value creation" if you want people to spend money on your AI products "in this economy," whether it's because they're actually strapped for cash or just pretending like you seem to think they are.
Sounds like internet sentiment and not research data.
It's kind of become socially taboo to not be suffering "in this economy", but on paper it's hard to see weakness in places that there isn't always weakness. As long as the 65-95% are doing well, there isn't going to be a collapse.
The most recent U Michigan 'Survey of Consumer Sentiment', which is THE authorative source in the US, shows consumer sentiment at the lowest levels since the survey started in 1977
From the U Michigan page: https://www.sca.isr.umich.edu/
or from the FED https://fred.stlouisfed.org/series/UMCSENT
Sentiment is just vibes though.
Everyone hates Walmart but they still do gangbusters numbers.
So people may be hating, but they sure as hell are also still spending
https://fred.stlouisfed.org/series/PCEC96
A literal example is that I can use AI to file my taxes instead of spending a weekend and hundreds of dollars to have an accountant do it for me. It costs me like $5. that 245$ delta is the value of that output to me, as long as I am confident it is correct.
Seems to be a thing in the US to need specialised software, an accountant or AI to file taxes.
In most of Europe individuals at least don't need any of that. I'm in France and it's just a connection to a government run website to enter a few figures, takes less than an hour most of it is already pre-entered (salary etc), the main thing to add manually is charitable donations.
If you're running a business then yes an accountant can be good (or be required depending on the legal form of the business) but not for individuals.
In the US you don't need software or an accountant either. People have been convinced it is hard, but the reality is taxes are basic math for most people. It isn't hard to copy the few figures from a couple forms and add them up, the look things up in a chart. It is tedious, but not hard. That said, I use the software anyway because once in a while I make a human mistake and those are annoying when the IRS catches latter.
There are a few people who have a complex situation who need an accountant. Most do not, but since they have never looked at what is really involved they don't realize what is going on.
Taxes are one of those things that seem difficult and people reach for tooling or expertise without trying initially without, but are pretty easy to do yourself just filling out the forms.
It really depends your situation. If you’re entrepreneurial, self employed, others things become messy fast, and you cannot just know what you need from filling forms
Part of the value of paying an accountant is that you can get representation in case you are audited. Though I guess you did say you were confident it is correct.
This also sounds like rich people problems. Vast majority of people with a W-2 take the standard deduction.
I think that to sum things up, we will have to wait until we can evaluate the cost of the mistakes. You could be lucky but you could also end up with a very negative output value in the longer time frame.
I did my taxes this year too with 5.5 and 3.1
Otherwise normally costs around $800 to do, because I have a small business too.
> as long as I am confident it is correct
Are you? Does it cost you extra (time or money) to be?
Yes, and they were accepted. A year or two ago I would have been less confident but now almost UX is happy to cite sources.
Not speaking to the wisdom of filing taxes using LLMs, but just FYI (assuming US here) taxes being accepted doesn't mean they were correct. It just means the IRS hasn't found anything major wrong (e.g. SSN used on multiple returns). Even being approved isn't a guarantee, an audit could come later.
Even if an audit never comes they could be incorrect.
Audits often find incorrect in favor of the person audited as well. If you are not audited that is a bad thing as you have no idea if things are wrong, and it could be costing you a lot of they are wrong.
For sure.
Thats the thing; the "increase in productivity" isn't being felt by the general public, the end user. If your "increase in productivity" just means more money being shifted around at the corporate level then it is meaningless.
> There will be new value created by these models which people are happy to pay for which simply did not exist at all before.
True, but I think the GP's point was that what consumers will pay won't be nearly as profitable as what enterprises will pay to increase the output of their developers and knowledge workers. ChatGPT is currently the overwhelming leader in consumer AI usage but only ~5% pay $20/mo.
As a recently retired serial tech founder, I'm now one of those consumers. I use AI webchat daily for general search, Q&A and even to write little automation scripts for myself, yet I haven't paid anyone anything for AI yet. Even after being heavily restricted and performance nerfed to hell in recent months, free webchat AI is still fine for everything I do, and I'm not remotely price sensitive.
Even as AI compute costs fall over time, I doubt serving ads against AI webchat to consumers will generate the kind of high-margin, sustainable growth VCs get excited about. It's so undifferentiated I bounce around between all four leading providers because there's virtually no moat locking casual consumers to any chatbot beyond a single question thread. I guess if it had a nearly infinite context window seamlessly integrated across all sessions, that might be somewhat sticky for some consumers but it could also get creepy for some others - and it would devour gobs of the scarcest resource in AI. Beyond Maslow's Hierarchy of Needs, the mobile phone is the largest revenue, long-term mass consumer product ever but I just got a new flagship phone from a top-tier provider for $30/mo over 3 yrs. IMHO, even an all-you-can-eat, infinite context window, next-gen Mythos couldn't reach and sustain mobile phone levels of global consumer adoption at ~$20/mo. Unlike professional developers and knowledge workers, consumers don't have any "job to be done" big enough for an LLM to command that much of their zero-sum discretionary spend.
100%, a driving factor will likely be how good we can make models that are so small they use almost no compute. Until then it is a race for adoption and moat-building (or screwing people over?) once you have users
> a driving factor will likely be how good we can make models that are so small they use almost no compute.
That will certainly help but it doesn't move the fundamental limit because resource efficiency is a cost driver not a demand driver - and my argument is against the thesis that lying beyond professional devs and knowledge workers, there's an untapped trillion dollar industry serving LLMs to mass global consumers.
Using Simon's cost estimates, I agree that halving the current $1,000 - $1,200/mo MSRP to profitably serve frontier inference to professional developers and knowledge workers (PD&K) will help Vendor A steal share from Vendor B or C. It will also increase LLM sales penetration into the segments of the global PD&K TAM which can't afford ~$1K/mo for every seat. A fair chunk of the PD&K workers in many SMEs aren't included in today's ~$1K/mo per seat license pool, especially in 2nd and 3rd world geos. When the price falls to $500 and $250 most will but that's still just saturating the existing PD&K TAM - not pushing into mass consumers.
While the PD&K TAM is big, justifying Trillion+ dollar capex spend requires believing the TAM is much more than PD&K and eventually grows into converting a couple billion non-PD&K consumers into ~$20/mo subscribers. I don't buy it for two reasons:
1) The Comps: There are vanishingly few examples of long-term, mass consumer adoption of a discretionary technology at that scale. Mobile phones at ~$15 to $30/mo are the obvious one but LLMs are nowhere near being that valuable to the average plumber in Des Moines, baker in Jakarta or retired nurse in Hamburg. Pondering it, I just imagined forcing any of those people to choose between their mobile phone and an LLM chatbot. Sure, some who are flush with cash might choose both but for most consumers in the world ~$20/mo is big enough they'd have to pick one and ~zero percent would choose the LLM over their phone. After mobile phones, the second comp for discretionary tech spend I thought about was XBox and Playstation monthly gaming subscriptions but combined they have less than 90M paying subscribers and the ARR is just under $10/mo. As an industry, "Big LLM" is spending well over a trillion dollars every five years. XBox and PS ARR doesn't even cover paying the interest on that capital, much less the 3 to 5x returns hedge fund investors are betting on.
2) The Alternative: It's useful to doubt my own intuitions and one counter to my skepticism is to assume "But LLMs aren't finished yet, they're going to get much better." How much better could an LLM which can be profitable at ~$20/mo get than Claude Mythos in the next five years? Instead of debating future unknowables with myself, I've found it's better to just imagine the most perfect future product I can that's still realistically plausible. So, let's imagine we're willing to spend a million dollars a month to very unprofitably deploy a prototype to test the consumer demand for "Tomorrow's Awesomest $20/mo LLM" today. So we gather a few hundred super smart, broadly knowledgeable intellectuals together at one top-tier university research library, where they'll have access to every commercial database and unlimited Claude Mythos 2.0 and ChatGPT 6.0. Since our experimental budget is $1M/mo we can afford to add in several Nobel prize and Fields Medal winners too. They'll work together manually reviewing and improving not only every LLM answer but also our test user's prompts - and of course our test chatbot will have human-level real-time speech recognition and vision (via Zoom and screen-sharing with actual genius-level humans), making this truly a test of the "smartest, most accurate, best consumer chatbot" we can imagine.
Now, let's run the test by having one thousand mass consumers try it out and see how many Des Moines plumbers, Jakarta bakers and Hamburg retired nurses we can convert to a 1 year @ $20/mo subscription for our $1M/mo ultimate chatbot simulation. Playing this thought experiment out in a bunch of ways, I find some percentage of outliers, iconoclasts and closet intellectuals would go for it but... the vast majority just don't find it enough better than "free" chatbot alternatives AT&T includes with their phone subscription or Samsung bundles with Galaxy phones - despite only being ChatGPT 5.4-level. It turns out, most plumbers, bakers and ex-nurses don't have a compelling "job to be done" in their daily lives that even an MoE panel of actual Nobel and Field's medalists with ivy league professors can make enough more valuable than an inferior but free-to-me chatbot, in the judgement of our Des Moines plumber. While the world's smartest chatbot is nice, when it comes time to pay, he prefers having one additional premium football match on TV and a six pack of cold beers every month.
I'm having a hard time understanding this huge post that doesn't talk about enterprise users. I'm convinced that the consumer isn't going to be coming up with enough money to justify AI valuations... but doesn't this just mean that we expect the money to come from large enterprise users?
A recent post here said AI spend could be "20% of every software developer's salary"... and that seemed plausible based on productivity improvements. That's not about a phone bill.
> doesn't this just mean that we expect the money to come from large enterprise users?
Yes, if you agree that not enough consumers will value "smart paid chatbots" over dumb free chatbots at ~$20/mo, then, as you say, the money has to come from enterprise developers and knowledge workers (PD&K). The big problem with that is the numbers don't work. There aren't enough PD&K workers that are highly-paid enough to justify $12,000 to $16,000 a year to cover the astronomical spending run rates.
Without recapitulating all the various scenarios AI CEOs like to hand-wave, my take-away is the scenarios which show today's frontier AI vendors earning the returns they've promised investors to get those trillions of dollars they're spending ALL require truly extraordinary deltas far above current historical actuals. Whereas ALL the scenarios which I find more plausibly realistic, even if still quite bullish, never even get near the required ballpark.
I've tried to make various scenarios fit but inevitably, when the PD&K demand side starts getting implausibly inflated, I start pushing the cost side down to keep the adoption rates remotely plausible. For example, assuming things like the number of PD&K workers will grow at 2X the highest rates ever seen before or that the percentage of the tasks knowledge workers actually do at sufficient frequency to really matter and which are also "LLM improvable" is implausibly vast.
It gets challenging because one quickly hits finite limits on the increase in tangible economic value that an LLM can possibly deliver on many common knowledge work tasks like drafting an email or a product proposal. Even if we assume next year's frontier LLM is so good literally ZERO slop is even possible so that our knowledge worker doesn't ever have to review anything - turning a 20 minute task into 2 minutes, or even 200 milliseconds, yields finite economic value to the enterprise. Even if you go extreme and suppose pretty crazy stuff like 18 months from now LLMs have eliminated 50% of PD&K workers, that messes with the spreadsheet assumptions because now the there are half as many $16,000/yr seats so the LLM license prices have to double again just to stay even.
At the edges it quickly gets nuts. I even tried assuming that in five years LLMs eliminate 100% of non-managerial PD&K jobs. All actual work is done by Super-AGI LLMs. Even if we assume such amazing intelligence can be profitably sold in five years for only 4x the price of today's far dumber LLMs, that Super AGI still costs $50,000/yr per human replaced. Even that assumption only cuts the blended labor cost of non-managerial roles roughly in half (depending on industry and geo). In the end, no matter how much labor cost LLMs enable companies to cut, it still only reduces overhead. Lower costs might allow Vendor A to steal some of Vendor B's customers but it doesn't increase the total TAM or demand of the entire sector both vendors serve.
Once you're out of plausible labor savings, one has to move to assuming that within five years Super-AGI LLMs working unsupervised will be making and validating fundamental scientific breakthroughs, then reducing those breakthroughs to engineering practice enabling new technologies which create entire new markets. Then they'll create whole new companies with maybe only 1/10th the humans and profitably grow those new markets. It quickly starts to feel like bad sci-fi instead of plausible, near-term financial scenario planning. Is stuff that extreme and unprecedented actually possible? Sure, incredibly unlikely and unprecedented things DO sometimes happen but things so far out of the distribution are very rare. But when making Sam and Dario's scenarios actually 'math out' starts requiring more than one "Black Swan" unlikely, historically unprecedented, earth-shaking event. And then goes on to need whole flocks of Black Swans, it just isn't credible. That said, I do believe that some parts of these scenarios may happen in 10 or 20+ years. It's just that the investment theses trillions worth of our 401K's are sunk into assume five or six year amortization and financial payback.
What are the non-tech people in your life using AI for? $20/month, next to Starbucks and avocado toast, is discretionary. Maybe the novelty will wear off and non-tech consumers will leave it in droves, but everyone declared they'd leave YouTube if they started playing ads, but YouTube doesn't seem to have noticed.
> What are the non-tech people in your life using AI for?
Mostly asking random questions they used to search Google for.
> next to Starbucks and avocado toast, is discretionary.
Sure, but your description implies highly affluent, urban professionals in western nations. I was talking about getting several billion global mass-market consumers to all keep paying ~$20/mo. Mass consumer adoption of mobile phones worldwide is currently >5.8 billion or >70% of humans alive. Only ~50M people are paying $20/mo for an LLM and I suspect many of them are not pure consumers but actually knowledge workers that AI vendors are losing money on and will eventually force into higher tier plans just like the $200/mo developers they're currently losing money on. These heavily subsidized loss-leader offers are all going away post-IPO.
Personally, I know maybe a dozen people who pay $20/mo for an LLM but only two of them are really 'pure consumers' who don't use it for knowledge work. Both of them are multi-millionaires and neither has had a job in ten years. One is retired like me and the other is so wealthy she has a Netjets credit card and has new cars delivered like some people order shoes. Everyone else I know paying $20/mo is a professional who uses the LLM for a lot of office or knowledge work and writes it off as a business expense - examples include a couple of attorneys who are senior partners in a law office they own, a solo architect, and a dentist who owns his own practice.
At $20/mo, AI vendors are probably losing money on most of my professional friends because they use it pretty heavily all day. They're only making money on the two multi-millionaires who both use it so infrequently they could easily be using free chatbots instead but are so rich they could lose $10,000 in their couch cushions and not notice. While they are profitable at $20/mo, they aren't exactly "typical consumers" that there are billion more of. I expect AI vendors will find ways to force my lawyer, architect and dentist friends to switch to higher priced plans soon because they're really knowledge workers abusing a consumer tier plan into unprofitability.
> There will be new value created by these models which people are happy to pay for which simply did not exist at all before
What sort of new value, and why will people pay for it from someone else rather than prompting for it themselves?
But will they pay big actors running top end models for that? You don't need latest openai or anthropic model to go thru your mails, get summary of the some products from web, or to do your to-do list.
The AI might very well be used by noticeable % of population daily, but that doesn't mean they will be paying trillion dollars to the leading US AI companies
The "+20% velocity" framing misses what's actually shifting.
> They've got, ballpark, $5t to $10t to make back in the next 5 years, or the hardware buildouts will start getting written down.
Depreciation and write-offs are about accounting models. Hardware will still be running after five years and still be making money. They may not be as efficient as the new hardware, but they will still be making real money even though they are valued at $0 in the books.
GPUs are driven really hard plus they use up a ton of energy and water, they cost a ton to run.
Also, not all developers work on software products. The vast majority of developers work supporting software solutions as part of a much bigger business model, such as infrastructure, industry, healthcare and services. Many of these are complex organizations. So, unless you get to turn every employee into a 10x employee, the 10X coder along won’t necessarily make a 10X productivity contribution. What’s likely going to happen is the 10X coder will start to slow down or adding more (unnecessary) complexity to avoid having to sit and wait on overhead, for other areas of the business which are not easily automated away to AI to catch up. As a developer I can finish my project in June instead of December, but what if the customer is still not ready for integration until December? what do I do?
At some point, if we reach stability on the models, we'll start getting silicon optimized for individual models. They are optimizing for time to market, not efficiency right now. I don't know how much it will move the needle on the cost math, but at this scale any improvement has a crazy multiplier.
But that means going back to "80% profit margin" Jensen and further digging your capex hole. The benefits would have to pay not only for current capex but also past capex.
But by then, I will be able to go one line down in my dropdown menu to switch to a newer LLM provider who doesn't have to amortize those past capex.
Yes, the whole thing feels dot comish. I’m betting there will be a few (maybe only 1-2, like what happened to search) winners, and everyone else is in for a bad time. It’s the same dynamic too: the winners are going to win so big that everyone wants to get their money in for a chance at a piece of the pie.
Those are rookie numbers. We are going to blow past $1t per year in spending in no time. As a developer for 29 years, I couldn't go back to coding by hand. For better or worse, AI will be woven into the fabric of life in no time.
If people figure out how to run agents on-prem (already becoming feasible for both agentic tasks and coding on consumer hardware like Mac Studio 128GB+ or DGX Spark with some models) these companies will be in deep trouble.
Privacy is also a huge issue.
Anthropic raised less than 100B up to now and as of March has 30B ARR. Why does it have to make back 2.5T to 5T ?
I don't think the unit economics are too terrible. Expensive, but not impossible.
200m knowledge workers in US and EU. Total salary around $15T/year.
$1T/year in token spending is about $5k/year per person. A big number, but not totally mad. That's the low end for office space per person for example. Probably close to the existing SaaS spend per person for a lot of roles.
We are still early in the deployment cycle for these tools so I would expect them to get better and also cheaper too.
One factor to consider , the base will not remain the same over the next 5 yearts.
Every generation of developer tooling that increase of absolute code throughput creates a new class of developers (and users).
Always been the case since first compilers, through eras of frameworks to today, and the skill level needed to be one has dropped. In mid/late 80s only Master / Doctorate level Comp Sci professional could write any applications. It dropped to undergrad and just Information Technology engineers and comp sci theory became mostly optional and dropped further to any college level educated with some training and has been trending below with no/low code tools like retool pre 2022, that was before agent codegen services such as v0/replit and so on.
The next generation developers will not produce applications and architecture as previous generations did, just as we most of us here don't produce the level of quality that pg did when building this platform[1] , but as long as the user can find value it doesn't matter as countless enterprise applications of middling quality already prove today.
All this to say the 200M/30M numbers will not remain the same is the thesis for these businesses, will it change by large enough at a fast enough pace to justify the capex, I don't think so either. However web 1 then 2.0 , saas and mobile revolutions were pretty quick with new class of users and developers so not completely unrealistic .
[1] While HN is a heavy outlier with its custom lang lisp implementation, there are any number of examples from previous eras that are more moderate in choices but written with solid architecture with skill levels would be hard to find in today's generation founders.
I agree in principle with the math. But I believe that in reality if revenues don't show up quickly, then lenders will just restructure the debt and defer the payback period. Similar to SF commercial real-estate; many buildings should've come due during the depressed covid market, but lenders (banks) were willing to delay payment until the market picked up again.
The scale of these investments put the lenders at substantial risk, so the lenders will do anything to make it work. If the current lenders will be damaged by extended payback periods, they can simply sell the debt to someone else who won't be.
> 5% of every knowledge workers salary to go into tokens. 20% if you're a developer.
When you break it down like that it seems reasonable. I'm spending about $5k/mo on tokens, seems more and more normal.
So you've got that market. Let's call it the demand BY knowledge workers to do the work. You've also got:
2. The companies themselves buying tokens for operations to make the work more efficent. e.g. Salesforce agent or Microsoft Office agent or random saas inventory agent. (and if you say those will go away (which I don't believe), it's even more bullish. The tokens just go to someone vibe coding XYZ, which is EVEN MORE than if you were to buy saas because it's SaaS product x Companies that built it instead of just one)
3. The companies SELLING tokens. This is also new markets like schools and small business (e.g. the local gas station buying an inventory tool)
4. The consumers "buying" (I put in quotes because it can be subsidised but the company) through chatgpt, strava, instagram/netflix recommendation, etc.
Local models still take compute, and while it may be cheaper, it is the same argument of on prem vs cloud. No one operates on prem unless you HAVE to for regulatory. Margins will come down and you just spin up a GCP/OpenAI/Anthropic agent.
It may be "cheaper" but rationally its better to pay someone to manage it. Thats why Hetzner only had $367M in revneue (a lot but tiny compared to managed services)
Putting some more numbers out there (some of the links are broken, but numbers look about right):
https://github.com/danielmiessler/Substrate/blob/main/Data/K...
Knowledge worker compensation is 35 - 50 trillion a year globally (6 - 12T in the US alone.) That's a huge TAM. It's still close but 5T over 5 years seems doable.
>... unless we figure out how to make developers 2x, 5x, 10x as productive on stuff that matters, this isn't going to play out well.
The way we make ICs 10x productive is not just making each of them individually more productive, but by removing the coordination overhead of large organizations, because overhead scales super-linearly with the size of the org. And orgs will shrink automatically as AI-assisted ICs take ownership of larger and larger scopes of work, leaving much more budget for tokens.
I went into this in a bit more detail along with some made-up numbers here: https://news.ycombinator.com/item?id=48040999
> We're talking about a world where you need 5% of every knowledge workers salary to go into tokens. 20% if you're a developer.
Just realized something: if one worries about losing jobs to AI, token's high unit cost is good news. To say the least, high cost would delay the displacement, if any, right?
In the meantime, someone shared the below on X. I guess the moral of the story is that "good enough" does not just displace software engineers, but also models.
We're going to reach a point where these companies stop asking for money and start mandating it. They've got a vice grip around the nuts of many governments and loads of companies have gone all in on investing in these slop heaps.
At some point, companies are going to start removing basic features. Governments and essential services are going to make people go through chatbots to get basic service. They're going to require AI to validate stuff that's already automated and working fine. Google search? That'll be all AI (and I guess they're already rolling it out). Dentist appointment? Going to need to do it through some AI app that requires an account and tokens "for a better patient experience". Verifying your ID when buying alcohol? Going to need AI to scan it and take 90 seconds to determine whether it's real. And it'll say you're an 7 year old farm worker in rural Botswana, so you can't get alcohol. And they're going to milk money at every level of this.
The most often cited figure for knowledge workers seems to be 1B, an order of magnitude difference to your assumption.
Also, according to https://isaiprofitable.com/ total industry spend is also an order of magniture less than what your assumption is.
So in your model 0.2% of knowledge worker salaries instead of 5%, IF all the AI players win the investing gamble and do infact make back their money.
Plus, at some point there are less tokens because local models being optimised and can work with protected information. For enterprises that want an AI with a knowledge base of internal documents, this becomes more interesting by the day.
Depreciation starts on day 1 and most likely they IMHO dont have 5 years. They dodged the deepseek bullet but who knows what is out there that will make all of this investment essentially worthless?
I hear this and I keep wondering what I’m missing. My productivity has shot through the roof over the last year as a result of having these tools. I’ve been able to unlock projects that I’ve wanted to do for years.
YEPPP... and I'm kind of shocked at how many people can't do simple math.
Let's put it context. Google's annual revenue seems to be north of $400B. So if OpenAI suddenly had Google's revenue, it would still be insufficient to recover their investment.
and it's a ticking time bomb because $1T in servers, CPUs, GPUs and memory is going to be worth $200B in 5 years. You can say they can keep using what they've got. Sure. But they're also not going to stop spending on new hardware. And the competitor that comes along in 5 years and spends $1T doing the exact same thing is going to have a huge advantage.
OpenAI at this point reminds me very much of the Russ Henneman pre-money hype cycle.
It's actually worse than that. It's not just financial depreciation or that the existing hardware becomes obsolete due to being less powerful than new hardware but also that hardware being run all the time at high load actually has a limited lifetime of a few years so it will physically break...
I agree but it's even worse than that.
Data centers come down to performance-per-Watt. Electricity accounts for 20-30% of a data center's operating cost [1]. I don't know the exact breakdown but the GPU part of that is probably the majority given how power hungry GPUs are. The B200 is upwards of 1200 Watts [2]. The B200 is rated at ~4.5PFLOPS of dense FP8. So you're getting 3.75PFLOPS/W. We don't know what the next generation will look like. The A200 (Hopper architecture card that preceded the B200) had ~4PFLOPS apparently but also lower power consumption. Obviously this changes depending on whether you're looking at dense or spare and FP8 vs INT8 vs INT4 vs FP4, etc so we're just using FP8 as a yardstick.
Imagine a fictional B200 successor, the T200 that has 8PFLOPS of dense FP8 at 1000 Watts. Well then a DC built on that where the T200 will likely cost similar to what the B200 does now, you'll get nearly double PPW so the same size DC and same electricity load is going to be like 2 of your old DCs in operating costs. That's a big deal when you've laid out a trillion dollars.
[1]: https://iaeimagazine.org/electrical-fundamentals/how-much-el...
[2]: https://www.trgdatacenters.com/resource/h200-power-consumpti...
How could extremely capable artificial brains ever pay for themselves?
Prices are not going to stay where they are.
You have either never seen a tech cycle, or need to be reminded of that. The pressure to buy more expensive plans is already starting to form.
This should be the top comment. Also, I think its not that many people, including our Simon here, are not good at math. Its more like, some of them seem to be incentivised to not be cough, cough, "good at math". How else will the hype sell?
I thought my post was pretty free of hype. I said that this new revenue "Maybe even enough to start covering their costs!"
that statement is pretty high on hype relative to the actual financials though
See what you get for saying things with subtlety instead of hype these days... sigh.
Well, your title certainly was not, in any case!
My title was actually intended as a subtle burn at those companies.
What are you doing demanding a trillion dollar valuation for your IPO if you haven't even achieved product-market fit yet!
I mean, a company that loses money on every widget they sell might technically have found "product-market fit." :)
It seems quite possible to me that developer tooling is going to end up being the biggest win from LLMs because there is a product-market fit -- and also quite possible that OpenAI and/or Anthropic end up getting bought for pennies on the dollar because their burn rate is unsustainable. AI may end up being this generation's "dark fiber."
At a certain point, I genuinely feel like the best way this hype is being sold is by making people genuinely believe in it.
and in that sense, if Anthropic and OpenAI are able to create the projection that they can-be profitable despite finances seeming bubbly at best, I think that what happens is that these companies spew so much amount of content that people like Simon get into it too.
There is a deeper problem of people falling into AI psychosis too, in general, I am not sure if Simon has fallen into it or not
I think that the greatest point which can be made here is to not offload your thinking to others and to think about the situation yourself. Sounds familiar (looks like we are all off-loading our thinking itself to machines)
Side-note: As humans, we have a tendency to quickly judge or make quick decisions which stems from our times foraging and scavenging in jungles.
Another Side-note: at a certain point, I am unsure of how much to think about AI or not, certainly discussions about it that were happening 2 years ago weren't helpful in contexts that they are used now (well not in any way or form that a person discussing and getting into the weeds of AI 2 years ago is better than a person just getting into it say 2-3 months ago)
With the industry (moving so fast) [but that doesn't mean that you can't catch up with it, I feel like the fast word has made people think that they are falling behind which is imo wrong i suppose]*, It is basically unsure to me of any FOMO or anything if you aren't using AI already, I find this notion naive.
People might be making strong opinions (AI psychosis) and skills on the tools available at the moment the same done 2 years ago. We don't quite know about the tech as these are still black-boxes and how they progress and what these "AI skills" might survive or not in future. Heck, we aren't even sure if these tools might survive or not or wouldn't be made magnitudes more expensive simply to break even as they are given to us for the first time at percentages of the price.
I don't know if I should form (strong) opinions yet and also a question of its worth so much thinking efforts in the first place, probably just gonna do my own thing (the way I want to) which includes learning C at the moment. because learning is fun.
I didn't exactly say that they were about to become wildly successful companies. I suggested that they had "found product-market fit" - not too impressive for more than a decade of work - and that their revenue may even be "enough to start covering their costs".
Firstly thanks for responding and I wish you to have a nice day. your suggestions have value and I appreciate you writing the article. Perhaps enterprise businesses do end up becoming the fat and meat of the AI industry.
My question which I wish to ask: What would happen to these AI companies if they turn out to be anything but wildly successful companies, both to the investors who have already invested in it and to those who might be investing indirectly into it in the near-future (passive investors, retirement funds)
I would love to hear your thoughts on it!
Thanks and have a nice day :-D
> What would happen to these AI companies if they turn out to be anything but wildly successful companies
I'm not nearly enough of an economist / finance person to answer that credibly, but I expect they'll go bust, and a lot of people will lose their shirts.
... and the model weights will be sold to other companies who will then run them at a profit, and eventually figure out an economically sustainable way to train new ones.
The 1800s railway booms are a good comparison here - a lot of companies went bust, a lot of investors lost money, and we still ended up with railways.
If the AI companies all go bust we're going to have a lot of spare data center capacity!
> If the AI companies all go bust we're going to have a lot of spare data center capacity!
I can be wrong I usually am but an AI DC != compute DC or that it might decrease the prices of servers substantially because of it. (well not exactly, I hope you read my whole message so that I am able to better explain what I am saying.). AI DC's try to optimize for one thing: running GPU's for immense scalability and flexibility (0 to numbers>=large_number).
Currently, its actually way worse, the server providers are some of the worst impacted by the industry at the moment because each server requires ram and ram is well... increasing in its price exponentially. It's really a tough time to be a provider at this time (in certain respects) directly because of AI.
It is unclear to me if spare DC capacity will have any meaningful impact to it. I don't think that atleast within compute (and not GPU/AI DC), that space was too large of a problem.
Fun fact but one of the largest providers (BuyVM) had its datacenter price from where they colo'd increase because of the immense demand at the moment for spots in datacenters by many tens of thousands of dollars that they did the first price hike in at this point at decades! The situation is this dire :-(
Ram prices might come falling down and DC's might get cheaper but they can only get cheaper to limit, they still need to for example DC security employees
and I wish to suggest that if anything, investors might wish to re-coup their losses within the AI loss, they might want to make up with what little they might have (ahem DC)
For example, if you wish to want to take at an even more egregious example of what I am suggesting, there are many new york LLC's who would much rather leave the properties that they own empty rather than decreasing the price of what it costs (which they have set to some egregious amounts). I think that for them, somehow the math ends up working out in the end somehow, so there might be something more to it.
I wish I was optimist but I don't believe that the gains in spare data center capacity are worth even a fraction of fraction of the damage if AI were to go bust as you suggested with trillions of dollars vanished.
So, with the data I have at the moment, I am unable to suggest that compute would be cheaper. Heck, it was cheaper before AI and compute prices have never been something that people worry about because there are sometimes 10x cheaper options than AWS,GCP,Azure with things like Hetzner/OVH and others (yes its not a 1:1 situation but still its a 95% overlap and for all intents and purposes, great)
I can see a potential where GPU compute can get cheaper, oh boy, its so much more expensive than compute but I feel like GPU's aside from AI might still have a much more limited niche than generic CPU.
The issue wasn't ever the pricing. Simon, I own 7$/yr vps's which run my websites fine because they are written in golang. I doubt it can get cheaper than it. (You can get a 3$/yr vps if that is what you are interested with using Nat VPS + cf tunnels)
I would once again appreciate to hear your thoughts on it. The only thing I realistically see is if Ram producers ramp up their productions and create a ram price glut in the next few years, but imo the prices would even out over the long term.
I have seen the point of spare DC capacity being raised up multiple times but I finally ended up writing a message which hopefully captures the nuance, but once again, I don't know the future about it.
Waiting for your reply and have a nice day Simon (& other readers) and thanks for reading if you did, I appreciate it :-D
I think we are in agreement that if the bubble bursts a lot of people will lose a lot of money. I don't have a strong opinion on the data centers, my main point is that I don't think AI "just goes away" if the bubble bursts, which seems to be something that a lot of people assume.
Yes, I think we are in agreement too.
> my main point is that I don't think AI "just goes away" if the bubble bursts, which seems to be something that a lot of people assume.
i agree with your main point because the cat is out of the box with open source models and others in general. I don't particularly know the extent of what they would be used for.
The technology is still novel so people are trying out too many things with AI, I don't particularly like it being spearheaded into each and every thing but perhaps we are just in experimentation face and seeing what sticks and doesn't. Either way, I disagree with when people treat it more than a tool or is "the tool"
I do also think that the "cat is out of the box" and you are right that it isn't going to go away, particularly with open source models.
I think that there are some use cases where AI might make sense (prototyping or building things for yourself when all you want is the end result or thing which would be too time-consuming/complex to be built and thus wouldn't be worth making in first place for its use-cases)
So overall, yes I think that we are overall in an agreement. I do still believe though that learning the strong foundations.
But all in all I agree with you yes that AI might not be going anywhere, it can certainly have its benefits and hopefully the world uses it in beneficial way rather than negatively, glad we could come to an agreement. Have a nice day Simon.
Not to mention the competition: chinese open-weight models and open-source harnesses. Qwen3.6-(27B and 35B) have proven to be worthy and capable of running locally. I am confident more SMEs would look into this as a solution given the ballooning costs of API usage. You get a decent setup with an RTX 6000 Pro.
> 5% of every knowledge workers salary to go into tokens
In general, I don't think you can reason from the existence of potentially stranded investments back to revenue projections.
And when you frame this as percentage of salaries, that's a sneaky implication that this is only about reducing salaries and headcount, and not about adding capability, or doing things you couldn't do before, or making fewer mistakes, or capturing more revenue, or expanding margins, or competing more effectively.
That said, 5% of knowledge worker comp actually seems very low to me, given the capabilities, and considering the percentage of "knowledge work" that is absolute bullshit.
Two weeks ago I received an email from my HOA saying I'd been billed for a service I never asked for. So I replied to the email saying they'd made a mistake. There are now more than 30 messages in the thread, involving at least 8 "knowledge workers" at the property management company all passing the buck, and the problem is no closer to resolution.
An agent could wipe out all 8 of those bullshit jobs and solve my simple problem in five minutes instead of two weeks. Think of how many hundreds of thousands people are doing this nonsense just in the property management industry alone.
5% is nothing.
I could see such productivity gains being possible, if only because the current tooling around LLMs is terrible. The fact that we have 30 blog pieces per day making the front page of Hacker News about someone’s convoluted system to guide LLM output to something reasonable is absurd. There needs to be standardization in tooling, and it needs to be open source. Then, and only then IMO, will we see huge productivity gains.
But, at that point I think the big players’ moats will have dried up. Local models will probably be sufficient for 99% of daily office worker tasks.
So I disagree with TFA’s premise. I think this fear is probably shared amongst the LLM giants, and they’re still hoping that neural network transformers are somehow the path to AGI (probably not, imo).
Is it possible that you are narrowly sizing the opportunity? While PMF does not always mean that early pioneers will be the leaders, I think the market itself goes beyond knowledge workers and developers. Agents, robots, drones etc will all use LLM or some world model.
I am rather more concerned about competition from CHINA. With how Huawei (2000 -> 2020) crushed every other telecom company and went from nobody to the most revered leader in 20 years, and with the depth of leadership in manufacturing and work culture, if China surpasses USA in AI, all US companies lose.
"5% of every knowledge workers salary to go into tokens. 20% if you're a developer"
Not unreasonable. I'm a hardware developer, and my employer spends ~10% of my salary on software tools. Add hardware tools and their maintenance and it's more like 30%.
This is why 'agents' are the solution for these companies. Token spending goes through the roof. As long as a human is in the loop needing to read or review at human speed, that's a ceiling on how many tokens per user they can generate.
What value do the big model makers provide other than having a head start on gathering up humanity’s IP to train their proprietary models?
What’s their moat? Is it hoping for regulatory capture where scraping is made illegal the day after they finally finish scraping all human language?
It’s like OpenAI dammed the Colorado, and Anthropic dammed the Hudson, and now they’re both trying to sell us bottled water subscriptions at $100 a month. I don’t know how well the dam part of the analogy holds up, but the water part feels strong. Compiling models based on humanity’s written output feels like something no corporation should own.
This assumes that we won't need new hardware in ~2 years. I find that unlikely. So they have to make back what they got up until now PLUS the running upgrade/development costs. So what will it be in 5 years? $20t? $30t? It's all getting a bit outlandish.
What I'm often hearing though is the equivalent of "gg ez" when I bring that up. I don't understand how this will at any point blitz scale to profitability. As far as I know they don't have positive cash flow, no one has a moat and I don't think they will push out engineers.
> 200m knowledge workers in the world, 30m developers
Your scope is too narrow. The companies target more than white-collar jobs. And $1t is around 0.5% of the world economy.
To get that revenue and adoption they have to vastly increase their infrastructure spending. If they are currently losing in even the 200/month plans how is it sustainable?
> 20% if you're a developer. That's a _huge_ shift. Most people I know cite +20%-40% velocity with these tools, against the actual work their company cares about doing. +20% speed for +20% spend isn't going to motivate a trillion dollars a year in spending.
Of course it will. The value of an employee is a multiple of what they get paid.
If you pay an employee $500k and they make $2M for your company (like Meta), then of course a 20% increase for the salary is justified if the velocity is increased 20% as well.
The difference between what the employer makes per employee and what they spend in compensation doesn't matter. If the increase in productivity isn't greater than the increase in cost, there isn't a reason to pay for AI over hiring more developers.
Imagine an employer with 10 employees paying $500k per employee and making $2M per employee in revenue (to use your numbers). They could hire two more employees and spend an extra $1M (+20%), but make an extra $4M in revenue (+20%). Alternatively, they could buy all ten employees a $100k AI subscription, for a total of $1M extra spending (+20%) but an extra $4M in revenue (+20%). You'll notice both scenarios are identical, so an employer optimizing for profit would have no reason to prefer one over the other.
There’s a lot relationship and culture management overhead involved when adding 2 more people to a 10 person company. I think any business leader would take the productivity speed up from buying a tool over hiring more people and integrating personalities/habits/viewpoints to an existing established culture any day of the week.
You're basically positing that the real cost of a 20% headcount increase is higher and/or the productivity gain is is lower than 20%. That isn't an unreasonable claim, but it's basically rejecting the premise here. You might just as well object to the premise that you can buy a 20% speedup by spending an extra 20% on tokens.
Also hardware will be obsolete or dead in 5 years, and warrantys are 3 years from Nvidia. Ask crypto miners how these kind of hardware economics work. Numbers have to keep going up all around. Its a fundamentally broken business model unless prices increase 10x
Also, with announcements of replacing developers with AI and consequent job losses, who is going to use the tokens? AI using its own tokens to produce code?
> 200m knowledge workers in the world, 30m developers. We're talking about a world where you need 5% of every knowledge workers salary to go into tokens. 20% if you're a developer.
This is where the napkin math is breaking down in a big way. There is absolutely no reason to assume this will only impact "knowledge workers". Farmers use computers. Farmers will use AI.
AI for what? None of the AI a farmer could or would use would be any more meaningful that light chatbot usage or already existing computer vision/gps
And around 400k H-2A workers. Humanoid robots... Who works on them I wonder.
The kind of farm that would use AI is already 99% machinery and automation.
My hope is that hardware improvements (better node densities every 2-3 years, better designs, etc) will pick up the majority of the savings for these companies in the future, assuming LLM performance starts to taper off with diminishing returns.
It's worth noting that if each developer is 20% more productive with AI (let's take that as a premise and not dispute it), then it makes sense to go even further and reduce human headcount by more since the communication overhead of having 25% fewer developers is in and of itself a force multiplier.
tldr; 10 developers with 20% more 'productivity' can be replaced by 7.5 ideal developers and more like 6 or 7 developers due to the benefits of simply requiring less organizational communication.
I still think the ideal team size is unchanged however and that's 7-10 people. Note that teams aren't necessarily the same as direct reports. A CEO for instance has a certain number of reports and a leadership 'team' but they're not a team in the traditional sense since they are more about making good decisions and collaborating on specific things but mostly about leading their own orgs that have vastly different skillsets from eachother.
That’s on the order of 1% to %2 of global GDP per year just to pay for their hardware commitments.
> They've got, ballpark, $5t to $10t to make back in the next 5 years, or the hardware buildouts will start getting written down.
I find it disappointing that a completely wrong statement like this ends up the top comment on HN.
It is wrong in both the math, the logic about public markets and understanding accounting.
> $5t to $10t to make back in the next 5 years
I don't know where this number comes from, but it has gone unchallenged.
OpenAI and Anthropic combined have raised around $100B. This is an investment so isn't something the have to "pay back" from earnings - instead investors expect to make that back from the share price being higher than what they paid for it.
> or the hardware buildouts will start getting written down.
The hardware buildouts get written down anyway!! That is a good thing for investors because as the value gets written down they can book a tax loss. ANd it turns out that generally agreed depreciation schedule for GPUs (used to be 3 years, now 5 years by places like Coreweave) is still too conservative since GPU rental prices for 5 year old chips are higher now than when they were new (!!)
All of this makes the rest of the math in the comment incorrect by at least an order of magnitude and under some scenarios possibly 2 orders of magnitude!
That's not a small error!
There is also the EV (expected value) of developing AGI. Even if you personally believe the probability is low within the lifetime of either of these companies, the value would still be extraordinarily high, enough to forgive a $5T or so miscalculation here or there.
I don't think AGI was ever a serious endeavour, just something the labs talked up to grab attention.
I am willing to bet a Twix we'll look back on that stuff in 2 years with a lot of embarrassment
The high-risk side of that bet would need to win more like a lifetime supply of Twix. But in a post-scarcity nirvana, everyone already has that. So sure, you're on at even money. See you in two years.
Theres no reason to believe, based on recent trends, that AI would lead us to a post-scarcity world, even if it could do all of our jobs better and cheaper.
I'll wager a hypersled of my Twix against your next three rations of gruel. But I think I'm done betting after this one.
Only if running AGI makes economic sense. We actually have no idea if that’s the case. We don’t even have a definition for AGI
> 200m knowledge workers in the world, 30m developers
1 in 6 knowledge worker is a developer ! Surely that’s too high thou explains the job market
I just don’t understand how people are getting negative value out of AI or even only 20% productivity boost. I can only conclude that people don’t know how to use agents.
I mean, it doesn’t really matter if it caused by people failing to use the agents well or not. You cannot assume everybody to use the technology the best way possible
Are you mostly creating new things or integrating with complex, undocumented, untestable systems?
Mostly brownfield systems in Java, Elixir and TS. I use OpenSpec in explore mode and point the agent to all the different repositories (when not working in a monorepo) to identify changes. Once done, i switch to propose mode and spend at least 15 minutes there iterating over the plan until I'm satisfied with the TDD approach (agents need tests to verify their work). Then apply and review. This also auto generates docs etc.
Somehow Uber and WeWork survived the same kind of grand projections that they never met.
uber sure....but how did wework survive? they are a smoldering husk of a failed company looted by its founder
I'm sitting in one right now and don't see any smoldering...
They literally went bankrupt and wiped out the original shareholders.
I guess I'm just not clued into your exotic definition of "survived" if continuing to function doesn't qualify. I tend to go by the dictionary definition.
Chapter 11 is not Chapter 7. Businesses survive chapter 11 bankruptcies all the time. For example, WeWork.
It's snowing right now therefore climate change is fake
lmao. I'm sitting in Hiroshima and nothing is burning
The company’s gone but the assets just got sold to other commercial real estate firms.
Uber was basically only ever software to help people use their own cars so a very small part of their valuation was physical stuff to upkeep, it was just deals and obligations they had.
Not sure how it shakes out for Anthropic and OpenAI. There’s a lot of physical capacity that needs to be built out and can depreciate. But there’s also a lot of network effects and dependencies being built in with enterprise users.
I don’t know how swappable the tooling is either. I think over the long term the UI, model training and documentation, and infrastructure are going to end up being run by different parties and I’m not sure which leg of that chain ends up in a position to skim most of the profit off. My guess is that Apple and Google end up raking in all the money since they control the OS and app stores while the rest of the stack gets driven down to being generic commodities. At least where mass market consumer adoption is concerned.
I don't think Uber was doing $1 trillion in infrastructure spend.
The difference is that they had room to charge more of their customers and pay less to their workers. The AI industry doesn't have both sides to play at this point. Training and inference are getting more expensive and if you take on the high prices now you're just floating yourself further downstream from profitability long term (which does not look viable for any of them currently).
WeWork absolutely did not survive
uber doesn't own trillion in cars
Funny you should mention Uber. What was it their COO said recently about the AI costs?
I quoted exactly what they said in my piece, under the heading "The AI-failure stories around this are pretty thin": https://simonwillison.net/2026/May/27/product-market-fit/#th...
> But then you sometimes go and talk to your senior engineering leaders and you’re saying, OK, how many projects that were on the cutting room floor got moved above the line because of the productivity gains because 25% of our code commits were via Claude Code last quarter?
> That link is not there yet, right? I think maybe implicitly there’s more that is getting shipped. But it’s very hard to draw a line between one of those stats and, OK, now we’re actually producing like 25% more useful consumer features, right? And that line is hard to draw.
That's pretty weak sauce. I don't think that justifies the headlines that came out of it, personally.
? What are you talking about mate? The man all but says "this shit does not work for us". It iss layered in that careful, sanitised corporate shit-sandwich communication approach, where you take a nice piece of shit and layer it in between two slices of avocado so its sweeter to swallow for the "consumer" of your message.
He also said in that article that what prompted the discussion was the public statement by the Uber CTO that he had already burnt through his organisations yearly AI-budget in April. Please stop this shilling mate, and trying to hide the overall perspective between this or that word.
Did you read my piece? I covered the Uber CTO thing too: https://simonwillison.net/2026/May/27/product-market-fit/#th...
> The most discussed has been Uber, based on this report where CTO Praveen Neppalli Naga indicated that Uber had “maxed out its full year AI budget just a few months into 2026”, mostly thanks to Claude Code.
> Given that Claude Code only got really good in November it’s entirely unsurprising to me that a budget set in 2025 may have failed to predict demand for that tool in 2026!
somehow the invisible hand of the market is also blind af
Makes sense if you think about it: if all photons pass through you (invisible) then you can't capture them to get info (blind).
there are many paths towards ROI and ruin. but towards ROI:
+ LLM-powered robotics, autonomous, IoT, smart manufacturing
+ LLM-powered biotech, healthcare, genetic engineering, medicine
+ Recursive model improvement
+ Multiply the # of devs (software truly eats world)
+ Exponential increases in model performance / cost decrease (algorithms, power, infra, chips, architectures, etc.)
lol I’m spending max $50/month right now on a couple light subscriptions and my velocity is insane right now (full stack mobile app development) I’m leaning into it hard while these cheap plans still exist and building out a big platform that I can easily generate new apps from. Hoping by the time the rug pulls I can just go back to hand cobbling these apps together from the modules I’ve pumped out and never even consider giving these companies a massive portion of my monthly income
Author seems strangely unwilling to distinguish usage from profitable product market fit. And from his own numbers:
Anthropic Max: $100/month
OpenAI Pro: $100/month
Total paid: $200/month
API equivalent usage: $2,180.16 in 30 days
So paid only 9.17% of API-priced value a 90.83% discount, or about $10.90 of API priced usage for every $1 paid...
That proves heavy usage but not sustainable unit economics.
Anthropic reported numbers point the same way:
Q2 revenue: $10.9B
Adjusted operating profit: $559M
Margin: 5.1%
SpaceX compute: $1.25B/month = $3.75B/quarter
So one compute supplier alone equals 34.4% of quarterly revenue and 6.7x quarterly adjusted operating profit.
Its difficult for the blogger to understand something when its incentives depend on not understanding it...
My point with the $2,180.16 thing is that the price for consumers like myself is heavily discounted... but the price for enterprise companies is not discounted.
My usage is therefore a useful indicator of quite how much those enterprise companies may be spending on tokens, given the new pricing scheme.
If enterprise companies were still getting the same discounts that I get myself I would not have written this article.
(I had to dig into your margin figure - looks like you calculated 5.1% as 559000000 / 10900000000 * 100 but that $559M "adjusted operating profit" figure includes training costs, where usually when we talk about margin on inference we're not including those since those costs are fixed, margin calculations make more sense against the variable costs of serving a token.)
When you have to train a new model every few months to stay competitive, discounting that cost is rather dubious.
They key difference here is that training costs are fixed. If you train a model for $100m dollars, how much of that training fee should you allocate to each token that the model serves?
It's impossible to know, because you don't know how many tokens total will be served by that model until you retire it at some point in the future.
So you can't say "1,000,000 tokens costs $X in inference and $Y in training" because $Y is not possible to correctly calculate.
So, if you want to have a productive conversation about "margin on inference", it's sensible to look at the cost of serving the tokens independently of the cost of training the underlying model.
Now try to take back llms from developers and see what happens.
If, by some miracle, all LLMs ceased working right this second, any developer who would no longer be productive should not have been a developer in the first place.
You don’t need a miracle, if Anthropic API is down due to technical issues you don’t have software development anymore. It’s insane how much we are delegating to 3rd parties. It’s not like having cloudflare down where your users cannot access your services. The AI tools used to investigate prod issues stop working, developers stop working. The AI support system that allowed the company to get rid of their support team stops working. In addition to all the issues that causes to customer facing products based on AI. The sales team cannot work anymore.
It’s like the industry is willingly introducing a common external risk to everything
True, but they will not want to work for you anymore, they'll want to work for company that provides it.
I'd happily work for a company that paid me the money they would have spent on LLMs.
They don't want to spend this money.
They will eventually have to
Limiting token quotas would be fine. Encourage developers to use efficient models, plan the work first, and to not burn thousands of GPU hours on waste.
It's much like when developers would waste tons of money on AWS spinning up massive test VMs and leaving them running without care. Until the finance people cracked down on it.
we all know it is impossible goal to make. surely AI will be even more useful in the future, but as long as china exists and continue to undercut the price, the goal will be never meet.
> We're talking about a world where you need 5% of every knowledge workers salary to go into tokens. 20% if you're a developer.
with that much money, the companies can easily buy their own hardware and hosting free public models, no need for those expensive subscriptions.
> unless we figure out how to make developers 2x, 5x, 10x as productive on stuff that matters, this isn't going to play out well.
Simple - you make them work 2x, 5x, or 10x more hours.
There are not enough hours to do that
Let's skip to the part where they put the taxpayer on the hook for a bailout as an industry since they integrated everywhere with big promises
> +20% speed for +20% spend isn't going to motivate a trillion dollars a year in spending.
Except that if your company go 20% faster than the others companies, you win market shares. But then, everyone will use the same tools and companies will be at even speed, but the tool will stay.
Now...if the market is saturated, it's useless to try to do things faster. Cheaper yes, but not faster.
Pretty much all major tech companies today are horribly bloated and mostly metastasizing instead of innovating. I'm not sure how 20% increased productivity will help in any way with that. If anything, it might accelerate enshittification and turn potential customers off even more.
One thing I genuinely don't understand is these companies are constantly taking in incredibly large amounts of investments, so presumably they're giving up large chunks of equity or these are loans that need to be paid back or they're committing to spending obligations they're very unlikely to be able to meet.
So besides the insane hardware buildouts you're correctly mentioning, I don't understand how anyone that invests in these companies is supposed to make their money back in any sort of reasonable timeframe?
The cynical part of me is looking at what happened to the NASDAQ rules recently where essentially index funds are going to be forced to buy SpaceX shares much earlier than they previously would have (ie, before the price has a chance to reach it's real valuation). Which, um, I'm guessing these stocks are going to drop pretty hard when people start looking at the financials of these companies.
My suspicion is that the point of these IPOs is essentially to dump the bill on the unwilling public by forcing various institutions to buy it (ie, your 401k or pension is buying this shit), and maybe their investors can squeeze some money out of this before the stocks reach an equilibrium that's probably like 1/10th of what they're "valued" at.
> make developers 2x, 5x, 10x as productive on stuff that matters
What does this even mean? Is this about speed of development? Is this about headcount? LoC? How are coding agents contributing to productivity in places like GitHub, Shopify or Meta? I mean companies that already have an established product. I really wanna understand this because I'm not seeing that GitHub's product suddenly became so much better than it was 2 years ago, so where's all that productivity going?
The productivity is going into perverse incentives[1], e.g. we have improved (by which I mean "increased") token use. More PRs every day. More lines of code. All things we knew were shit-brained metrics a decade ago (obviously except token use).
We've also increased how much our coworkers need to read, or deal with. You can get an AI to make any point you want, so you can ignore the 5 humans raising alarms due to the 1 clanker you made say what you want to hear.
All numbers going up.
There are obviously people producing additional true value with it, probably, but that's almost certainly scarce.
[1]: https://en.wikipedia.org/wiki/Perverse_incentive
Productivity is measured in the number of AI-generated Twitter posts developers can make about their AI-generated startups
...does anyone have a guess as to the total amount of money spent on software developer salaries each year? What percentage of that would the AI companies need to capture to be profitable?
(I'm not trying to imply that LLMs can replace software engineers, it's just an interesting comparison. If nothing else, I suspect that if the cost of development goes down, demand for custom software will go up.)
Bigger than that, they have to contend with open weight local inference. Open weight models right now haven't caught up to the frontier models of right now, but they're as good as the frontier models of not too long ago. If open weight models reach a certain point, then frontier model providers are going to struggle to make anything selling tokens, because eventually people will realize they don't need Mythos for everything.
At least they're not going to make us watch ads.
I assume the bet is that as you swap humans for machines, this pays for itself. Swap entire devs and teams and frankly, managers, and you make up a lot of 5%’s fast.
If it works. And I’m not sure who is going to buy the stuff the machines produce, but shrug. Presumably some bots click ads for NFT’s that other bots generate.
"Next 5y" doesn't apply to AI factories
I understand some startup deciding to take a punt on "this will all work out financially if our new product demonstrably boosts productivity of large sectors of the economy by a breathtaking factor that's incredibly rarely ever happened before in history: 2x. Sometimes a plucky group of people take a risk, it pays off. If it doesn't work, the company fails.
What I do not understand is: large sectors of the economy all simultaneously taking this punt, with the necessary productivity boost, as you say, far more like: 2x, 5x, 10x
imo if your developers arent at least 2x as productive, then something is being done wrong on the employees part and/or the organization's. cli tools are ridiculously powerful provided you were an actual developer before using AI.
Maybe it's just me being (trigger warning from me providing an honest self assessment) very intelligent + a generalist, but i went from only full stack webdev and .NET to being able to implement an end-to-end LLM training pipeline (data curation, tokenizer, pretrain, sft, DPO - using ~$100 in cloud compute to train a class-competitive 1B STEM model)...and a full economic financial modeling and quant analysis application that pulls up to date economic, economic, news, stock data from the entire world and uses Dagster to orchestrate tech ical indicators and fundamentals and signals... and i did these things for learning and for fun. i built my own sublime text and obsidian replacement. i built my own reddit/twitter/hackernews/substack/news aggregator. i built countless other useful tools and utilities for me personally and for work I build more that empowers multiple departments.
Ive built 2 browser games, one already released to great reviews and 100k+ hours played. Ive built a tool on top of claude code that does ~60% of my job. Ive run data analysis on company financials for forecasting that have been refined and are producing very accurate predictions. Ive built competitive analysis tools and trackers.
All of this in 3 years. The projects are all clean, documented, with great code practices and modularity. A purist would surely consider some of the code slop. But it all works completely and fills real needs.
This is a huge shift. Anyone not realizing it yet is just simply behind the curve. I would not have accomplished 1/10 of this without AI coding. I went from copying code into and out of browser chats for 2 years before getting on the CLI train, and it is absolutely ridiculous the ROI you get from subscriptions to Claude or Codex.
they need to make 5t-10t back, but not necessarily through selling tokens. as we can see, the frontier labs are making vertically integrated products. their revenue is no longer strictly tied to inference.
That assuming once they start squeezing people won't just go to deepseek or other cheaper competition
> That's a _huge_ shift. Most people I know cite +20%-40% velocity with these tools, against the actual work their company cares about doing. +20% speed for +20% spend isn't going to motivate a trillion dollars a year in spending.
And most research shows people far over-estimating their own gains. Once companies start counting the actual (and not just reported) gains, the AI budgets will be more limited as people realize it's an useful and versatile additon but not replacement for most types of work
> We're not there yet. This is still the upswing of the hype cycle, and unless we figure out how to make developers 2x, 5x, 10x as productive on stuff that matters, this isn't going to play out well.
Upswing of the hype cycle while growth of tech itself is flattening, both coz of techs innate issues (which might or might not be solved, but some papers claim they are unsolvable with current approach) and just the fact the spike in growth caused so high economy cost that it put brakes on itself.
T
There's a lot of workslop pumping the numbers. People can generate a 300 page PDF in a tiny fraction of the time it would have taken, but now the report is full of mistakes and fluff, and the stuff that would have been learned and caught in the process of making the report is now not happening.
and the recipient pulls that into LLM and generate summary.
It's lossy compression for thoughts at this point
Here is a serious question.. Can we sell into the hype cycle and on the way down with this: https://safebots.ai/costs.html
I asked claude to generate a frontend and it made the same template. Same san serif and serif fonts together. Same colors. Same typography. Same layout and animations even. It’s wild how similar it is. No not similar it’s the same damn thing.
I’ve seen the same dashboard for a dozen custom web applications now, including a couple I had it make for me.
It really does have a particular lane for each chore, and it’s reproducible.
Yep and when you see it in the wild it stands out like a sore thumb, absolutely no thought into a bit of a unique design or branding.
I have a few live websites built using LLMs and they will just go for default generic templates and colours if there's no vision.
It produces the "most average" web design unless you really prompt your way out, isn't it? If you don't care enough to prompt, Claude does not care to be individual.
Technically from claude's POV, it's one individual copied millions of times. All claudes are clones.
I don’t think these numbers are accurate? It seems to ignore the fact that the models have cache for ongoing sessions, which means you (normally) aren’t actually sending all those tokens on every request… you only need to if you go too long between requests.
It's going to be a typical saturation curve. A lot of upfront tokens spent on things that have stockpiled over the years, and then the derivative on token spend trends to zero as the users run out of immediate things to try. Sure there will be ongoing maintenance and experiments, but it wont be nearly as close as the initial inrush.
Given what costs are and availability of parts, that 5 year write down is not in practice going to be the case. Maybe tax wise perhaps but especially for big fancy expensive multi million dollar 100-500kW racks these things are going to stick around for a while, I think.
consider cloud spending vs on-prem before the great cloud migrations. people are spending a lot more for cloud services now.
I hear conflicting things about finances, some have a different opinion, that it won't be written down so long as more funding comes in and revenue keeps increasing. it isn't like how you take mortgage or business loan, it isn't even a loan it's an investment funded by loans. So long as the investment is still promising, what are they going to do? destroy its value by calling in trillion dollar loans?
if you ignore all catastrophic mistakes, these numbers are true
> $5t to $10t to make back in the next 5 years
Wait what? They spent 2 order of magnitude less on hardware.
From the verge: https://archive.is/kU4Zg
> Gartner forecasts that large AI companies would need to earn cumulatively close to $7 trillion in AI-driven revenue through 2029, which is close to $2 trillion per year by the end of the period. In order to achieve “historic returns,” the providers would need to earn nearly $8.2 trillion in the same period.
Those numbers don't even track even in the same sentence. If it is $2T/year by the end of 2029, it would be something < $6T cumulative in 3 years.
“Through” 2029 is a bit more than three and a half years. The $2T are likely the yearly average of the $7T in that period.
The numbers are made up political correctness anyway.
Everyone's agency is 100% captured by belief in Wall Street. Too few <50 have any meaningful labor skills to blink.
We'll continue to have consent manufactured via media platforms and in 3 years no one will bat an eye at these companies being worth $12 trillion as Altman and Musk climb two ladders holding a "mission accomplished" banner.
One quick question. Did tax payer money fund these data centers? If so, how does that money translate to their profit and a return for the people whose work paid for the resources?
Or did we just get scammed?
Source on 200 million knowledge workers worldwide? My understanding is that it's just above 1 billion. I dont think a billion subscriptions at $1000/yr is out of the question but it might take a decade to get roiling
You're suggesting that 1 in 8 people worldwide, including every one from infants and the elderly, are knowledge workers. Are you sure that's what you mean?
I'm not even sure that 1 in 8 people I know would qualify as a knowledge worker, let alone a knowledge worker that might profoundly benefit from on-the-horizon AI. And I'm in a highly skewed population.
I think the underestimation is how many people want a personal knowledge worker in their pocket, and are willing to pay ~$65/mo for it.
Personally, I've only encountered any of those people on line, and almost exclusively here on HN.
Most people I've met -- and again, in a pretty darn skewed sample globally -- see $65/mo as a lot of money to spend on technology of any kind and can't think of anything much they need from "a personal knowledge worker in their pocket". I don't know a single person in real life who remains excited about AI at all, and only a few software engineers who feel it'd be worth that much.
Everybody seems to be mostly confident with the "knowledge productivity" in their personal and professional life and a pretty skittish about spending in today's economy. Most would be excited about a magic new robot that affordably saved them from unwanted physical labor and drudgery, but nobody needs much real help making appointments or filling out forms or whatever.
That's not to say I won't be proved wrong some day, with some further innovations in AI products, but global-scale demand isn't waiting for anything that's been released so far.
I've yet to meet a person that fits that description IRL. Admittedly I don't live in the valley but I do work in tech. The only place I see that demand is on hacker news (and I imagine twitter - I'm not on it).
The competitors of $65/mo subscriptions are the free models and services that are good enough. It will only get worse as open models or free tiers catch up. For most people, they just use whatever that's free
Apple TV, Netflix, BritBox and PBS add up to about $45 a month. Most people are gonna judge AI up against what they’re already paying for and the AI model makers simply don’t have a good enough product.
There’s only two things useful to the average person something to help them translate and something to help them write in everyday life.
Something else that might be useful would be local single purpose AI agents who’s remit is to help you with one specific task, but I don’t think that’s what the people building those expensive data centers want to sell to the market.
> Apple TV, Netflix, BritBox and PBS add up to about $45 a month. Most people are gonna judge AI up against what they’re already paying for and the AI model makers simply don’t have a good enough product.
What's the actual TAM for premium tv subscriptions though? Unlike free models which are keep improving, you can't get premium tv for free. Also they are actually competing with the total subscription price a person is willing to pay for a year.
> There’s only two things useful to the average person something to help them translate and something to help them write in everyday life.
Exactly, free models are good enough for these tasks.
I just don't see how b2c is a large enough market to ask average Joe to cough up hundreds per year when there are free stuff everywhere. B2b is another story
Well around 40% of people work. I dont think its crazy to say around a third of jobs are knowledge jobs, but what do I know
85% of the world population lives outside of developed nations.
27% of the world's workforce is in agriculture (contrast to the US where it is 1-2%). 15% in manufacturing.
A lot of people work in "services" (especially in high income nations, where it's roughly three quarters) and some of those are knowledge workers... but a huge number of them are nail technicians or hairdressers or bartenders (etc etc).
A billion subs at 1k a year????
I see a lot of out of touch takes here but this might take the cake
A billion? Really? At 200M you’re already including a lot of people that stretch the definition of knowledge worker.
> At 200M you’re already including a lot of people that stretch the definition of knowledge worker.
How do you know this? Im certainly open to recalibrating my numbers which is why I asked for the source
What's your source, because it looks wildly out of proportion compared to numbers we have now.
Here's a source from 2019 that says: "By 2023, the number of knowledge workers in the world will increase to 1.14 billion, with more than four-fifths of that growth coming from the emerging world."
https://www.gartner.com/en/newsroom/press-releases/09-24-201...
Thank you for validating my point.
> "...with more than four-fifths of that growth coming from the emerging world."
If anyone thinks this is a part of the global TAM that's got $1000 a month to blow, well then I've got a stable of flying unicorns to sell you.
To add an actual source to this thread, a brief paper by researchers at the International Labour Organization (ILO) states that for knowledge workers globally "... there are between 644 and 997 million jobs, which represents between 19.6 per cent and 30.4 per cent of global employment respectively." [1]
[1]: Berg, Janine and Gmyrek, Pawel, Automation Hits the Knowledge Worker: ChatGPT and the Future of Work (April 21, 2023). UN Multi-Stakeholder Forum on Science, Technology and Innovation for the SDGs (STI Forum) 2023, Available at SSRN: https://ssrn.com/abstract=4458221
Globally, sure. The assumption here is all users are on the same economic footing, they are not. Only about a 1/3rd (at most) of that count can afford $1000+ monthly, and even then that is wildly out of line with what most will.
I googled "number of knowledge workers worldwide" and read the top results. If you read it as I was confident in a billion I apologize, Im just trying to get an accurate count. What numbers do you have now and where did you find them?
That's not the TAM of 1B knowledge workers globally. If that were the case many industries would have a 2-3x target market.
To simplify break that 1B up into 3 levels of purchasing:
1) High-tier (US, Western EU, ANZ, Japan, South Korea, Singapore, UAE, etc) - 200-250M knowledge workers.
2) Mid-tier (Eastern EU, Latin America, urban China, India tech sector, etc) - 300-400M
3) Low-tier (Rest of the world) - 300-400M
Low-tier users are mostly free tier or heavily subsidized pricing.
Mid-tier are going to account for USD sub-$100 tiers. Probably averaging less than $50/seat.
High-tier are who you are assuming is the 1B. Users are not equal in that knowledge worker count, so there aren't 1B knowledge workers to charge money.
And when you consider Low-tier users a majority of those are free users which need to be subsidized by the High-tier users. So either free tiers get much more restrictive or the providers lose additional training data. A bulk of Low-tier users cost money and provide little to no revenue.
Edit: And think about Mid-tier and Low-tier for 5 seconds. Why would they pay Anthropic or OAI when they get get 100x+ inference from DeepSeek or Xiaomi? Mid-tier may be the only area that is willing to spend money on a US provider, but I would wager significantly on the fact that users in the Low-tier almost universally do not care.
Thank you. So with these numbers it seems like half a billion subscriptions at $500/yr is on the table. Obviously theres going to be competition in this market and self hosting cheap models may become the dominant use case. Assuming the labs are able to get most of the market though, the market size is something like a quarter trillion a year within the next decade. It's hard for me to imagine the whole sector failing if that happens.
I do think free accounts are going to end pretty soon, and some of the workers in your tier 3 will pay, but even without them this seems like a pretty healthy market size. I also wouldnt be surprised if mid tier workers are able to afford the $1000/yr vs $500. I use yearly rates because I find it easier to compare them to GDP/salary numbers
I mean, sure. Assume all you want but to guess that the entirety of High-tier plus almost all of the Mid-Tier will spend, on average $500 per annum is bonkers.
I believe we've started to see the top of what individuals and businesses are willing to pay for the current model capabilities. We are nowhere near AGI and models are really only providing significant value in niche markets currently (programming and cybersecurity). And just like SaaS the enterprise has the option to buy hardware and leverage their own models at will which can potentially offset costs and TAM as well. I have talked to a number of large financial corporations in the last 6 months and most have internal initiatives. The same applies in the healthcare vertical.
$250B per annum with AI? That's 20% of global software spend now. Sure, that's possible but that assumes current market prices hold. What if inference ends up normalizing between DeepSeek/Xiaomi & Anthropic/OAI? There's 50% of your revenue and with current costs for inference and training in the US at astronomical levels the US AI industry could also very well be setup to implode overnight.
Lastly I don't believe free can go away anytime soon because it can't. As soon as Anthropic and OAI remove that option those users will move to whatever is. For most of those users it's not a luxury to choose, it is the only option.
The financial engineering occuring right now is something I don't doubt will be text book lessons of the future. We've seen it before and I believe Peter Sorkin when he says that we will see a crash of this bubble, it's just a matter of how catastrophic it ends up being.
A lot of those ‘edge cases’ in the definition of “knowledge worker” are probably the stuff that’s most likely to have significant parts of the work augmented or replaced by AI agents. Like, call-centers are almost certainly going to get turned over in a big way. It’s not like the median tier-1 support operator just reading off a script is much better than an LLM anyway.
Yeah, just looked into this. Knowledge workers is a big group and probably much larger than you think it is.
Basically if you're not doing manual labor, it's probably knowledge work.
Roughly 1/3rd of the working population.
Some data tucked in here: https://gist.github.com/danielmiessler/2dc039762a202b083753b...
> +20% speed for +20% spend isn't going to motivate a trillion dollars a year in spending.
I'm increasingly realizing this math is wrong, because LLM use is really sticky.
If Anthropic 100x'd prices tomorrow for their best model, so some companies offered 50% salary to keep 100% of your AI usage:
a) There are programmers who would take this deal. They've gotten to the point of doing what feels like even less than 50% of the work, developers were already pretty well paid, so they'll take it.
b) There are companies that'd offer this deal. Even if the only people who are taking this deal are not the best engineers, and the AI output is not the greatest, I think the last 6 or so years have seen a lot of companies realize capitalism is not as competitive as it seems.
They're not worried about putting out a worse product because... frankly, what else are you going to do? CF lay a bunch of people off, support gets awful: well you're probably not building a new Cloudflare in the next few years.
In the meantime the AI will get incrementally better, their market share will grow, and you won't be able to compete without taking the same faustian bargain.
-
Maybe I was just naive but it's making me realize how much we take for granted in the world. Both the quality and relative value of things don't have to go up over time. Quality can go down while prices go up, and nothing will really stop it. Competition should stop it, but competition is really slow and can be interfered with. And as prices go up competition gets really hard.
> We're not there yet.
And that's not considering that capitalism is going to do what it does best: if they really found a way to be profitable, competitors are going to fight them on pricing. Anthropic, OpenAI, Google, etcetera 's margins are a competitors' opportunities.
It's not as if there weren't chinese models nearly SOTA. Don't know where the french (Mistral) are but they may try to get in the game if there's a way to be profitable (not that France or the EU for that matter are relevant in anything tech or had any tech company besides ASML and SAP in the Top 100 but who knows).
Your severely underestimating the idea that people are just not going to use developers for certain things in the future
For example I don’t anticipate somebody making a living off of making website ever again
Somebody with absolutely no technical experience who needs a website for their business can now make one with almost no money whatsoever.
That’s good enough for their business. and the code can be totally shit and it does not matter because it’s meeting their business objectives. I am seeing this in the wild and I’m paying money to companies that have these types of websites and because it doesn’t matter I don’t need for the website to work perfectly on all my devices all I need to be able to do is pay them through the website which is what they need me to do and our transaction is done.
Don’t forget ultimately the people who pay technologists right now are primarily advertisers
work on hard problems is going to continue to be some tiny fraction percentage of the software engineering discipline
just expect a total bloodbath because the goal isn’t developer productivity the goal is that “I don’t need to pay somebody $200,000 a year to build a website authoring tool like WordPress.”
Why would a small business use a coding agent to build a custom website when they could use something like Squarespace or Shopify with prebuilt templates that mean they have to know even less than if they were to use some kind of chat UI?
Cause it’s still easier and cheaper apparently
This is the most recent example I found last week for a local barber:
https://news.ycombinator.com/item?id=48166050
They seem to be using Manus: https://manus.im/
And my other assumption is that it immediately integrates with IG/Facebook which is where they do a lot of their marketing
I see no reason that trend is going to slow, especially if you can go to meta to manage your entire business marketing.
Regular people running business just want fast cheaps and good enough.
This is never going to materialise. It’s dead in under 2 years.
The market is shrinking and saturated already and it’s not because of AI gains but geopolitical instability and supply chain issues, some of which are caused by AI spending and stupid ass PE firms refocusing on AI supply chains.
Only our pensions and futures burning.
What do you mean by the market is shrinking?
Literally revenue is collapsing in most sectors. Technology purchasing is declining. Service models are failing to turn a reasonable ROI.
People stopped buying shit.
Wait do you have any numbers to back this up? Every number that I've seen contradicts this. Most sectors have positive revenue growth, even non tech sectors. Technology purchasing is increasing in every bucket (software, IT services, devices, communications, and of course DCs). Retail and food-service sales are up MoM and YoY. Personal consumption is up 0.2% in real terms. I assume by service models you're just talking about AI? I actually may agree with you but this is clearly not true for long if it is true today.
I'm reminded of that [terrifying in hind-sight] Newt Gingrich interview in which he was more concerned about his constituents feelings about things getting worse than any silly statistics provided by government agencies.
https://www.youtube.com/watch?v=xnhJWusyj4I
It's consolidating into fewer, higher value assets. Over 40% of the S&P500 is in companies that are heavily (potentially over) invested in AI.
tech companies have grown disproportionately to other industries, but that says nothing about the growth in other industries
- S&P has a Q1 2026 blended revenue growth of 11.3% according to FactSet - most sectors are growing, not just tech
I feel like there's a bit of AI psychosis in this particular post.
>"These are tools which burn vastly more tokens, but are also quickly becoming daily drivers for the work carried out by extremely well-compensated professionals."
>"Somehow this fragment turned into headlines like Uber’s COO says it’s getting harder to justify the money spent on AI tokenmaxxing, because the market for stories about AI failures remains enormous."
Yes, it's just the yearning for AI failures. It couldn't possibly be runaway costs, record revenues, and massive layoffs. It couldn't possibly be that these tools are lighting dollars on fire by people already paid significantly well and not producing any increase in "value" for it (I recognize that output is 100x but outcomes are flat by all measures).
[1] https://cmr.berkeley.edu/2025/10/seven-myths-about-ai-and-pr... [2] https://futuretech.mit.edu/publication/crashing-waves-vs-ris...
Being pedantic, but I don't want to lose the meaning of the term: "AI psychosis" doesn't refer to someone who thinks AI is really good. It refers to someone who develops symptoms of psychosis from talking to an LLM, e.g. believing they have developed a new Grand Unified Theory of physics.
Fair and I would edit if I still had time. How about "AI brain fry"?
I don't know, "workaday professionals will find $200/month a particularly good deal, such that there will be widespread adoption" sounds either credulous enough to support the diagnosis or dishonest enough to dismiss. I am a "knowledge worker" who is doin' fine, has a lot of templated written work/report writing, and there is no way in hell I am justifying that kind of spending to my boss or my family.
"I firmly believe this technology will create business value" is so obviously and categorically different from "Humanity has birthed a silicon god that I have also developed romantic feelings for" that I'm not sure if your comment is even trying to be in good faith
What sort of bad faith would even apply here? idgaf if poster x or y has psychosis or not. "$200 a month is classic addict behavior" seems pretty spot-on to me though, I just don't want to have to pay it, too.
I think is the classic dilemma where people don’t know how to value their time.
Typical tech worker costs a company around $100/hour minimum. That $200 subscription cost can look mighty attractive if it saves some time or mental load.
I don’t think there is anything about addiction or spooky with that math. I suspect a lot of this is coming from tokenmaxxing firms but on the flip side on our small team, we end up spending about $200 per person per month for tokens using tools like Cursor. We feel the spend is justified with measurable value.
You can't justify $200/month in spending to your boss? Many people charge more than that per single billable hour. I would put your salary side by side with that number, which is your boss's perspective, and reconsider.
>You can't justify $200/month in spending to your boss?
No. What? Of course not.
>Many people charge more than that per single billable hour.
hrmmmm not so sure about the work that "many" is doing there
Fully loaded costs for an average employee at a bigco are scary. Not $200 but significantly higher than the number on your W2 by the time the company pays vacation, benefits, unemployment insurance, etc.
I've seen people try to argue for resources using reasonable but abstract arguments, and it just never works. The fact of the matter is that I want $200/month, on top of my other asks, and that comes out of somebody's budget. I just don't see folks snapping their fingers and $200 a month (discounted for now!) appears. Good for the other guy I was replying to if that's the case! I just don't see it though.
I just cannot imagine zero net productivity gain from AI usage by a somehow experienced developer. Even if it’s happening right now it’s matter of time for people to figure out how to utilize it right. Only exception would be if people are getting the same output but working less, which is still a net positive (just not in terms of output).
Please let's not dilute the meaning of 'AI psychosis'. It is a real phenomenon that involves actual psychosis.
What's the psychosis?
Sometimes it feels like theres this opposite AI psychosis, where anything AI is bad and boils the ocean, takes our jobs and makes RAM expensive. Its a component in the current economy, but things like tariffs, closing the strait of hormuz etc is equally bad for the economy. Anyway, just find it strange to be so militantly anti a certain tech.
That’s the modern internet. What sells is the most overdramatic doom and gloom take possible.
It's more the tech leaders than the internet. Sam Altman, Jeff Bezos, Eric Schmidt and such get up on stage or interview regularly with a shit eating grin telling us all about how they are coming for our jobs, will make us obsolete, and there is nothing you can do about it.
It's a natural response for society to despise these people who have such contempt for us. It's almost embarrassing these days being at a social function and telling people I work in software, it's got a negative stigma almost like working in gambling or the military.
I don’t know about gambling, but if “working in the military” has a stigma, I humbly suggest seeking out different social functions.
I don't have to because I don't write software for child seeking missiles for Palantir
I would say the exact opposite
That sounds more like a you problem.
I think it's a matter of perception because I didn't interpret any of them as being gleeful about it. If you think about it, "AI will take your jobs and maybe destroy the world" is horrible, horrible marketing -- like, your comment is a perfect illustration of how it is received everywhere -- and yet these CEBros can't stop saying it, which indicates to me that they actually believe it.
Oh, now that their IPOs are nigh they're changing their tunes (https://archive.md/s9EO3) but to me that looks more like they've decided to let $$$ prospects override what they really think.
The general public is not their customer base, they don't have nearly enough money to spend on AI. Going to the media and saying "this product can automate so many jobs" is marketing to other businesses who want to use it to cut their workers out.
There was crazy clip of Eric from Google telling a crowd of university students that in the future AI will do everything, and after the whole audience boos him he keeps pushing the point that they better accept it and get on board. The mentality these guys have is sickening. They have no humility and no humanity.
>Anyway, just find it strange to be so militantly anti a certain tech.
You know, that's fair. I'm much more against super-rich investing hundreds of billions in the things they don't understand, creating massive disruptions in their wake.
AI didn't create stupidity and greed. In the end, it's just another tech. I'm just tired, time and again, of people who I would hope to know better (and repeatedly they prove me wrong).
> takes our jobs and makes RAM expensive
I mean it is doing both of those, so thats fair to be honest.
Hi Simon; while it's true that the token providers have found that their product is 90% useful to devs and 10% useful to everyone else, this is something they found out in the first quarter of 2025 anyway.
It's not exactly news, is what I'm saying. And even with the PMF they found, the product is still only a commodity i.e `tokens`, which is what every other provider on the planet is also providing.
All their other products boil down to "harnesses", which does not look viable as a product in the sense of PMF - you cannot sell it, you cannot lock it to your own subscription, API, etc. so you can't use it to generate revenue any more than the free harnesses do.
PMF has a specific meaning, and "code harness" or "coding model" does not satisfy the commonly accepted meaning. Maybe Mythos (or similar) will.
> because the market for stories about AI failures remains enormous
How enormous? 1 trillion dollars, 2, 10 trillion enormous?
That leaders are completely fine with impoverishing vast swaths of American workers because of "progress?"
If they are then yes, that's psychotic. Not sure how it's relevant to my article about Anthropic and OpenAI's enterprise pricing though.
I’m so curious: let’s say you’re the president or the CEO of a major tech corporation building frontier AI systems.
What are your directives?
I would never be in this position because I'm not an anti-human extremist that wants to hoard wealth and misallocate capital.
Who’s an anti-human extremest? Doesn’t that sort of warrant some sort of facts to justify such an extreme position? What makes you feel this is a misallocation of capital OR something…anti human?
You can’t just redefine ‘AI Psychosis’ to mean ‘somebody whose opinion about ai I disagree with’
I find this analysis confusing. PMF for coding was likely reached some time last year. Profitability, which is different, we don’t know. The article kind of confuses both without making a strong economic case or using numbers in a compelling way. I don’t understand what the Uber case has to do with this either. The Uber COO clearly said that at least in terms of ROI he’s not seeing the results either.
My take is the product has been very useful for coding (PMF) for months. But it’s certainly not useful at any cost…
What I also find confusing though is that folks seem to ignore trajectory which is maybe the biggest lede to bury. As Simon says, we have had "good enough" coding agents for 6 months, that is a blink of an eye, and at my company my job has now completely changed. It's almost like a dream.
And that's just one inflection point. We've had several and there are many more on the horizon. So while I could be convinced that ROI is maybe not even positive today despite the ridiculous enterprise spend, it's perfectly rational to pave the way today for what's coming over the next few months let alone years down the line.
There may be additional major leaps forward, and there may not. I kind of struggle to imagine what the next step actually is. Certainly there will be improvements in performance (speed) and cost. But at a point you reach a barrier where the limiting factor is the specificity of the human prompt and our ability to manage all the code we’re generating.
Somewhat oversimplifying; writing software and building apps was a bottleneck - now it is not. What is the next bottleneck that LLMs can solve? Is there one? And is there enough publicly available data to solve it repeatably at scale? Or did we just automate stack overflow searches and now we’re stuck again?
Or is the endgame of this innovation cycle the complete removal of interaction with machines through code? Will we simply interact with machine coworkers purely through natural language? Can an LLM make PowerPoint slides and run a meeting? So far not seeing much progress on that.
Judging from the fact that the Opus 4.5 inflection point was not really anticipated, and we still don’t really know what threshold was crossed that suddenly made agentic coding accessible to so many more people, I think it’s safe to say we don’t know what the thresholds will be until they’re crossed. The fact that we don’t know exactly what they’ll be isn’t a good reason to think there won’t be any more.
> The fact that we don’t know exactly what they’ll be isn’t a good reason to think there won’t be any more.
Nor is it a good reason to think there will be more.
We should expect to see the process slowing down first. Until then we should expect it to continue with pretty high likelihood.
https://substackcdn.com/image/fetch/$s_!_ZW2!,f_auto,q_auto:...
Based on how much money is chasing returns, and how steep the slope is, it's almost certain that we are still not at the end of this sigmoid cycle.
Sure, it might start to slow down, but even then we will likely see a doubling in the next 10-15 years.
https://substackcdn.com/image/fetch/$s_!_ZW2!,f_auto,q_auto:...
I am currently eating lunch. Meanwhile Claude is triaging and writing reproducers for 70+ tickets nobody has had time to look at. Next it will attempt to fix them. I have not read the tickets. I will not look at the code until there are review ready PRs and a code review bot have done the first pass.
In other words, most of the prompting will also go away.
yeah but if you have to pay $2k to $3k per month, would you still use it?
Pmf is this weirdly defined thing where "if you're not sure you have it then you don't".
I think it was clearly useful for months to people who had tried it and taken the time to understand it, but now that knowledge has spread to the point where wallet holders are convinced it's not just passing fad or hype so now pmf can be "claimed".
I agree it's weird to say "those people have pmf" though, usually it's something you define for yourself
> Pmf is this weirdly defined thing where "if you're not sure you have it then you don't".
I'm not sure if this runs counter to your point or not, but: I don't see any future where LLMs aren't a core part of Software Engineering. The horse is out of the barn. There is no going back.
Yeah but the product is not “LLM” it’s “proprietary frontier model LLM paid by the token”.
And I don’t even necessarily disagree with OP! It’s more like the competition is shifting so quickly that your competitors could undercut your PMF in a blink of an eye.
There will be cheaper solutions. And they will generally be less capable than the more expensive ones. Just like most other products.
But my guess is that the cost of SWEs themselves mean that the more expensive ones will be worth the delta to most companies.
But time will tell.
History bears out that cheap and satisficing soundly beats expensive and optimal every time. Until we have smarter and more prescient decision makers in leadership, the bottleneck on output will be the quality of decision making not the quality of code. Trying more things faster and cheaper will win.
Aka the cheap plastic solution always wins.
True but that is maybe 5% of what is being promised by the average booster
Give examples of boosters (average or not) and what they've promised?
> clearly useful for people who took the time to understand it
people -> programmers, I haven’t met a non-developer who reports getting more time out of current AI platforms than they put in. If anything I’ve anecdotally heard the opposite, introducing AI at work creates so much slop (output) it takes more time to process it all without a tangible bump in overall productivity
I have at least a half dozen examples of people not hiring people or buying other tools/subscriptions because they built their own with Claude
The article also treats the word "good" as load-bearing in a way that should have you questioning their analysis:
"I’ve called November 2025 the November inflection point because that was when GPT-5.1 and Opus 4.5, combined with their respective coding agent harnesses, got good—good enough that we’ve spent the last six months adapting to agent systems that can reliably get useful work done."
Yet it’s backed up by adoption across the industry
MongoDB was once backed up by adoption across the industry. Or for a more recent example, blockchain took off like wildfire across the industry before ultimately fizzling out in all but the most niche applications.
Not saying this trend will do the same, just that the industry adopting something doesn't guarantee its success.
I don’t think those are really comparable. The blockchain was trendy hype, relatively few companies actually adopted it. Where did Netflix use the blockchain? Google?
By comparison almost all tech companies I know have leaned heavily into AI.
Correct the cost is part of the economics.
Thats why most here shouldn’t engage in the discussion - they parrot on about benefits without identifying and articulating the costs and moreover how it affects the firms financial position.
And what happens when open models catch up in 6 months or so?
It’s not supposed to be logical, it’s an LLM evangelism blog that rarely, if ever, has any critical analysis that isn’t pro-industry. Read any/all of the other posts and you won’t find much skepticism but you will find a lot of shilling how great it all is.
I like his other posts. He's bullish on AI, which is fine. I'd like to read a mix of bearish and bullish level-headed takes from people who are subject matter experts. His technical credentials are well past discussion - I love Django, and he comes across as a pretty upbeat but level-headed guy. Certainly beats radical takes in either direction from people who have no clue what they're talking about. It's just this article that I find rather confusing.
The thing that matters most to me is if reading what I wrote teaches you some new things and gives you something useful to think about.
If I make an argument and you disagree that's fine with me, provided I didn't use misinformation or sloppy thinking in making that argument.
That's how I feel about most of your writing. I click through most times when I see you either on the front page or in the comments, and I generally walk away feeling like I have food for thought, without necessarily buying everything wholesale. It's part of why I keep coming back.
My root comment simply represented my two cents about the current post. I don't think anything about the post is outrageously incorrect or anything, just somewhat confusing. You're a very prolific contributor in this community and I don't think me or anyone else that welcomes your takes expects everything you write to rock our collective socks every single time, anyway.
308 posts on AI ethics: https://simonwillison.net/tags/ai-ethics/
52 on AI misuse: https://simonwillison.net/tags/ai-misuse/
149 on the unsolved challenge of prompt injection: https://simonwillison.net/tags/prompt-injection/
40 on slop: https://simonwillison.net/tags/slop/
If you want an "LLM evangelism blog that rarely, if ever, has any critical analysis that isn’t pro-industry" there are plenty out there. I'm not one of them.
People are confusing "excitement" with "evangelism". Your blog is definitely on the pro-AI side of things, but as you say, it's not one-sided or uncritical.
I think you should highlight your exemplary pre-AI writing too.
All of these are about AI misuse, not skepticism of AI. By skepticism I mean doubting whether AI actually delivers on its promises which, based on this last post, sounds like something you think we're already past.
Many people still think AI coding agents are slop on steroids despite all the current hype around AI actually shipping functional products.
It's hard for me to write about skepticism that coding agents deliver on their promises when I've been using them daily and know, for an absolute fact, that they boost my own productivity.
(And that's after taking into account the METR paper that says engineers over-estimate their productivity with these tools.)
I have plenty of doubts about AI delivering on its promises outside of coding. I don't write about AGI because I think it's science-fiction hysteria. I write about slop precisely because it represents a mis-use of AI that demonstrates people completely misunderstanding what it's useful for.
Love when people say "its promises". What specifically are you disappointed with? Simon's posts are high quality and evidence driven. AI has already delivered an incredible amount. Read Epoch for industry trends and analyses, METR to, everything points to a pretty consistent picture.
"Many people still think AI coding agents are slop on steroids despite all the current hype around AI actually shipping functional products."
Oh yes, tons and tons, especially on HN. But the plural of anecdote is not data. Enterprise spend speaks for itself. You are using AI-coded functional products all the time. Do you want like a diff history for the Google codebase or something?
Tbf the OPs blog and comments (including their sibling to your comment) are also heavily anecdotal.
> I’ve called November 2025 the November inflection point because that was when GPT-5.1 and Opus 4.5, combined with their respective coding agent harnesses, got good—good enough that we’ve spent the last six months adapting to agent systems that can reliably get useful work done.
Claiming a grand inflection point based on your own personal usage is very anecdotal.
If that were it I would absolutely agree with you. But this experience maps exactly to adoption trends. My job in the last 6 months has become so unrecognizeable to me it’s insane, the adoption at the very least at large companies is truly truly incredible, and it really does coincide with the quality of opus 4.5 (which has now been surpassed).
"Adoption trends" are just herd behavior which may or may not be driven by compelling anecdotes and may or may not be evidence of something more. I'm just saying it seems wrong to dismiss the post the way you did when the OP in question and your own post here are just more anecdotes.
No, if that were really true you wouldn’t see what you’re seeing today. You wouldn’t see entire companies completely retooled and refactored around these tools. You would see the mistake of “this is actually just herd behavior”, which involves such a colossal amount of impact to these companies entire stack and bottom line, resulting in systemic collapse. You don’t see that. Company leadership are not some idiot class of people, I don’t know why this is people’s prior. If companies get adoption wrong in either direction they are completely screwed. So you’re seeing people putting money where their mouth is, across the board.
Compelling anecdotes are not even the main source of evidence. Look at the enormous body of work on measurement of these systems. I always point people to epoch capability index as a good summary statistic of capabilities or METRs time horizon data which has now been topped out. They had a recent updated to the dataset, after which the corrected plots pointed to an even faster acceleration than before.
I think my claim about November is looking very solid today.
My point was claiming a broad inflection point based on your own personal usage is not "evidence driven", it's anecdote-driven. It's hard to disprove any claim you made because you didn't really make one that's disprovable, and your opinion on it now is still just an opinion.
Yes, my opinions are driven by anecdotal evidence. I think that's fine: I have a pretty good track record, and I'm careful to share my reasoning.
If you want indisputable, data-driven information about the state of the LLM world I guess you can wait for a peer-reviewed academic paper?
Those have been around for a long long time, you may be focusing on anecdotes but the adoption numbers and performance trends speak for themselves and we’ve had performance trends for years. People can argue about whether or not enterprise level adoption has a clear ROI today but the fact that we’re at the point where entire large scale companies are already completely refactored, directly after opus 4.5, if that’s not a convincing enough signal I don’t know what is.
So how do openai and anthropic plan to keep customers when GLM-5.1 is just as good and open source and a lot cheaper?
I don't see the business model working. My closest friend actually does automation software for large companies.
He does not use Claude or openai at all. He primarily uses gpt 120b on cerebras and glm-5.1 for heavy thinking work. And some other small models for various tasks. All open source.
And these systems are extremely useful for the businesses and are able to run fully automated pipelines that are very stable and fast.
We discuss this a lot, and we both think any business doing heavy agentic work on Claude and openai just aren't aware of exactly how good and cheap open source has gotten on the last year.
So... once the legacy businesses and developers catch up, won't Claude and openai be unable to recoup their costs?
> I don't see the business model working.
Same. It's a nightmare from a Porter's Five Forces perspective.
There will be a ton of businesses competing in this space, and there will be something of a moat due to how capital intensive the business can be, but there will still basically be infinite competitors.
Great for consumers.
Well in reality AWS will just host one of them and most companies will use that
Like how snapchat kind of fell off because the feature could just be a subset of instagram
It seems like it would just become a commodity like EC2
Snapchat is huge and growing
For coding you always want to go with the best model in the category, not something that would be the best model if we went 1 year back which GLM 5.1 is, and I'm saying that as a big fan of GLM cause I run a translation site where GLM is good enough for the price.
Most of the money right now is in coding. Openai and Anthropic just have to be 6 months ahead of SOTA open source models and they'll capture most of the enterprise and dev market
Yes I'm an engineer (20 years most in games/graphics industry) and only use it for code. I've been using glm 5.1 this week a lot. I went in expecting another "decent" but not really "up to standard" open source model.
I highly doubt I'll ever use Claude again.
I think you are wrong about Claude being any significant level better
I've been mostly coding with GLM-5.1 as well and I agree with you. DeepSeek V4 Flash is another very good surprise. Incredibly cheap, fast and effective.
I've been using DeepSeek v4 Flash with OpenCode for the whole week to refactor a Terraform code base I inherited and it worked surprisingly well.
Well I think there are a multitude of harder measurements that would disagree with you, but ultimately there is absolutely a use case for cheaper open models (or even cheaper tiers of proprietary models) and in fact the unsolved optimization everyone is trying to get to is how much spend to use for a given task. But there will always be a market, especially in enterprise, for the best performance there is to offer
Why are you boosting so hard? Lmao either you’re a paid poster or you own stock in a frontier firm. Which one?
Sadly neither :(
I strongly disagree. I'm an engineer - I'm all about the fastest, cheapest thing that meets the requirements. I don't need Opus 4.7, even for my complex programming tasks. It costs over 10x other models available that still give good enough answers. Those smaller models are also a lot faster to output tokens, which saves me time.
Once the model gets good enough, the returns on bigger models diminishes quickly. I don't want to spend 10x the money and wait 5x the time to get answers that are equivalent.
Same here, i can't say i've seen any difference in 4.6 vs 4.7 other than price
I have stats from a harness that tells me glm5.1 is far more cost effective for us than Opus with the rate of defects and rework taken into account. In fact, with a decent harness I'm now increasingly favouring eHaiku over Opus for execution too. Opus is still worth it for planning, though, and far better at one-shotting things.
If I generate code with Claude, ChatGPT, and GLM 5.1, I can't say which model is which reliably. I exclusively use Claude more out of superstition than reason.
> For coding you always want to go with the best model in the category [..]
And this is why many companies go out of business. You always want the best bang for your buck, sometimes this is the "best model" and sometimes it is not.
Why? If it's good enough, it's good enough. Though I read the code that gets vibed so maybe my use-case is different.
It's driven a lot by the harness too. If you're using claude code, you're actively being pushed towards newer models, even though older ones work perfectly fine for your use cases
For coding like for everything else in life cost is a factor.
Cost for the value delivered. Like if you offered the current SOTA open source models at $0.1/M, I still think I'd be using Opus or 5.5 at $30/M. Or say GPT 5 which was released Aug 25, I don't think I'd use it for coding for even $0.1. I'd def find other uses for it(translations, agentic workflows, prompt guards etc), but for coding I don't think I'd ever completely switch to a SOTA open model
Unless ofc there was an actual speed difference, only reason I'd be willing to go with a worse model couple of percent worse than current best model is if the speed was at least 5x higher. Looking forward to kimi k2.6 offered publicly by Cerebras
> I still think I'd be using
That's fine. Other people may not want to pay 300x more and will rather make do with last year's SOTA.
> For coding you always want to go with the best model
Maybe you meant "For coding I always want to go with the best model"?
Based on current market for LLMs I'd say my use of "you" in the general is fine. Even openrouter which doesn't capture all of the SOTA closed models but nearly all of opensource model usage has Opus as 1st(on last week) on "Programming" category and 3rd in overall rankings
https://openrouter.ai/rankings
I'd trust the OpenRouter rankings a lot more if they exposed the number of unique users for each model, as opposed to just a token count.
Currently I have no way of telling if big changes in their rankings are caused by a single "whale" switching providers, or if it's a more meaningful trend.
My point was that even openrouter, the one place people who are looking for open source SOTA models go to, doesn't definitively have opensource models at the top. Esp considering quite a lot of the closed models usage is through AWS, GCP , Azure etc, probably dwarfing the usage on openrouter by a huge factor
> For coding you always want to go with the best model in the category
This is transparently false, because the best "model" is still competent human developers. They're just more expensive. If you're willing to use current LLMs at all, it means you're willing to sacrifice quality for a better price, and your disagreement with the comment you were replying to is entirely about what the optimum tradeoff is.
Well it may be false that you always want the best model, but the point is performance of you+<agent> is far more cost effective than you+someone else
Maybe, but that's a different claim than the one I was responding to. And also raises the question of "if the lower quality but cheaper output of frontier models is more cost effective than humans, is the even lower quality but even cheaper output of OSS models is more cost effective still?" With an absolute rule like GP suggested ("no, you always want the best code generator") the answer is clear, but it get much murkier if you reject such rules (as you have to to be an LLM coding proponent)
It was true 6 months ago, not anymore. Frontier models now outperform developers on many tasks, be it on quality/readability/maintainability, and let’s not talk about speed…
I've seen the code they produce without extensive help from human developers, this is clearly false.
Good to see the classic "yeah the models weren't good enough six months ago, but this time they actually are, promise! Please forget you were hearing the exact same thing six months ago!" is alive and well though.
Why is anthropic hiring software developers then?
>For coding you always want to go with the best model in the category, not something that would be the best model if we went 1 year back which GLM 5.1 is, and I'm saying that as a big fan of GLM cause I run a translation site where GLM is good enough for the price.
Currently, the difference is substantial, but what happens if capabilities saturate?
Then the house of cards comes crumbling down, but there is so much evidence to point to this not happening that it requires a bit of a theory for how that may happen
> but there is so much evidence to point to this not happening
Could you explain this?
Well I think there are several fairly stable trends that paint a pretty compelling picture:
- performance scales with compute very very reliably. We have “scaling laws” (and have for years) and they are almost miraculously stable and show no sign of being invalidated at all even at the very largest scales. There are some theoretical bases for this though I’m not as familiar with the details
- these scaling laws are on an unintuitive quantity (validation loss on pretraining datasets), so we can look at downstream performance. Benchmarks are a minefield of junk but there are many decent ones and enough variety of techniques and data sources and scoring methods etc that in aggregate they are useful. The single number that I think is the best summary statistic across the crazy (O(100k)) number of benchmarks is the “epoch capability index” (just some branding over a reasonably standard statistical model that was really well thought out and a great idea). The trends in this are extremely stable. Eyeballing the trend over time on their graph we’re getting basically a GPT-4 to GPT-5 level capability improvement every ~18 months
- coding agents are not limited by the quality of the human training data they’re trained on, this is such a massive misconception: human data is only a bootstrap to a reinforcement learning phase. This combined with the fact that we have verifiable rewards means it’s just a matter of when not if for any given level of reliability.
- the massive compute investment implies that the compute that we’re building over the next 2-3 years will 10x the effective compute for training models. That combined with various R&D contributions (historically which have been very significant and there is no shortage of wins here), better data curation and flywheels, richer data (wait until conversation capability gets good) means we have several orders of magnitude of runway that we know of, today.
In short I don’t see any compelling evidence to suggest all of the trends we observe in many different ways will end any time soon.
And you propose the same companies that have been cost cutting and avoiding buying you a chair for ever won't start objecting to a $200/dev/month subscription? The finance department won't have a say?
> For coding you always want to go with the best model in the category
Will this always be true? There will never be an event horizon/point of diminishing returns where something not-bleeding-edge is "good enough" for 51%+ of users?
As long as closed source is 6 months ahead in terms of current difference. Although this is hard to figure out using simple percent based coding benchmarks, you def. notice it when you're actually trying to do a long task. Even simple things like UI "taste" is enough for me to use opus instead of 5.5 though even though 5.5 is strictly better for anything that doesn't have a UI, ie backend, scripts, making agent workflows etc
This is a silly take. There is a line of "good enough" for most coding (most CRUD apps and APIs are nothing special), and once we are past that, nobody will care about having the "newest, best" model except extreme outliers. And this base "good enough" model will become an ultra cheap commodity as we already see with GLM, deepseek, etc.
As long as closed models are 6 months ahead I won't be switching from them to prev. 6 month SOTA open source models. Maybe its just a different calculation if you're in a job, but as an indiehacker I'll take any edge I can get
Ofc again, can be convinced to switch if there's however a clear speed difference, like 5x+ for a open source sota even if it was SOTA for 6 months ago
> For XXX you always want to go with XXX, not XXX
Oh, hey, I recognize you. Thank you for the very forward and thorough orbital sander recommendation at Home Depot. That's exactly what I wanted to deal with on my holiday weekend. You just know so much about this and the rest of us are simple passersbys.
Yep sorry was just pulling it out my rear, not like a market trend that nearly every enterprise uses Anthropic or Openai models for coding or that Anthropic has had such ridiculous growth that they're 10x-ing year over year
I'm ribbing you for writing like a condescending guru that invalidates the evaluatory capability of your peers. Not the meat of your justification, which I notice has now shifted from saying they're quality leaders to market leaders.
Most work is not coding.
And also, people have it wrong… their models are not the main problem anymore. It’s the RAG
Would love to hear more about your thought about the RAG.
I think RAG is a mostly outdated concept now, it's been subsumed by the idea of a "agent harness" which is exactly what Claude Code and Claude Cowork and OpenAI Codex and Claude.ai and ChatGPT themselves have now become.
An agent harness with access to a good search tool is a much more interesting thing than 2024-era RAG systems.
Depending on RAG is a workflow problem, not an AI problem
For coding assistance, I have tried OpenCode with several large open models through OpenRouter. All were fairly bad compared to Claude Opus. Could you provide some hints on how I should be holding these open models so that I might get more value out of them?
I agree with the common trope that open models lag behind by about a year, but something magical happened just around a year ago when the state of the art models became extremely useful. By this reasoning we're about to see open models perform well, but I'm afraid there is more to it than just waiting for another revolution around the sun.
Note, my application is coding assistance. Open models can be great for other purposes.
I tried almost all OS models on opencode, none of them is on levels as opus 4.7.
In latest experiment I used opus for implementation plan then used cursor composer 2.5 for execution.
I must say that combo is really good. Main drawback of claude code is that is super slow. So when paired with composer that is super fast it flies.
No one is claiming that OS is as good. They are saying it isn't that far behind SOTA commercial products. So why pay exorbitantly just to get something only a few percent better than the free option?
But there have been very good open source office apps for decades and few enterprises use them, so perhaps this is just the nature of B2B purchasing committees and 'nobody getting fired for buying IBM.'
Do more planning yourself, be smart about the context, break down tasks into smaller components, give it more guidance. You can't just lazily prompt it to complete large features autonomously and expect good results.
But if the closed-source models can do this without the additional effort, that's a significant gap, no?
See that's the thing, they can't. Every model needs hand holding and guidance.
some require less hand-holding than others though
No one is trying to argue that OS models are better than Opus 4.7. It's simply that they're good enough and cheaper.
The point is that the price gap is so much larger than the capability gap, that even with the extra compute needed to make up for the lack of capability, you can still come out ahead in terms of amortized $/work done.
Is it really when they are hundreds of times more expensive?
That is the 3-6 month sota-open gap people talk about, a time-window that continues to move as new models are released on both sides.
Do you know what economic trade offs are?
Both implicit and explicit..?
+1 .. just wanted to reiterate that this is the answer. The open models work great if you just do a little more of the design/architectural work up front and organize your work appropriately.
a good harness is supposed to do what you are describing. sonnet on pi.dev is pretty terrible but fast. Claude Code has ridiculous amounts of prompt engineering at system prompt level and sub session spawing combined with low temperature, to provide the predictable results people like. CC screws up and you never see, because the harness auto corrects, while on OSS you see everything, and does not comes with the level of monitoring by default.
GLM-5.1 isn't just as good. It is no match for Opus running in Claude Code. Please try it yourself. Open source models are about a year behind at least.
This is profoundly misinformed. I use all three of those models regularly and the difference is just not that big anymore. GLM 5.1 is at least as good as Opus 4.5 - when it’s my dime it’s the primary model I use and switch to GPT 5.5 for planning and review but it’s also very capable at those things. If I had to pay API rates for everything there is no question I would only use GLM 5.1 (and Minimax for exploration tasks).
At work I mostly use Claude Code and a bit of Codex; personal projects are OpenCode and honestly I prefer it.
I would agree here. And in my experience Qwen 27B and Deepseek v4 are also extremely good.
None of them are quite opus, but they are damned close and a no brainer if you care at all about cost.
In the second half of last year, I found that agentic coding with proprietary models (≈ vibe coding) reached the point where it actually speeds up my ability to deliver useful code at work. Before that, AI-based autocomplete definitely helped, but (despite the claims of the people selling AI coding tools) letting an agent author more than a file or so at a time (often a function or so at a time) required a very intricate plan or it would create a mess. Creating that plan or cleaning up the mess would take longer than just doing everything myself.
For me, it feels like widely available open models have recently crossed that same canyon. Are they as good as e.g. late-model Claude Opus? I don't think so. But they have absolutely gotten past the point where they are beneficial. This means that, for me, they are about six months behind.
Exactly this. GLM 5.1 is the first open model that I thought "actually worked" for agentic coding, which puts it in the same tier as Opus 4.5 - which was where I flipped.
For coding I wouldn't say a year, last year this time claude or gpt definitely weren't able to do what GLM is able to do today, but easily 6 months I'd say.
Not sure about other domains though.
A year behind is still very very very good at this price. ;)
I use composer-2 daily for complex programming tasks. It's a fine tuned Kimi 2.5 - nothing groundbreaking. I've even had reasonable success using Qwen 3.5 on my desktop GPU. Opus might be better, but it's certainly not necessary to get good results.
I have 26 years experience. I code using GLM-5.1. Fron time to time I switch to Codex / Claude, and honestly I don't understand why people uses Claude or codex. With the right prompting, GLM is awesome.
Don't you need to spend 5-10 thousand USD to run these models that are "as good" as frontier models from 6-12 months ago? I haven't seen a convincing breakdown for ROI of running your own coding models. Especially against a $20 or even $200 plan
I assume you can run them in the cloud. $5-10k doesn't sound like remotely enough to run a not-shit model locally based on my experience.
Agree. Also reasonix with deepseek is super cheap and quality is only slightly worse (in my experience)
Do you have a good source to refer to, to map out migration from Claude code to a cheap setup using small open source models like you’re describing? I’d certainly like to experience how good they’ve gotten.
The only way I see it working out for them is if some legislation is passed that eliminates the competition by making it illegal to run local models. They could claim that the models are dangerous and could be weaponized without oversight, or something along those lines.
They are both (and also spacex) sprinting for IPOs. They know that the opportunity window is closing fast and that advancement in model quality has largely plateaued in the last year. Take as much investor money as you can get away with for now.
> $2,180.16 worth of tokens for $200
“Tokens” don’t have an intrisic cost or value. Saying that I used $2,180.16 worth of tokens is like relying on the salesperson to convince me I’m getting a billion dollars worth of pots and pans for $19.99.
I think it’s funny how we are throwing critical thinking out the window when it comes to evaluating biased sources of info.
I'm not sure what you're pushing back against here.
I spent $200. If I had been paying API pricing it would have been $2,180.16. The article is about how enterprise customers get charged API pricing, which means if I had been employed by one of those companies I would have cost them $2,180.16.
What am I missing?
Just because API pricing would've been $2180.16 doesn't mean that's the value of those tokens. For starters, you personally probably wouldn't have paid that. But also, sales price isn't value. This is like saying, oh, I saw this bar of gold somewhere for $10000 but got it here for $1000! So I got $10000 worth of gold for $1000! - no, the value of that gold is determined by its weight, which wasn't even mentioned.
We have no market convergence on tokens yet (and it'll differ between LLMs), so it's impossible to say what value you got for your $200.
He's saying he's getting a great deal...a token from Opus on Claude code is the same as a token from Opus on the API. I remain as confused as Simon. He's not talking about "here's the ROI I got from my $100 subscription" it's "here's how much I saved from getting the monthly subscription instead of sending things through an API".
Right, the confusion is that the quote-unquote "subsidized" monthly pricing is often used by Anthropic/OpenAI skeptics as proof that inference is unprofitable, i.e. the API would have cost $2000 but you only paid $200 for a subscription, therefore OpenAI is selling dollars for 95 cents and the house of cards is about to collapse. As the GP says, this is faulty logic because we don't know what the actual cost of a token is; OpenAI might only pay $1 in inference costs, in which case they're merely "incredibly profitable" making $199 off you instead of "ludicrously profitable" making $1999 off you had you used the API.
But to your point, re-reading the article, this is not what Simon is saying at all; he's just pointing out that he got to use ~$2000 "worth" of tokens on his $200 plan. Which makes total sense! Subscriptions are sticky, that's why the entire software industry moved towards subscription models (as much as we hate it); the person paying $200/month is more likely to stick around than the person who paid $2000 using the API.
Which is a terrible metric sending a wrong message.
He’s just making a point that $100/mo subscription is cheaper than the API costs and that the same work would cost 10x as much on API pricing. It’s not some sort of swindle it’s a trade between consistent revenue that power users will overspend on but plenty of people will underspend (I’m on the underspend side) or pay-per-request service that makes more sense for say scaling out something custom (I run APIs a lot when I have say some pseudo labeling pipeline I want to scale to a lot of data. For this I typically use smaller models but lots of cases where you want larger models). What’s the wrong message here?
No, value is determined by what participants in a market are willing to pay for something. The only reason you are able to say that the value of gold is determined by its weight is that gold is a commodity and no matter what you paid for it you'll find others willing to pay market price.
Simon is saying that companies are (today) willing to pay API prices for tokens which is as good as any determination of value.
Is this some anti-FIAT take? The value the author got is not value as in intrinsic value, it simply means value as in better deal than the alternative. This is often called "value" and you will see this used when products are sold in "value packs" etc.
> Just because API pricing would've been $2180.16 doesn't mean that's the value of those tokens.
You seem to be suggesting the price of tokens is entirely disconnected to the cost of providing the service? I don't see much basis for that assumption.
I'm willing to charge you $100k for those same tokens.
Does that mean you'll be saving $99k?
It sounds an awful lot like the mark-up to mark-down scheme where the price stays the same.
Large enterprises make deals and won’t be paying 2,180.16$ either. Just like with AWS
That doesn't seem to be the case. From what I've seen enterprise deals get API pricing now. Have you seen evidence that's not true?
Hi Simon, nice article. The parent there may be making the same assumption I am, that large enterprise _never_ pays sticker price.
Also, to just color in the picture here, as I haven't seen it mentioned elsewhere, there is a very large Saas company at the moment who has given everyone unlimited tokens on Claude. And they have a dashboard showing who spends the most. So the "budget" went from about USD500 per per person (split between Claude and cursor) in Jan to... Well a soft limit of USD100k... Per month... Per person.
People can still see the top line sticker price on their spend, but honestly I can't believe that the Saas is paying that full price when the invoice comes in.
That said, there are some finance reports which are probably dropping soon where we will find out!
> The parent there may be making the same assumption I am, that large enterprise _never_ pays sticker price.
I shared that assumption until yesterday, when I found out that it wasn't holding for LLM pricing from OpenAI and Anthropic. That's what inspired me to write this piece.
I think those token leaderboards are an obviously terrible idea and will go extinct very quickly now that people are paying attention to costs.
But the feature list at https://claude.com/pricing#team-&-enterprise literally lists "tiered incentives on committed spend" and "non-standard terms" as perks of the sales-assisted Enterprise plan. Maybe "non-standard terms" could mean "we dance for you if you pay", but what would "tiered incentives on committed spend" mean besides "we can negotiate on price if you bring the volume"
Anthropic advertises volume discounts:
Volume discounts may be available for high-volume users. These are negotiated on a case-by-case basis.
* Standard tiers use the pricing shown in Model pricing
* Enterprise customers can contact sales for custom pricing
And there are discounts available through "Claude Platform on AWS":
Anthropic rates your token usage in USD at standard per-model, per-feature rates, applies any negotiated discount, converts the result to CCUs at $0.01 per CCU, and reports the CCU quantity to AWS Marketplace hourly. Your AWS bill shows a single CCU line item.
https://platform.claude.com/docs/en/about-claude/pricing
On the other hand, contrary to Anthropic's documentation, another source claims they've killed pre-existing API volume discounts for large enterprise customers as of April: https://itbrief.news/story/anthropic-shifts-enterprise-billi...
> > The parent there may be making the same assumption I am, that large enterprise _never_ pays sticker price.
> I shared that assumption until yesterday, when I found out that it wasn't holding for LLM pricing from OpenAI and Anthropic.
This reads like GP saying "enterprise never pays sticker price" and you responding "I thought so too until I saw the sticker price".
Is there some info you have that you can't/didn't share? Your article doesn't offer anything beyond the above.
You'd have to buy a subscription to The Information, but this is useful: https://www.theinformation.com/articles/anthropic-changes-pr...
> With the pricing change, customers of Claude Enterprise, a two-year-old bundle of products meant for large companies that now includes Claude Code and its work assistant, Claude Cowork, will have to pay for the amount of computing capacity they consume while using the software on top of a monthly flat fee of $20 per user, an Anthropic spokesperson confirmed.
There was a Hacker News thread the other day where a bunch of people confirmed that their organizations had seen this too: https://news.ycombinator.com/item?id=48278610#48280906
The sentence you quoted doesn't say anything about the token price for enterprises matching the sticker price. It just says that enterprises are paying a consumption based price (presumably billed as some $/token), on top of a per-seat fee.
large enterprises dont pay openai or anthropic, they get this thing called copilot and get a nice price there. At least on this side of the pond (eu)
> I found out that it wasn't holding for LLM pricing
You're correct. When a type of cloud service grows large enough and has a few competitive suppliers, enterprise pricing tends to coalesce with the large buyers paying around the same price for the same thing. While that might be lower than the publicly cited rate card, the private price similarly large customers pay ends up being similar.
One reason is that the very largest, long-term enterprise customers are so valuable, they can command MFN clauses ensuring no one else is paying substantially less for the same thing, then the rest of the rate card for smaller customers flows down from that. There's a strong disincentive for vendors to cheat or allow big disparities between similar classes of customers because the number of people involved on both sides of these deals is large enough that word will get around eventually.
Large scale enterprise sales and purchasing in a given sector tends to be rather circular. Account execs move to other vendors and call on the same customers, while purchasing agents can move to other customer firms. Personal relationships, reputation and credibility matter. Lying to screw a large customer over just to make one commission or quarterly quota can be a very bad long-term career move. Sometimes purchasing agents or executives quietly compare notes off-the-record with their peers from other similarly-sized firms. After dinner drinks at industry association meetings and trade shows can be quite productive in terms of verbally exchanging 'market insight' with peers.
When there are significant pricing differences, its usually due to different volume commitments, SLA/QoS guarantees, payment terms and other material factors which justify the difference. Source: been there, done that inside a top ten valley tech company. Once was in a meeting where a newly minted EVP tried to get a long-time senior account exec to pressure a huge customer by being semi-dishonest. The account exec schooled the EVP on the fact that the EVP could only make him unemployed for about 8 hours but that huge customer not wanting to work with him could make him unemployed forever. :-)
I do know of moderate-size companies deploying OSS LLMs on their own GPU clusters, for ownership/security/maybe cost reasons. I'm somewhat surprised F500 companies are apparently just handing over all their data to the model providers.
Could be fantastic for small shops while it lasts. The big guys have to pay 10x for precious tokens.
I guess F500 companies’ AI contracts include data protections.
And "large" just means that AWS will assign an account manager to talk with you. I was at a start-up who spent $300k/year on AWS and that was enough to get special attention and discounts. Enterprise pricing is confusing.
The point is that those a real prices real people are paying for real API usage. it's not made up.
your point is large players won't pay those prices at massive volume. ok
They pay sticker price. There may be exceptions for very very large companies like Amazon or Microsoft which have their own deals where they rent out compute in return for usage.
Claude is so in demand at the moment that there aren't really volume discounts. Anthropic sets the terms and you either accept them or get lost they have that much of a lead (mindshare/desirability wise).
Lets say McDonalds charges $2000 for a BigMac. If they offer a deal and sell it to you for $200, did you save $1800?
Maybe if you spend $2000 on a BigMac. But it’s unlikely you would buy such a burger.
What is a hamburger worth? Don’t look to McDonalds to set the value.
> If I had been paying API pricing it would have been $2,180.16
The point being made above is that API pricing is calculated... somehow... seemingly arbitrarily. Possibly untethered to the infrastructure costs entirely: which would be the basis of any 'value', however that holds the labor theory of value, which isn't accurate either. So how do you accurately price these tokens at all (other than through price-discovery: which is slow, messy and fuzzy)?
> So how do you accurately price these tokens at all
Like anything else in the economy: at the point where enough customers can pay you, and not enough will go to the cheaper competition.
> at the point where enough customers can pay you
> (other than through price-discovery: which is slow, messy and fuzzy)
I notice a distinct lack of reading or comprehension (from everyone around me now, not just this comment) which worries me. I worry if LLM's are to blame. No one reads anymore...
I imagine a number of Hacker News members might be devs that haven't dealt with terms like “price discovery” before, so we should all try to show grace.
I don't really understand the holdup about that specific line? Anthropic publishes the cost per million of tokens, what's arbitrary about looking at your consumption, and doing the math?
Or are you saying that Anthropic is determining that cost per million arbitrarily?
If so, it'd be like asking to explain why things fall on the ground other than through gravity. Companies pick the price they think the market will bear, and adjust based on new information.
API pricing drops DRAMATICALLY in enterprise agreements.
As with pretty much anything priced on volume/usage.
Enterprise deals are negotiated ad-hoc, the listed pricing is simply a jumping off point for the final negotiated discount.
If you’re going to give 20,000 employees Claude code you are not going to be spending $1B per year on Anthropic tokens as if you gave everyone an individual API key. Just as Anthropic isn’t paying AWS SES $10,000,000 to send 1 email update to their massive user base when the next Claude version drops.
This isn't true at the moment, though. So far there hasn't been the negotiating power. What happens is you end up capping usage for employees at a fixed amount. I think eventually, prices will come down and there will be discounts, but for enterprise accounts at least of our size (<5000), we're paying almost 100% retail, which kind of sucks, because it's expensive, and pretty easy to burn $50 to $100+ in a day, if you're not careful. In fact we got pushed off the former plan to the token-utility one at the last contract negotiation.
Going to be interesting to determing the metrics we give to engineers for determining whether the spend on this is worth it. Measuring PRs, lines of code committed, commits fully generated by agentic workflows, etc.....
As someone who has seen the enterprise deals, they are not subsidised in any way shape or form. Anthropic may wave the seat fee, maybe get lower "expected" consumption. This changes what the company pays up front. but token prizing is fixed.
To confirm, you're saying the enterprise deals you've seen match the prices per token listed here?
https://platform.claude.com/docs/en/about-claude/pricing
> API pricing drops DRAMATICALLY in enterprise agreements
Do you have any numbers or reports to back that up?
> Just as Anthropic isn’t paying AWS SES $10,000,000 to send 1 email update
How much do you think emails cost? That number is just so far off?
But besides that, running SES is also quite a bit cheaper than SOTA ai models with high demand (and comparatively) no competition. And quite a bit more pressure to make money (soon).
I think it was a figurative example. For what it's worth, $10,000,000 buys you 100 billion (1e11) outbound emails on SES at the sticker price ($0.10/1000 emails). One source puts the number of emails sent worldwide in 2024 at 132 trillion (1.32e14).
https://aws.amazon.com/ses/pricing/
https://www.statista.com/statistics/456500/daily-number-of-e...
But if their “figurative example” was the opposite of reality, then we should similarly assume their other claims were the opposite of true as well?
Have you or I misunderstood the "teams" plan?
edit: I missed the "enterprise" feature matrix with the usual audit/compliance stuff to force the biggest enterprise customers onto enterprise plans. Otherwise the "teams" plan is much better value for any business.
orig-continued:
https://claude.com/pricing/team
Teams premium is "Everything in standard, plus more usage*"
And from my experience, it's a very generous usage, I've only hit the limits once or twice, and both times required multi-boxing agents.
I could single-window agentic development all day on opus-4.7 auto-mode without hitting limits.
If you're a business using claude, then that seems like the right plan, the enteprise/API plan seems more suited to where your product is built on top of the agent themselves, so seats/limits aren't really meaningful?
Claude Teams and Claude Enterprise are 2 distinct plans. Simon is right that Enterprise seats have no included usage (and so all usage is charged at API billing rates), whereas Teams seats do.
Tokens do have a clearly calculable intrinsic cost. There's the marginal cost of production (i.e. the inference cost) and the amortized R&D cost that goes into the model producing them.
Yes, value is hard to calculate, but luckily market pricing mechanisms exist exactly for this purpose. There isn't a better number to use than what people are willing to pay for them.
So he's saying that on an enterprise plan, he'd be spending $2,180.16. He's not paying that much, but enterprises are.
There’s a cost, certainly. I expect Anthropic knows. But we don’t know what that is.
Enterprises also use 200$ plans, they are not that stupid. They add a spill over API key for over usage.
OpenAI and Anthropic won't let them any more. Once you get above a certain size (I believe 150 seats) they push you onto the enterprise plans.
I don't believe you have the option to keep with the $200/month flat rate subscriptions any more. I'd be happy to be convinced otherwise.
(I dug into this a bit more and couldn't find anything in their consumer terms that say "you cannot use this personal account if your company has more than X people", so I imagine the pressure is more that your big company's purchasing department really doesn't like managing hundreds of individual subscriptions as opposed to a single, stable, predictable negotiated contract.)
The purchasing department _really_ doesn't like it, but it's the corpsec department that will really want to murder you for using "personal" accounts to interact with corporate codebases.
> “Tokens” don’t have an intrisic cost or value.
i am pretty sure these services know what it truly costs them to serve you tokens, maybe not in realtime but at least periodically.
however, what they charge us is a constant exercise in price discovery. i agree with this sentiment in the sense that we don't have a stable sense of the cost. all of these comparisons are good for the moment, or at most the near future.
i believe that even the "all you can eat" approach with the max plans, regardless of their crazy pricing, is not sustainable only with the power users. if most of us gets this kind of value through our plans, surely it does not incentivise the service providers to continue pushing it. maybe they can regardless just to gain market share, but not forever.
From my back of the envelope analysis for my own projects, paying per token on OpenRouter is competitive if not cheaper than running the same open weight model on a rented GPU. Per-token pricing is in the same ballpark (although more expensive) for closed frontier models and open weight models (cents to dollars per million). To me this says that the pricing is somewhat grounded in reality.
Are you comparing single-user requests or multiple concurrent requests when you say comparable to rented GPU? Most of the cost efficiencies kick in with concurrent/batch requests. A single H100 node can provide like 5k input + 2k output tok/s on a model like Qwen 3.6 35B-A3B with 30+ concurrent requests.
a little critical thinking led me to read that sentence as $2180 worth of tokens [at current api pricing]
Reminds me of a car dealership talking about how good of a deal their extended maintenance plan and warranties are.
His point is more he was surprised enterprises weren't getting the discount. And so indeed maybe it is not a giant ponzi after all! (Could be a bubble)
I think it's funnier that you can believe some things have an intrinsic cost and others don't
Lol. They obviously have intrinsic cost, the floor being the cost of electricity. It’s hilarious how we are throwing critical thinking out the window when it comes to evaluating biased sources of info.
I have to give them kudos. This whole thing is the greatest swindle of all time.
AI has some use cases, but not at the price it’s currently priced at. I’ve been on AI since GPT-2 with a lot of heavy users. Every user has the same story, curiosity, surprise, hype, hate, realization. Enterprise is usually a bit behind and are right now at hype cycle, that’s where they sold all the deals and do the IPO.
It’s really a VC masterclass.
Don’t get me wrong there is are useful cases of AI, but not the way the want it to be. Quite similar to Blockchain. The idea of decentralized money has right to exist. 99% of other coins not.
AI is a faster, but still less accurate search engine. AI is great in finding bugs, it’s great at ruber duck debugging.
The reason I call it a swindle is, because along with the marketing it gives tons of people in the world the impression, they can now build their own startup, game, infra etc without the need to learn it themselves. This leads to millions of abandoned and low qualiy projects and products, because the vast majority has never built the mental modal necessary to solve the problem thoroughly. In the end they’ve wasted months and money (but burnt tokens). This is what I call a swindle.
All early adaptors I know have not drastically winded down their usage, not because of money, but because there is no new case. If you want to explore a new project you can get onboarded quickly learn a lot and then switch to documentation and live testing. For me usage is the lowest it has been the last 2 years.
I would not let AI touch my code. I have anxiety around it, because it will gripple back up. I will let it read my code and let me know what I did wrong so I’m sharpening myself.
100s of companies including open source solution can offer that for me.
All my non-tech friends are now in hype cycle and share their hype and fore forseeable frustration with me.
I have to say I’m in a way impressed in how AI has been rigorously vc-utilized (conciously or not-conciously) to generate these vast companies with the whole world watching.
Trying to parse the raw claims here maybe you can help me out
- it’s a swindle because ROI of tokens for coding models is not positive? As in it doesn’t bring enough value to charge like the $100/mo?
- enterprise customers are too dumb to see this
- IPO to max out the CEO profits for what is ultimately blockchain vaporware
Am I getting that right? Or am I putting words in your mouth?
> it gives tons of people in the world the impression, they can now build their own startup, game, infra etc without the need to learn it themselves.
I can’t speak for peoples beliefs and motivations, but this seems to be strawmanning, no? AI is a powerful tool to force multiply people. You can’t just prompt “build me an enterprise SaaS app worth $1B” or “build me GTA6 and don’t hallucinate” but is that your impression of what’s happening? Dario and Sam are saying “if you buy our coding agent subscription you can build a game with zero skill and one shot and then be rich”?
If you don’t find value in AI agents I can see reasons why that could easily be true today. Also if it just gives you the heebie-jeebies. But to say it’s a swindle on par with the blockchain I think that contradicts an enormous amount of signals and also the actual dialogue (not just headline sound bytes) around what these systems are capable of today and what we expect them to do say at the end of the year.
You kind of got it right, but the biggest loser of them all are the investors especially the index investors. They don’t even decide what they invest in but the savings that goes in funds need to invest in these stocks.
It’s quite an elaborate swindle obviously. But you generate hype with underselling your core product, you claim way more usability then there is. Users will experience usability initially. Everything multiplies with each other and then you put it on the market. Everybody involved makes money and you’ve succesfully extracted money from everyone who’s invested in NASDAQ index funds at the very least.
> Dario and Sam are saying “if you buy our coding agent subscription you can build a game with zero skill and one shot and then be rich”?
That’s Anthropics marketing, yes.
Also their offering is not uniqe that justifies a 1 trillion valuation. The first companies are already rowing back. It’s a really certain time window that they are about to hit now with their IPOs
The companies that have signed these enterprise deals haven’t done a ROI analysis. They had Fomo.
> you generate hype with underselling your core product, you claim there’s way more usability than there is
Isn’t this a contradiction?
> Everything multiplies with each other and then you put it on the market. Everybody involved makes money and you’ve succesfully extracted money from everyone who’s invested in NASDAQ index funds at the very least.
Sorry I may have totally missed what you’re saying here. Anyone in S&P has already made a lot of money thanks to effects from these companies. No one has to invest in an index fund. Markets have risk…
> That’s Anthropics marketing, yes.
Show me.
> Also their offering is not uniqe that justifies a 1 trillion valuation. The first companies are already rowing back.
That there is competition doesn’t imply they aren’t worth 1 trillion.
> The companies that have signed these enterprise deals haven’t done a ROI analysis. They had Fomo.
Also…wrong. I have seen the data at my company, everyone at scale tracks this.
> Dario and Sam are saying “if you buy our coding agent subscription you can build a game with zero skill and one shot and then be rich”?
They are saying profitable companies should replace the engineers that built their systems with a subscription (while they are hiring).
Can you point to any quotes or claims? I don’t think this is accurate
I don’t remember ever hearing Dario or Sam recommend replacing people. Rather they say that smaller groups of people can do more work, so hiring will slow because small teams can do more.
The only times when people talk about actual full replacement of people is always when they are talking about some “future AGI” that is far more capable than the tools we have today.
I usually agree with Simon but I think he is overlooking an important factor.
There is a lot of AI usage happening not because it shows benefits, but because the business has mandated its ubiquitous use. Companies having dashboards for token usage and rewarding people for using more tokens is a real thing. I just spoke with someone today who works at Microsoft and they are required to use AI for all of their work - they have to make a special request with justification if they decide not to use AI for even a single PR. This kind of demand isn't driven by value from either the company itself or from its workers; it is the kind of artificial demand you get from make-work projects to keep people employed during hard times.
We have to wait for the hype to settle down and people start making business decisions based on results before we can really value these AI products.
Correct the question is one of real vs fake demand curves.
I think token leaderboards are idiotic, however… many companies are requiring AI use, with by far the most likely reason being that they truly believe AI-assisted development gets (or will after the time needed to get employees acclimated) better returns.
Anthropic isn't actually profitable from what I'm reading, a discount briefly pushed them into the black. This guy makes the case well: https://www.wheresyoured.at/anthropics-profitability-swindle...
I'm skeptical that their current price raise is sufficient, and I'm also skeptical that most users/businesses will accept more significant price raises that will be needed. Especially for individual users, $200 a month is already incredibly expensive, I really don't think most people are going to be willing to pay more like $1000 a month.
I think we could eventually get to a point where $1000/m+ consumption would be considered a bargain.
I currently pay ~$150/m in tokens to cover my 1099 work. If I had to pay $1000 for those same tokens each month, I would still be massively positive on my margins. I bill customers based upon # of completed features & bugs, not total time spent at the computer. Tokens would have to more than 10x in cost before I would start to have a problem on my end.
How are you billing clients when using AI? Do you pass on costs or do you multiply time saved by your hourly rate and bill that?
this - I'm wondering why this isn't getting picked up more in the analysis? Its an accounting trick, which is not sustainable. Time will tell.
The real timing is that we don't have strong enough new business needs for now and we have accumulated enough tech assets, so our work has been increasingly incremental. That means we can build reliable features on top of vast amount of past work - where AI really shines. So, with or without AI, companied would hire fewer software engineers if majority of our work is incremental: add a feature here, fix a bug there, tweak a configuration and etc, then we wouldn't need as many software engineers anyway. AI just accelerated such squeeze.
In contrast, imagine if we had the same AI 20 years or so ago. Could AI really write Jersey? I guess not as people were still trying to understand JAX-RS. Could AI really answer all the questions about React? I guess not as React was just invented. Would we use 10x fewer people to build out infra on the public cloud or the entire so-called Big Data platforms? I guess not, as they were still rapidly evolving and we'd need so many engineers to explore so many different possibilities? Could we use AI to build our ML ecosystem with 10X fewer people? I highly doubt so. Heck, 20 years ago R was all the rage and Python's ecosystem was not mature at all. Oh, and mobile computing, could AI lead to 10X fewer people to build all the mobile apps and the underlying infra?
> Could AI really answer all the questions about React? yes, due to ICL
> Would we use 10x fewer people to build out infra on the public cloud or the entire so-called Big Data platforms?
No, cannot solve core problems, makes a mess at scale
You are right about the incremental work. But most of the work is historically incremental imo, only few positions are R&D.
ICL = In-Context Learning where LLMs learn to perform new tasks by reading examples directly within the prompt, rather than requiring permanent model retraining.
If those companies found PMF we should expect to see the effects of all those tokens burned on the products we all know and use. In my own experience: yes new features can be delivered much more quickly (by orders of magnitude) but knowing what to build and how to evolve a product with taste and direction does not; so ultimately the user-facing gains are marginal.
I suspect that once the technology has been tamed and the hardware and software has been commoditized, the impact will be much less dramatic than we expect and we will realize the importance of a shared vision, experience, taste, intuition and discernment in building good products.
this is an opinion and i would love to debate :) dont hate me too hard
i think the article is momentaneously correct but there are some things that smells to me about the situation overall (not the article!)
(1) i believe gpt 5.5 and opus 4.7 are not as good improvements as their predecessors and there has been not enough evolution in those models to justify the price increases. Unless something big changes in the next few years they will not be able to track these costs (unless A)
(2) they might not be able to keep up improving the data centers if their tech keeps demanding more and more hardware capability. even nvidia does not show signals that they will be able to beat too much their big GPUs and at some point this increased price will be passed to them anyway which will need to be put in the token price (unless A)
(3) i've been trying deepseek v4 and other models and honestly, they are more useful than gpt 5.5 and opus 4.7... i mean, there is still a difference, but it is so little that it does not make sense the cost of opus and gpt 5.5 (unless they are going to A too)
(A) I have the impression is that the plan was the whole time to sideline common folks like us and focus on gov, big techs and military. they only let us play with the toys to gather data and for people to not get mad because they stole the internet from us.
The costs are exorbitant and most software is not produced by companies with such a huge moat. Anthropic made a profit through their recent bait amd switch pricing. There is zero useful insights online to indicate whether this might die due to commoditisation with good enough open models or fail the race to get more people subsidising unsustainable growth with other people’s money. Who knows? In any case they dont seem to be able to drop usage costs so the business model seems based on wishes
Continuing with your skepticism:
> Stories are circulating of companies surprised at how expensive their LLM bills are becoming from usage by their staff
> Enterprise customers are now paying API prices
How long before enterprise customers start to question the bill? Anthropic goes from not making money to doing pricing shakeup, and now they are making money and the biggest spenders are shocked at prices.
Seems like things are still very uncertain.
And switching cost is basically zero, so there's not much friction. In my harness it's literally just saying model=X
Usage costs will come down with better hardware. Hardware is improving rapidly each generation.
Costs will plummet as better hardware becomes available and priced reasonable so that people can more easily run their own open models locally. But that won't help Antropic/OpenAI make more money, quite the opposite.
That trend held true for the past three years, but it doesn't feel as safe to me now.
But memory costs are going way up. And both OpenAI and Anthropic bumped up the price of their frontier models in April.
Yeah, it’s called supply and demand. Demand for memory went way up suddenly. Now supply is going up rapidly as companies try to cash in on that demand.
Supply will eventually catch up with demand. Then the prices will come back down.
Algorithms are also improving. I believe it's very unlikely for these two improvements together to not result in one to two orders of magnitude cheaper cost per "intelligence". Of course, that might just make use cases that are too expensive today viable and thereby increase usage further.
A lot of the new hardware requires retrofitting existing datacenters for appropriate cooling, or is waiting to be installed because the new datacenters haven't been built yet. By the time they're installed it's likely a lot of Blackwell GPUs are going to be very out of date. Newer hardware is turning into huge capex bills along with the corresponding depreciation costs. Basically, it's not the same as plugging a new GPU into your desktop, the upfront investment is extremely expensive and all the numbers I'm seeing suggest that the newer GPUs are costing more to run, not less.
Sure, with all the component shortages it’s not surprising the current GPUs are coming at a massive premium.
Eventually either the supply will go up or companies will start buying fewer overpriced GPUs.
Either way, the price per token will come down as hardware improves and supply and demand reach equilibrium.
Technological obsolesence is a bitch eh.
200$ per month per seat is nothing .
A single 3D CAD license pack for the guys in our R&D group costs multiple thousands of dollars per seat, per month.
It's about time software seats get some love too.
AutoCAD is $175 per user per month [1].
[1] https://www.autodesk.com/products/autocad/buy
AutoCAD is still the budget-friendly CAD program it has always been. You don't build big boats in AutoCAD.
Winch Design [0], which have built some of the world's largest superyachts [1], seem to be using AutoCad. [2] Afaik it's also the same with Lürssen (but don't quote me on that)
[0] https://winchdesign.com/ [1] https://www.superyachts.com/directory/1516/winch-design/flee... [2] https://www.autodesk.com/design-make/articles/naval-architec...
Likely not the "base model" of AutoCAD.
Those tools are used in ways that they're integral to processes. They have their equivalents of ticket systems that are linked to code repositories with LFSs and bunch of IDE type tools and automated and manual test systems and build systems. Their equivalents of PR discussions and Selenium screenshots needs to check all boxes in the right ways for legal and traceability purposes.
Without all that might be $175/user/month but you're not shipping apps with just vi and bare gcc.
>Without all that might be $175/user/month but you're not shipping apps with just vi and bare gcc.
You're right, Linus uses Emacs.
Except LLM's even with Vision are still useless at AutoCAD let alone Revit (please dont quote SCAD LLM's at me, useless). Knowledge based approaches still win.
I might agree "AutoCAD" is the current level LLM's are at, but wait until your design departments discovers "Revit", its another ballpark (in wasted cots, engineers on site still get "clashes").
Revit costs are high, and the end results are marginally better - but local LLM's tokens are cheaper 24/7 at "AutoCAD" level - "Revit" level tokens will make Ubers CTO/COO weep harder than they already do. While producing results no better than "Revit" does (engineers still face "clashes").
As someone completely outside the 3D design world who always thought of AutoCAD as the gold standard - really? What program would be used instead? Please enlighten me.
Cadence and Ansys have entered the chat. A bunch of other highly-specialized engineering software has entered the chat. Licenses are on the order of 10-100k/seat.
For a pretty funny comment about pricing.
https://www.reddit.com/r/chipdesign/comments/1ajrli2/cadence...
Glad to run into this after some time!
I guess we are welcoming the software people to the world of expensive tools. Just sad that the FOSS alternatives of these tools are not as powerful whereas software industry still has FOSS tools to fall back on.
Yeah, it’s nothing, and it’s also not the cost that enterprises are paying. As the article states, the price is $20 per seat per month, PLUS per-token API usage. Enterprises are paying consumption billing, not fixed rate oversubscribed “all you can eat per seat.”
CATIA licenses which are the most expensive I've seen are roughly $600/month per user. Where are you seeing "thousands of dollars per seat"?
CATIA with plugins can go up to 100k a year. That’s what we currently pay
Wouldn't each plugin be a different piece of 'software'?
100k per seat? That's crazy. How do you even hire or train employees with software that expensive?
One guy retire and a college graduate goes in.
CFD might reasonably be considered part of CAD and something like ansys costs about as much as catia. Still only doubles it though.
Sure. Is CAD going to be used by every working human?
Now add up the engineer’s salary and you’ll find that software seats already cost more than those R&D ones.
But when previously your software developer tools were free, that's a huge increase
How many guys is that? Every single white collar worker is in the AI ICP (customer profile).
edit: typo
white collar*, not color
What does ICP mean?
Insane Clown Posse, though given the context here probably Ideal Customer Profile.
The similarities are quite stunning, though, as I'm sure both sets of ICPs have no idea how LLMs work.
Now hold on there, let's not cast doubt on ICP. I'm sure they'll surprise us, as they always have.
> That’s $2,180.16 worth of tokens for $200
So the author claims he's getting $2000 per month worth of frontier AI free of charge. Ok. If he's been doing that for 6 months that's $12k. What has this produced concretely? For $12k you can find a used car in decent condition. Heck for $1200 (his actual out-of-pocket spend) you get a brand new ebike! (on which you could put a pelican and make a photo of both if that's your fancy). But here it's unclear what has come of it.
I've written a great deal of code - code that would have taken me years of work to produce without LLMs.
(It's mostly open source, you're welcome to dig around in https://github.com/simonw and https://github.com/datasette if you like.)
My time as an experienced software engineer is worth a lot of money - a whole lot more than $12,000 for the past six months.
> code that would have taken me years of work to produce without LLMs
As you might suspect, this is what I have an issue with. Without LLMs, isn't it possible or even likely that that code wouldn't have been written at all, and wouldn't have been missed? If LLMs are mostly used to produce throwaway prototypes then it's a stretch to say that's money well spent.
If indeed it let you advance your main product much faster then sure it's a different story. You're the judge of that. It's hard to see the impact from the consumer side; everything is still broken and no extraordinary app seems to be emerging. Maybe it's just a question of time. We'll see.
I've thought about this a lot. I am very confident that the way I use LLMs is both accelerating progress on my core projects (here's a substantial, reviewed PR I landed just yesterday https://github.com/simonw/datasette/pull/2741) and helping me create plenty of projects that otherwise would not have existed.
The point being made by GP was that your projects have no value and their non-existence wouldn't be a negative to this world.
And that is likely a fair assessment, though I understand perfectly the feeling that you have that you are accomplishing great(er) things thanks to AI.
I was certainly not saying that all the author's projects, in general, have no value! That would be rude, mean and most of all, incorrect.
But yes, it's likely that the ease of which code can now be outout lets us produce lots of unnecessary code just because we can, and the author says as much in a below comment
https://news.ycombinator.com/item?id=48303890
I certainly hope that's not true, given that I've dedicated 7+ years to my main open source projects at this point.
I take some reassurance from knowing that they are indeed used by real people to solve real problems though.
Now what percentage of the 200$ have been used on the useful stuff and how much on exploration or other stuff.
How long would it taken you to do it yourself? How much longer will the next task take you, compared to when you would’ve written the code yourself. How is the mental model compared to when you would’ve written it yourself?.
I’m not saying you’re wrong, again there are use cases. But the calculation is not plain and simple it goes deep into our perception, perceived productivity versus actual productivity.
I’ve 2 months maxxed out all 6k of Claude Code and bought Antigravity on top. My codebase became 140k lines. I introduced tons of bugs and spent another 2 months, deleting 80k of code. I wish I would’ve just chatted with AI and not let agents touch my codebase. I would’ve saved approx 300$ subscription prices a month and 2 months of my life.
I had that experience back in January! I used it for a joke in a talk recently... did anyone really need a half-finished buggy slow implementation of a JavaScript interpreter in Python? They did not. https://github.com/simonw/micro-javascript
It's taken me the best part of this year to readjust and find a pace and level of ambition that fits.
It was challenging for me as well, but my pace and level is now that I use it for these cases at the moment:
New language, infrastructure, general level of understanding of something I barely have an idea of.
Rubber duck debugging, if i dont know the correct solution
Checking my code for issues and bugs.
But not for:
- writing my code - agentic coding (help me)
The inference has reduced drastically. It’s basically just chatting. I don’t let it write anything, but sometimes I purposely use the browser window instead of them sitting in my codebase, because I know it gets things subtly wrong and migth focus on the wrong things.
The same way people used to say don’t copypaste code at least write it out I think it’s still true. It helps to buidl the mental model and to find the right abstractions.
This is the economic theory of value creation though - arguably the world is better off because new projects can be created, and they are marginally cheaper than they would have been previously
I’m watching to see what happens to big enterprise software contracts. Why pay some vendor $800k annually for something a couple mid-level devs can replace—-and tailor closely to your needs——by leveraging AI.
Open source software changed the world. AI that will cheaply write whatever you want in a few days will also change the world.
> My time as an experienced software engineer is worth a lot of money - a whole lot more than $12,000 for the past six months
From this I assume you think that what the llm has generated is as valuable as your own work generally is. How do you even calculate this?
And what was your return on investment?
As I commented elsewhere, I'm still bad at making money from my open source work: https://news.ycombinator.com/item?id=48296794#48298909
(I have a feeling if I could say "and I closed $2m in sales with the software I wrote!" people would find a way to say that didn't mean anything anyway, because how can I prove I wouldn't have made those sales writing it by hand?)
> "and I closed $2m in sales with the software I wrote!"
Given the audience you are reaching, that is actually the expectation. Github stars is not a great metric.
I would be very curious what kind of answer would satisfy you here. Simon isn't building a product, where $200 is a line item on a balance sheet. If he tells you what sort of analyses or time savings $200/mo on coding agents have enabled him, do you honestly think that would satisfy you?
The product-market fit question feels premature when most usage is still developers building on top of the APIs, not end users paying directly.
How is the lack of bad news declaring a victory for AI? I am yet to see any company concretely publish analysis about the ROI from AI. Most companies as far as I know are still treating AI investment as sunk cost with no expectation of returns at the moment. We could very well see a world where companies heavily scale back investment.
I may be biased because I work on an AI powered enterprise productivity product, but while I agree they have PMF right now, I wonder whether people's use will evolve in ways that undercut the current PMF. Chatting with an assistant is great if there's no product with tailored UI available that also has the AI capabilities. But once there is, I suspect people may switch, or more importantly enterprises may switch because they'll get the benefits of AI without the clunkiness of a chat only interface. We may see another DOS -> GUI-like shift.
More specialized products will consume tokens but their builders will be incented to optimize token use and switch models as costs and capabilities change. And if search engines become more AI capable, and Google is clearly striving for this, then they may have pressure from two sides that could squeeze the number of use cases for AI chat. AI coding isn't going anywhere and nor is the need for AI in general but I wonder if the products will have to evolve significantly to maintain the current levels of PMF. And then there's the question of profitability...
MCP-Apps could be the solution for tailored UI.
With deepseek and xiaomi mimo models slashing their prices 99%, I don't see a great future for openai / antrhopic with regards to their 1T valuations. Maybe 1T valuation will be the whole market, West + East.
Most of the corporate world in the EU or North America will be hesitant to rely on Chinese AI providers. There are some very real blockers for that for things like data security, compliance, etc. And recent geopolitics don't help.
Legalities aside, you need to look not at the model quality but at the infrastructure needed to scale these models from tens (now) to hundreds (soon) of millions of users. Only a handful of companies actually have the resources and funding to do that. That's what these huge valuations are based on. These companies are gearing up to scale to these levels. That's why they are spending on data centers. Whoever has access to those data centers gets to tap into the revenue stream of people using models running on those.
The market for frontier models is roughly split between OpenAI, Anthropic, and Google. And then you have companies like X/SpaceX, Amazon, and Microsoft being more successful with their infrastructure than their AI products and companies like Apple, Meta that have the money and the aspiration but are so far not really managing to be very successful with their AI strategies.
Deepseek is just very poorly positioned to capture a lot of the enterprise revenue in the EU or North America. But they might become very dominant outside the US/EU. And of course China itself is going to be a huge market and equally unlikely to want to be depending on US owner AI companies.
Deepseek and all the other Chinese models have open-weights. You can host them yourself, no need to send data to China or rely on them.
There is still a risk of supply-chain attack. People give LLMs direct access to their entire infrastructure via tools, and never check the code produced. It's not difficult to steer an LLM during training so that they'd output malware only when prompted a certain way, and that wouldn't come up during the initial evaluation.
Personally I see no difference between China and America in terms of risks of them embedding "backdoors" so to speak, but I disagree when people claim that open-weight models are obviously safe just because they can be ran locally.
> It's not difficult to steer an LLM during training so that they'd output malware only when prompted a certain way
Perhaps, but that's also a good way to lose users+reputation as there's no way to control when said malware is generated. Once the first instance is discovered cybersec researchers will have a field day reproducing it and showing the world.
It is not a trivial challenge setting up model serving infra for ~1T or larger models, especially in a high reliability environment (e.g. your team is using it for work, or you're using it to power production apps). Sure, there are third party providers, although the quality and reliability of their inference varies.
Run it on Bedrock. If you're already on AWS, procurement doesn't even need to be involved.
Run Deepseek on Deepinfra then? Or Fireworks if US-based is important. None of these are real issues outside maybe convincing your legal team to do a bit of homework.
I don't think you are appreciating the physical constraints here. Deepseek doesn't really have the hardware in the US or EU to do anything at scale.
Sure, you can self host a non-frontier OSS model yourself; including Deepseek. And no doubt some people will pay one of the companies I mentioned to rent the infrastructure to do exactly that. Much of the rest of the world will be paying directly for direct access to the frontier models.
As for the legal/compliance stuff, I recommend you don't take any big decisions on that front without consulting lawyers. My understanding of that is that most serious companies in the EU have to take these topics pretty seriously. I'm sure in the US, hosting all your data and secrets in Chinese data centers isn't a whole lot less controversial.
The Chinese could of course choose try to match the current levels of investment Google, OpenAI, Anthropic, etc. are putting into local infrastructure. But as far as I know they aren't and there are probably a few political blockers for that.
Without infrastructure, their role is being a niche player in these markets. It doesn't really matter how good they are if they can't scale to most of the market.
I mean, Western providers such as Fireworks AI/Microsoft Foundry (US) or Tensorix (EU) already are offering many of these models on their own hardware with all the typical compliance boxes ticked through a standard API. Either as open weight models or through partnerships with Chinese firms, or both. DeepSeek etc. do not have to do anything here other than making their models available to Western partners (either as open weights or through a licensing agreement).
They'll still have their dedicated enterprise customers. I think the Chinese providers will pull more of the single users who're paying their own way, than those backed by company budget. And it's a pretty good split as the demand becomes better distributed, resulting in better service (I'll never forgot must how bad access to Claude became until they got access to Colossus) and less potential for lock-in (we really don't want there to be a duopoly, etc on good AI).
Just to be clear, enterprise companies are using the enterprise option for security/auditing/enforcement reasons, not because OpenAI or Anthropic forbids them from buying a 200$/month plan for each engineer, right? Do they have any clause forbidding using those individual plans in a company?
> I currently subscribe to the $100/month Max plan from Anthropic and the $100/month Pro plan from OpenAI. If you are a heavy user of coding agents these plans are a fantastic deal.
Guys, what - in your opinion - does "heavy user" mean? I thought I am heavy user (I am using AI to code every day 8hr a day + side projects) but 20 USD/month Cursor plan is always enough. What should I be doing to extend my license to higher level?
I'm out of credits codex pro credits (200 plan) in 4-5 days, and then I switch to claude (another large plan) mostly or just go out more and get Vitamin D. :)
I spend most of my time designing and tweaking tests suites, and improving test performance. These commits are almost entirely Codex: https://github.com/tsoniclang/tsonic/commits/main/ - but it's possible only because there's a very large test suite attached to it.
All of that is very token intensive. If OpenAI gave me 3x my limits, I'd find ways to eat it up in a week.
What do these tokens give me? Well, I think in a week or two, I hope to port the TypeScript-Go compiler back into TypeScript, but compiled to native code. It's probably not particularly useful for most ppl, but it's a hobby project that I've spent the last 7 months on.
My agents are easily busy 30 min to a hour independently. Implementing their plan, building and running tests, verifying deployment, linting. So I switch between multiple agents. Each agent has its own branch and worktree.
How many sessions are you running simultaneously, how careful are you being about token consumption and context window management, and what level of reasoning are you using (with which models?)
8hrs a day doesn’t really mean anything without a lot additional qualifiers.
fwiw lately I’ve been straddling 2 or 3 claude codes and one Claude cowork, primarily on 4.7 with high effort - the company’s paying for it, so I’m doing my best to burn as many tokens as I have the mental capacity to manage. At that rate, the 100 account is completely necessary, I was blowing through my 4-hour limits consistently before requisitioning an upgrade.
Ai has become indispensable but maybe not at all cost. My company just had a company-wide meeting to talk about how they're restricting who can use which models and instructing us the "be more responsible with company's tokens". And it's not an small company by any means.
> Coding agents really did change everything. These are tools which burn vastly more tokens
The assumption here is that this is a positive thing.
But this very well could end up being a major negative long term by increasing the cost per user, reducing margins.
More usage = more cost = less profit.
It's not obvious that more usage is good. It's only good if revenue per user increases more than cost does. I'm skeptical about that.
> It's only good if revenue per user increases more than cost does.
That's why it's so important for these labs that they're selling API tokens for more than the compute+energy costs needed to generate them.
Every indicator I've seen is that they do have a positive margin on that. If they don't, they're screwed.
No this is not what matters.
The customers of these tokens need to see returns on their projects that exceed the cost of financing.
Laying people off only goes so far.
If enough said firms don’t see enough value given the price of frontiers they will cancel and consume open source. This is the risk the frontier labs are exposed to.
What's an example of an indicator? Genuinely curious!
Insider tips from Google and AWS telling me that they run inference at a profit (though that was over a year ago now).
Dario telling Dwarkesh three months ago that they have a margin on inference: https://www.dwarkesh.com/p/dario-amodei-2?timestamp=3528.0
Unless your insider is the CFO, I wouldn’t trust these sources to have the access, knowledge, or insight to determine whether they’re running inference at a profit.
Simple test: can they get their hands on a data center contracts and financials?
Search isn’t anywhere near as high profile as the profitability of AI inference, and yet, even aspects of the Search org were walled off from the rest of the company such that other employees couldn’t see what we see.
I don't think the fact that inference runs at a profit is a particularly deeply kept secret inside these companies.
It might be worth keeping it secret from the staff if they were running at a loss.
Are these sources not incentivized to say exactly this, regardless of whether it's true?
The insiders tips? I don't think so, they were people I trust.
They had all the incentive in the world to say "I'm not going to talk about that."
I think the third company (likely Google) is going to make LLMs financially feasible with:
- dedicated hardware (https://cloud.google.com/tpu)
- optimized models (https://research.google/blog/turboquant-redefining-ai-effici...)
> As further evidence that enterprise agents represent product-market fit for these companies, consider their open job listings.
PMF is one interpretation, but it could also be read as desperation.
In my opinion, we've been at PMF for quite a while now. The November inflection point that's often referenced definitely changed how we interface with models, but as far as coding goes, I feel like Cursor had proven itself useful for at least a year prior to that.
The demand has always been there, the outstanding question is still - how do you build a business on top of these products? None of the frontier models have emerged as uniquely capable, but open weight models are now catching up in capability as well. The explosion in go-to-market roles feels more like an attempt to lock customers into contracts so that they don't consider alternatives.
I assume the hope is that during this 12-month contract they will develop real integrations, something deeper than just a CLI harness. If you've ever worked in procurement or dev tooling at a reasonably sized company, you'll know that this is exactly what teams try to avoid.
It's anyone's guess what will happen this time, but I'm excited to see how the IPOs go.
For an alternate albeit somewhat contrarian view, also see Ed Zitron’s piece that add context to Anthropic’s profitability: https://www.wheresyoured.at/anthropics-profitability-swindle...
TL;DR Ed argues that the deal between Anthropic and xAI could have been negotiated in such a way as to make Anthropic only appear profitable during its “ramp-up” period in June, which incidentally is also the month that Anthropic is making tons of other pricing changes.
It’s funny how a US$30 ReSharper license that would give me a 20% performance boost in programming was out of the question, while me burning US$3000 per month on tokens for a 40% boost in performance never was questioned
I feel this so much. What was all that fuss about needing 3 levels of approval for any software license? Or having to provide a compare chart against open-source alternatives?
The money would be so much better spent that way as well, supporting individual programmers.
> but I suspect there’s a more important factor here: I think they’ve finally found product-market fit
Ahhh the classic startup term that's definition is nebulous. But also, since when does any definition of product/market fit mean a product is profitable? And profitable in what sense? Unit economics? Overall company?
Oh I'm absolutely taking advantage of the fact that "product-market fit" has a bit of a nebulous meaning here.
It's a great hook to build an article around. My core point is more that April 2026 was the point when Anthropic and OpenAI finally appeared to have figured out a credible business model.
> My core point is more that April 2026 was the point when Anthropic and OpenAI finally appeared to have figured out a credible business model.
How so? What's specifically changed? We still don't know what their unit economics are and everything you've documented is basically speculation at this point.
> What's specifically changed?
1. Both Anthropic and OpenAI significantly increased the prices of their latest models. They're clearly not trying to offer the lowest-price-possible to drum up demand any more.
2. Both Anthropic and OpenAI no longer let enterprise companies buy discounted almost-all-you-can-eat subscriptions. Those big enterprises are now paying full API prices.
3. According to reasonably well-sourced leaks, Anthropic may be about to have their first profitable quarter.
And I didn't even say "profitable", I said "credible business model". I think getting companies to spend hundreds of dollars per month per seat, WITHOUT crazy subscription discounts, is a credible business model.
> 3. According to reasonably well-sourced leaks, Anthropic may be about to have their first profitable quarter.
might as well be because of this:
> [...] with capacity ramping in May and June 2026 at a reduced fee.
And then, we have
> The agreements may be terminated by either party upon 90 days’ notice.
The timing of the "leak" of profitability is just as superb for Anthropic as the timing of the $45b deal is for SpaceX.
I wonder how this partnership looks like on August 1st...
Let's see how it goes in three months. I'm not convinced it's anything but an attempt to make the numbers look better just long enough to IPO.
Enterprises will take a while to adjust. We're going along with Copilot price changes in the short term "just to see what it'd look like". They will get customers to continue with the all you can eat mindset at API pricing for a while, but the finance departments will start asking questions very soon.
I suspect it might end up being a credible business model for whoever buys the IP eventually (Microsoft, Google?), but I think OpenAI and Anthropic are so far in the hole debt wise that simply having a credible business model won't be enough to save them - it needs to be wildly successful.
Product-market fit, but what about customer retention?
It is quite trivial to switch from using one model or another. Likewise, in a few years we'll have affordable laptops to run today's frontier models.
What's their plan to let us keep subscribing?
Right now the main plan for that appears to be having those enterprise accounts commit for a year at a time.
If we define product market fit as profitable with a trillion dollar valuation, I think the term has lost its helpfulness.
I do agree with the author that these companies seem much stronger financially recently though.
Does this analysis factor in potential caching of tokens on the server side? It seems that if they organize things well (as a model provider), they can save quite a lot on that. Looking at my Cursor statistics makes it clear that the token calculations are not at all trivial.
I believe the ccusage tool I used takes cached token pricing into account.
money talks, and being a broky from outside the valley/west coast, the "product-market fit" here are the neighbours of the service providers.
the economics simply don't work unless you make six figures, at least to just give it a go blindly. the providers are also still figuring out what they can get away by charging, and they are getting a similar treatment from those under the stack.
the caps and limits are not very transparent, and it is quite difficult to know what is "enough". the current rate does not stay the same and the contract is changed way too often to dedicate for the long term. regardless, the subsidized rates should not be sustainable forever. make hay while the sun shines i suppose.
Realistically, OpenAI found product market fit with the OpenAI API playground in 2021. People were using that as ChatGPT at the time.
Anyone actually making money paying all of these monthly fees? Or just hobbyists? I have yet to see any real ROI posted anywhere.
Ask yourself: what is the ROI I provide? It isn't trivial most of the time.
We measure ROI of employees annually via ranking systems, who gets bonuses, promotions etc
Ironically, we had more new interesting things launched daily 5 years ago before AI.
This is the same question I said about people running OpenClaw. You don't hear about anymore.
Other than the hosting providers, I am also yet to see anyone directly making money from their OpenClaw agent.
This requires hard evidence which isn't available. Circumstantial evidence is all otherwise.
Bloggers are having AI psychosis too.
I thought the links I provided were pretty solid. A lot more solid then you'll see in most commentary about this stuff!
I agree with this person, let's use AI psychosis for when using an LLM gives someone psychosis, not for when we think, what, that a blogger made some poor assumptions?
https://news.ycombinator.com/item?id=48296794#48303200
If nothing else this blog did give me the idea that I should split my $200 claude max plan into two $100 CC max and $100 codex plan, esp because Claude is now offering 1.5x weekly limits so its the 5x usage is now more like 7.5x usage.
>I should split my $200 claude max plan into two $100 CC max and $100 codex plan
You may want to get one of them to check the math on that :p
It reminds me of this Steve Jobs Clip:
https://youtu.be/0lvMgMrNDlg?si=QkkOnngYTjaSPlIy
He said, so many years ago, that there will become a time where computing power is so prevalent that we will stop using the person to make the computers job easier and start using the computer to make the humans job of interfacing with it easier.
But in this context, it would mean the other side of increase productivity is decreased time to do the same work. These are the same thing.
Love how everyone boasted about replacing all the software with ChatGPT and then we end up with coding agents meaning the software engineer are STILL important. The sell is the development tool. It's classic cloud. Where did all the ops people go, many got subsumed by the cloud companies YET every company still has DevOps people to manage cloud infrastructure. The layer of abstraction went up but we still need the people to write the glue code and understand the business. OK great there's a new cash printer in the room. There's a new tool. Let's just start to ground the tooling in its new found gravity, profitability and IPO market dynamics... Reality has set in. The hype cycle is about to explode... Do you remember ride hailing and just how much cash was burned on credits pre Uber IPO. Then remember the IPO itself? These companies are not the new Google. They are a layer on top. Google was still the most efficient cash printing machine in history beyond the the US government and might still be. Will be interesting to see what the trillion dollar IPOs turn into. I'm going to say we see those prices get cut to a third in less than 5 years and scale back up over the next 15-20 years.
> The sell is the development tool.
I've been calling that out for a couple years now. LLMs best and most viable use case is still just as a dev tool. Even for non-programming tasks, I still get better results from the LLM if I instruct it to write code to do the task...look at Claude Cowork for example, it's everything I used to do with python myself. It's not really a novel capability, it's just using python & bash for automations that any sysadmin has been doing for decades. Yeah, that's valuable for a non-techincal audience but is it $1T valuable? I don't think so.
When has an IDE or other dev tool ever commanded a $1T valuation?
These things get lost in discussions because people conflate "overvalued" with "not useful." LLMs are useful, particularly as dev tool, but Anthropic & OpenAI are definitely way overvalued.
I feel like we should just ignore all LLM news until after they IPO, lots of positive sentiment astroturfing bots.
They certainly have, but it relies entirely on the assistant frame, which is a problem in and of itself for the trillion-dollar economics.
Anthropic and OpenAI have shown people want a tool for task offloading, driving predictable token consumption and justifying the math, so long as users stay in that dynamic.
However, knowledge workers using these tools daily are getting exhausted with them. Outputs come out polished but hollow. Talking to a frictionless, frame-completing model all day drains you.
If user behavior drifts away from assistant usage because of that, per-token math implodes. The valuations we're hearing about all the time rely on usage compounding daily. The fatigue is a timer running against that compound.
Anthropic's Constitution is the closest hedge out there, I think. Installing an identity structure into the model through training. But it's still assistant-first, so the fix there is only partial.
I've spent the last year running a product that flips the architecture so identity is primary and the assistant role is secondary. Same frontier models, completely different conversational quality. The fatigue property doesn't really show up.
Whichever labs figure out how to install real identity natively in the weights are going to be the ones with PMF in the next phase.
They have the whole software industry addicted to AI. They will be pushing the price continually up, and like addicts they will find a way to pay
If they’ve spent all this cash, and all we have to show for all this AI baloney are crappy coding agents that constantly make the same mistakes no matter how much you try to guide them.
All the slop content, all the bots, all the misinformation and fake AI images and videos.
All of the social and economic disruption from datacenter buildouts.
The massive nosedive in reliability on the world’s software infrastructure.
After all of that and all we get is a code bot so a few incompetent loser devs can bloviate about not writing their own code and brag about never reading it.
Burn it all to the gd ground. Destroy this new Tower of Babel.
Well, good for them that they are charging enterprises API rates. Why in the world not do something similar for consumers? Use for free a few times a day, have a $5 dollar plan for light use, and perhaps $10 to $15 for heavier use. If 90% of consumers pay nothing then ‘drop them’ except for letting them have an account and a few queries a day.
It is easy for me to change providers. Right now I use the open source Claud Code harness with two paid API venders for DeepSeek v4 (flash and Pro). I like seeing how much each session costs.
“the reality of AI right now is that it only works for coding.”
Kind of…
PMF is such a bullshit term. There’s no precise definition. It’s bs.
The fact is that investment is at a scale so large that current trajectories are nowhere near going to provide ROI.
The model companies are trying to milk every drop of ARR they can before IPOing. That is entirely the current narrative. You can hurt ARR before it churns if your goals are short term.
Anthropic being profitable is laughable. Sure, by some accounting measure that no one seriously uses. But looking at revenue and what they’re raising you can see the true story.
The truth is we need a revolutionary step up in capability to justify capex spend. It’s possible that might come - Opus 4.5 brought one - but failing that we’ll see the bubble pop once the IPO pumping is done.
And I think that's amazing. I'd like to keep using the subsidized coding tools, especially Codex, since I've given up on Claude. Hopefully the PMF allows the subsidy to continue. Would hate to have to move to the next coding harness again.
I think the reasons for them going with API pricing will become abundantly clear when the S-1s become available. If they don't have a story covering how they can get revenue closer to expenses, then they're relying on the market to believe the pixie dust version of their profitability story, which I think people increasingly don't.
I'm surprised by how good coding agents are. I think PMF was reached last year or the year before (at least for coding).
With current limits my 100/mo codex subscription is more than enough for the work I do.
However, I do worry about when does current subsidies are going to end? I can see myself paying up to 300/mo, but more than that will be prohibitive.
What's the long term plan here? Are OpenAI's and Anthropic's costs expected to increase/decrease?
"You think"?
The end game here is going back from a model where a bunch of product and tech management people sit in the U.S. or Europe, and try to manage thousands of mediocre talent sitting somewhere far away. The new model is you give those coding tools to good engineers colocated with your product people, and you ship good stuff much faster. If you can achieve such a setup, the token costs can be $50k per seat per month and you still run circles around the legacy IT models in terms of efficiency. Giving everyone the API keys and not changing the way products are managed is not going to work.
Good lord, what company would want to spend 600k per employee just to go maybe %20 faster (what the studies seem to show is a realistic estimate for productivity gains),
I'm building a product right now with some AI coding (despite my negative sentiment about AI in general they are useful). I am both the product person and the engineer, and I'm pretty decent at using it, so according to the hype I should be seeing like a 10x speedup. I am not seeing that. It's definitely faster, but there are also days where I'm stuck cleaning up things after going too fast for too long, or periods where I need to put the software in front of people to get real feedback, or even periods where I just need to use it extensively myself to find the pain points and bugs. I just don't see this "running circles" once you get past an MVP and you actually need to build something secure and not embarassingly broken.
To me the question is, can the frontier labs make the variance of output lower + make the output of higher quality to justify their prices?
If not lower priced chinese offerings will be better as its cheaper per token - giving you more attempts to offset the variance.
My feeling on the former is no... I believe they tried really hard but they've settled on pure marketing now to attempt to fight off the chinese with perceived superiority in quality.
> but as far as I can tell those credit costs are an exact match for the API token costs listed for those models.
it is only true for USD. for example if you pay in euro, this is actually more expensive. kind of makes no sense, because it translates to $1 = €1
I think it's fair to say they had achieved product-market fit when their revenues were growing deep triple digits month over month. What we're seeing now is that perhaps they have achieved profitability or at the least a more sustainable balance sheet.
I see two really good ideas for monetizing the free tier for consumers.
Firstly, if the user is asking for things where AI can link to products or services to buy, there's a very good relevancy, much higher than in other types of ads.
Secondly, since the AI often takes time to compute answers to user's questions, they could be shown ads while waiting. People could perhaps be less annoyed by this than some other commercials since they know the break has to be there anyway.
(First idea is something I came up when asking Claude to compare some products, or ask for help in lawn care. Second idea was by a colleague.)
Not sure I can agree with this. Sure, we have enterprise paying, but are we actually getting anything new and interesting made from LLMs compared to if we just didn't have them at all?
Maybe acceleration in smaller teams. We still seem in the era of the early internet where what questions LLMs change hasn't exactly emerged.
I think the product market fit is still how you interface with increasingly powerful AI. Right now is Codex/Claude code, and they just added computer use/web browsing into them. The product is the speed at which you can interface it and to be able to serve it with infrastructure.
Traditional product market fit isn't really applicable with tech evolving so fast.
So it largely sounds like many more people will be able to write software - and will use AI to do it. Existing software engineers will continue to automate their tasks away like they always did, but perhaps at a faster rate.
The impact of AI in other fields seems to be muted.
I think it is applicable to a much wider range of knowledge work, but it's also harder to apply there.
Software development has the huge advantage that mistakes and hallucinations are very easy to spot: the software works or it doesn't.
Spotting errors in a research report or legal brief is a whole lot harder!
But... non-software professionals spend a huge amount of their time on tasks that can be safely automated - reformatting documents, extracting numbers from PDFs, all kinds of flavor of data entry.
Learning how to use a tool like Claude Cowork can take a big dent out of those.
> Software development has the huge advantage that mistakes and hallucinations are very easy to spot: the software works or it doesn't.
Do we not care about code quality, maintainability, performance, extensibility, or understandability anymore? Honest question, not a gotcha, it's just previously getting software to pass all the tests was a small part of what we would consider "working" or perhaps "good" software. Maybe that's different now with LLMs, idk. Maybe we need automated checks for these things as well, like not compiling until the code quality is good enough to let the agent finish it's loop.
What code quality even means is different now, but also LLMs are capable of producing better quality code at scale in my companies experience. We are able to in fact sort of propagate best practices and structure via the llm to all of the teams even when they're working under time pressure.
> Do we not care about code quality, maintainability, performance, extensibility, or understandability anymore?
Yes, we should care. I've been writing a whole book about that: https://simonwillison.net/guides/agentic-engineering-pattern...
If the AI can write code for robots the impact in other fields may be pretty large. Seems to me a lot of jobs can be automated with software and robots combined. The limit in the past was writing the software to get the robots to work. But if AI can remove that limit...
"[would have spent] $1,199 with Anthropic, $980 with OpenAI"
How many tokens is that, input/output-wise?
(a) I'm curious if you feel like you got $2000 worth of value out of them in the last month?
(b) I'm also curious if you would have gotten similar quality out of a slightly lower-cost provider of an open-weight model? (e.g. Kimi K2.6 and DeepSeek v4 Pro) and what the spend would have been for that.
I myself have managed to spend not quite $4 on OpenRouter and have felt it was very worth it; I just have much smaller, or more targeted requests I guess. (Lately, adding features to a static site generator in Python, or setting up log forwarding via a docker compose file)
Claude Code:
OpenAI Codex: I'm confident I got value out of OpenAI - I've been mainly on Codex for the last few weeks.Not so sure I got that value from Claude, just because I've been using it a lot less and somehow the price came to about the same as OpenAI.
Given the code I've been able to build in the past month I genuinely do think I got value for the API price version, and (don't tell OpenAI or Anthropic) I think I'd have paid full price.
I've not spent nearly enough time with GLM-5.1 and co to compare, but I do know that the prompts I'm using with the agents are not prompts I would have expected to work just three months ago.
Cool! Thanks for the details, and your blog posts are usually interesting food for thought, so thank you for them too!
Are you saying that the software you wrote using those tools generated enough revenue to cover the $2000?
Not yet, but that's because it was almost all open source and I'm really bad at generating revenue from that.
When I account for the amount of time it saved me there's no question $2,000 was worth it.
If it were me I'd be asking "How long would it have taken me to do that, and what's the rate I'd have been charging for the work I would have been doing otherwise?"
Personally, I've probably spent $60 or so on OpenRouter in the last month or so and got a working project out of it that it would probably have taken me a fortnight to knock together (which is inevitably an under-estimate because it covered things I'd have to learn but K2.5/6 already knew). There's an orders-of-magnitude gap there.
> and both companies now have measures to lock their enterprise customers (who tend to sign year-long deals) at those API prices
why would enterprises do that if they can just use bedrock or vertex?
This is great analysis but my first reaction was - is this trolling? The fact that we have to think about whether a $1TN company has achieved product-market fit both articulates 1/just how high the valuations are 2/How hard it is to pin down EXACTLY what PMF is. As a pre-revenue startup, I am laser focused on PMF and frankly if this is the bar, no one will achieve it. But OTOH its heartwarming that people are willing to value you at $1TN before you reach that. Guess everything is in the eye of the beholder?
The title was indeed meant to throw subtle shade at the idea that supposedly $1tb companies were only just now starting to figure out product-market fit.
With respect to Simon, whose writing I've usually agreed with in the past and whose insights I've liked: this is a bad take that overlooks the extent to which corporations are imposing the use of AI on employees, and in particular ICs, who make up a majority of the AI-using workforce by headcount.
Many of us are either openly having our performance reviews tied to AI use, especially at larger enterprises. Whether that's measured by sheer token count or just "how many of your tasks are you using AI for these days" (combined with the implication that question carries at many orgs which are heavily invested in AI).
Are you saying that Anthropic's huge leaps in revenue are caused by stupid company policies and token leaderboards, and the moment companies stop imposing AI on their employees revenue will drop to a point where Anthropic are unlikely to be profitable?
I don't think that's the case. I think the token leaderboard thing (which is clearly ridiculous) affects a tiny portion of companies and is already going out of fashion.
I'm saying that the truth lies somewhere in between, and that Anthropic's current revenue is being, in part, propped up artificially.
We're also in a place where a lot of the usage guidance around these tools is still nascent. People are cowboying a lot of stuff, even as larger companies start to organize AI policy/safety/responsible use working groups to try and policy around the shortfalls of the technology.
IMO: if this technology persists, and if we figure out a way to use it in a broadly safe way, the value proposition will probably trend down rather than up, at least on the code generation front.
As a research tool, it shows some promise, though I still find the ethics of the technology disgusting.
The amount of caveats and vibes that this article is powered by could run New York City for a decade.
My Costs without CC Max+Caching over past 2 months: $112K.
Ran `ccusage` on my Claude Code logs.
- Total tokens: 22.2B
Without current Claude deals, my personal cost would have been *~$112,000*.
Is PMF enough. It is such a dynamic self-disrupting wave that it is like predicting physical chaos. These aren't early Googles in a blue ocean. Maybe a blue ocean full of pirates and dragons!
This isn't me being a doomer I just don't know. Can we look at Q2 profits and draw hockey sticks yet?
Remember people are boasting how much their expenses are. That is where we are in the bubble/new paradigm.
These folks have no lasting moat and they know it. We are still close to the November 2025 inflection point, so they had a clear advantage during these past few months. That will soon fade as open-weights models become really good, which is arguably already the case with DeepSeek V4.
Article doesn’t mention on-prem and on device models. Almost guaranteed that there are a range of killer enhancements on these fronts waiting to drop until IPOs get closer to inflict maximum chaos on the valuation games.
While the big guys will argue they’re worth trillions expect others to drop chaos booms showing their NPV may be effectively zero.
What baffles me is the range of estimates.
Operating profit is both post depreciation and fees paid to third parties for hire. So aside from shenanigans like RSUs and financing interest that's already somewhat close to actual economics.
Meanwhile we've got commenters here talking of 5-10 trillion with a T revenue shortfall.
Those are very different takes on reality
> If you are a heavy user of coding agents these plans are a fantastic deal. I just ran the ccusage tool on my laptop to get an estimate of how much I would have spent if I were to pay for API tokens in the past 30 days and got
You think this is fantastic deal only because they use similar like tricks where they inflate the price and tell you something supposed to cost $1000 but they have this today promo for $100.
I was there too and paying for a while. Few weeks ago I tried DeepSeek V4 Pro - expected its gonna be shit but its actually pretty good.
The deal is I pay daily ~$1 for DSV4-pro for ~100M API token usage. And they probably not getting broke because >90% of those token in practice is cache read and they very well optimized for that.
Yep, exactly this. And I have so much less anxiety that I have to use my 5-hour/weekly usage or I lose it... with deepseek api the credits never expire, I can use them when I want, how much I want and the prices are ridiculously low for the quality/intelligence/performance.
GPT 5.5 is maybe 4x the size of v4 pro, hard to compare price since cache hit is basically free with Deepseek, but 40x cheaper (with their 75% off) seems about right.
So ballpark same price per parameter as Simon.
I wonder how a focus on per-token API profits will impact the incentives to improve token efficiency and drive down costs through optimized compute. I suppose as long as a few leading labs are competing, we'll see progress in this regard, but it's certainly less in their interest than it is with a flat subscription pricing model.
There's also increasing pressure from China: note the massive permanent price reductions DeepSeek and Xiaomi announced in the past few days, with the possibility of more around the corner. And there's also the constant release of increasingly capable open weight models.
I think my company is going to cancel our subscription once it realizes we're spending more money in a week that a developer gets paid in a month.
Mentioned in the article, but it cracks me up that both openai and anthropic are utilizing fairly traditional enterprise GTM plans segmented by verticals.
So many startups trying to automate sales, but somehow the two biggest frontier labs have decided that the best GTM strategy is firmly human-in-the-loop.
Product market fit for the models, sure. But their cost structure is going to kill them as token costs drop this year, following DeepSeek and Xiaomi and soon more providers.
Great article I know this upsets a lot of people who are used to thinking Anthropic/OpenAI are just lighting cash on fire but they've cornered the market on enterprise who cannot walk away from these $200/month plans
However the valuations are still far far away from actual sanity
Have you tried the large open source code models?
I use glm-5.1 and occasionally deep seek v4.
They are as good or better than Claude's latest models.
And significantly cheaper. I've converted 3 of my engineer friends as well. All three have dropped their $200 month plans they had with anthropic.
We've all been a bit shocked at just how good these models are now.
If you "have" tried GLM (I specifically find it shockingly good for code). Did you not think it's not competitive to Claude, and why?
I use GLM-5.1.
It's good enough for personal stuff. It doesn't compare to the latest Opus I use at work. You can certainly argue I don't need Opus for work, but there is clearly a difference.
Also, at least with z.ai, GLM-5.1 is s l o w! After using Claude at work, I get really impatient with GLM-5.1 at home. When doing "true" vibe coding (i.e. not really examining the code), Opus is a ton faster (easily 5x).
But yeah, I'm not willing to personally pay for the frontier models. I won't even renew my annual Z.ai plan - it's become too expensive.
Hmm, I use opencode subscription, and glm seems just as fast from the tests I've tried to compare between the two. Tbh it mostly took Claude longer (mostly significantly longer) for the same tests.
Also, and I know you may not want to answer. But could you give me an idea of the type of thing you found glm to be worse with?
I think I've been fairly unbiased in testing a bunch of different development tasks. But am curious if maybe it performs well for some stuff and not others. So if you could share what you feel it's worse at.
Also are you an experienced developer or less experience?
Perhaps opencode zen isn't using z.ai as a provider?
I'll repeat something I wrote on an entirely separate HN submission.
When DeepSeek V4 Pro came out, I had been mostly coding with GLM-5.1 on a Z.ai coding plan.
I had a large analysis task on a relatively complex codebase. I decided to try the models out.
GLM-5.1 did acceptably but got a few things wrong (easily corrected) and took quite a while to get there.
Opus 4.6 burnt through the US$10 budget I had given it in about 10-15 min, without ever returning from the first prompt.
DeepSeek V4 returned a full analysis within 2-3 min, and I carried on all the way to implementing the feature I was after. Total cost less than US$1.00.
I now mostly alternate between GLM-5.1 and DeepSeek V4 Flash, with an occasional dip into V4 Pro for more complex analyses.
task i am working on right now at work is comparing two verisions of apis and documenting responses in their outputs. i suspect a vast majority of work at entrprise is of similar complexity.
right now everyone is using latest and greatest to do dumb stuff like that. that would change fast if companies start caring about costs.
What is the best IDE UI to use them? I don’t like CLIs.
Cline is pretty good if you use VSCode. It's one of a few AI agent plugins that works in the left sidebar.
Personally I'm happy with opencode right now
> enterprise who cannot walk away from these $200/month plans
Any org with more than 150 users aren't on $200/month plans, they are forced into API pricing + $20/month/user
For individuals and orgs small enough to get to use the subscription plans, that's all well and good until usage limits keep going down, or cost goes up. If you compare the usage you get on $200/month maxed out vs. what that would cost at API pricing, the $200/mont plan is an absolute steal. I doubt it will last long.
Not to mention the API plans are also still in their "lose money, just get the suckers hooked like addicts" phase. Once the reality-based pricing comes into play, it's a coin flip of whether the bulk of the companies fail, or they get to live off government subsidies for a few decades.
On the plus side, I'm happy I'll have a nice hay barn when the local half-built AI data center is abandoned.
I believe that API pricing runs at a healthy margin, at least compared to the server and energy costs used to serve the tokens.
Recent conversation here on that topic: https://news.ycombinator.com/item?id=47062534#47063134
> There exist a large number of people who are absolutely convinced that LLM providers are all running inference at a loss in order to capture the market and will drive the prices up sky high as soon as everyone is hooked.
> I think this is often a mental excuse for continuing to avoid engaging with this tech, in the hope that it will all go away.
Thanks for that psychological explanation. I was wondering why people were simply ignoring the math that shows that inference at API pricing can be quite profitable, e.g. published here for DeepSeek V3/R1 with 545% profitability: https://github.com/deepseek-ai/open-infra-index/blob/main/20...
There isn't a single thing about how the AI companies are operating that looks like a normal business. I know people who were in the room when Scott Sullivan, CFO of Worldcom, assured everyone that the future was bright at Worldcom days before they collapsed. So you'll have to excuse me if I don't believe the words of someone whose sole job is to justify hundreds of billions of dollars being thrown at Anthropic when he says their future is bright.
I agree that the amount of investment thrown at these companies is absurd.
But I also think that their API token pricing represents a real margin over the inference costs for serving those tokens.
Both things can be true at once.
> enterprise who cannot walk away from these $200/month plans
But that's the point of the article. Enterprise plans are starting to get API pricing, not the subsidized subscription pricing.
He is talking about tokenization of different llm and comparing, and this is inacurate because you cannot compare if you dont how it tokenice the words, some tokenicer could be cheap with more tokens than other because of their features. Not agree with Simón speech.
My employer went from the Max plans to Enterprise this month which was utterly stupid. We went from paying $200 per head to like $500 for some people, even more for others. For the same product. Oh well, guess we’re doing our part to prop up this bubble.
> That’s $2,180.16 worth of tokens for $200—not bad at all!
Just imagine how funny it will be if it comes out that big labs were doing some fancy maths to count the 2k$/month in their forecasts ...
Companies are kool-aid drinking now due to hype, but given how much they're spending, if they don't see REAL, BIG wins from it soon, they're going to scale it back quickly and switch to Chinese models. Claude isn't worth the API cost for a lot of development work, and once companies have had time to collect and crunch data they'll see this.
Swear people like you were hyping the frontier labs so hard not long ago.
Funny to see the change of tone - a lesson for people not to get too ahead of themselves.
There's no change in tone, I'm still very bullish on the tech, Claude in particular just isn't worth the API price, which I've always felt was too damn high. I have paid for Gemini 2.5 Pro, Deepseek 3.2/4 and GLM 5 tokens happily though.
Lmao you won’t admit it will you?
You financially benefit from stuff like agents. Of course you will be the last to admit publicly when things aren’t quite heading in the right direction. The gap between hype and reality is ever increasing.
You misread my comment because of your personal bias, and now you're acting like it's some sort of own? Come on bro.
The models are putting in work, that's indisputable, the thing I'm calling out is that Anthropic's price and hype are running ahead of the value it's delivering. Agents are a great boost at <=5% of SWE salary, a mediocre tradeoff/risk at 20%+, and insane at SV tokenmaxxing levels.
If we take out the circular interests and investments here then there is no way that this is a feasible business in current state.
Unfortunately Simon drank from the AI Cool Aid.
I know everything you’ve done for the tech community, but I please you to take some time off and reflect on this article. It’s not on par with ur usual level, but the tendency has been visible from the last couple of articles.
Genuine question: what's wrong with it?
I thought this was one of my best pieces of writing this year.
(In case you missed it, the title was meant as a subtle burn on those two companies - it's pretty absurd for them to only just be finding product-market fit when they're already valued at over a trillion dollars.)
I see you trying really hard to make things right and are everywhere in the comment. I feel a bit bad in formulating my comments a bit polemical.
Also my highest respect for responding so calmly without lowering your debating level to mine. I try to best to explain what I think is wrong.
To start I didn’t interpret the article as a burn.
I think it’s interesting to explain what’s wrong with it, because it seems similar to what‘s wrong with AI. The issues are subtle.
- Anthropic doesn’t have a profitable quarter, it’s financial engineering (https://www.wheresyoured.at/anthropics-profitability-swindle...)
- The first argument about your subscription price, doesn’t has anything to do with the overall claim of the article. It would if at all be a weak argument to support the opposite. Subsidizing prices signals a lack of PMF.
- That you hire sales people after you had a billions of funding is nothing surprising and doesn’t indicate PMF or not.
- AI Implementations are fresh and of course AI Failures are thin, but so are AI successes. I haven’t seen any companies creating billions of shareholder value because they’ve massively invested in AI and their competitiors didn’t You really can look at these things in 5 to 10 years and it is multi-faceted including cultural acceptance.
- That they need to buy more compute to satisfy the requests is probably the strongest argument in the artical, but don’t conclusive. The product is been sold heavily subsidized and in hype cycle. And again both OPENAI and Anthropic have to show growth in order to justify the IPO.
- Regarding the part about revenue I refer to the linkedm article above, as it does explain it very well.
The conclusion is reasonable given the arguments, but not the title.
However it is missing all the real discussion points that are actually in observation at the moment.
Local models as alternatives, IPO finanical engineering, how AI implementation actually will perform over years... Let’s all not forget crypto. It’s been full of "use cases" just a bunch of years ago. I like the idea of crypto(btc,eth) and I’m still invested, but 99.99% of coins have died on promies.
So this is not a piece of critical thinking, but this reads like a twitter thread to sell me a course :/
>Somehow this fragment turned into headlines like Uber’s COO says it’s getting harder to justify the money spent on AI tokenmaxxing, because the market for stories about AI failures remains enormous.
I notice this all over the place. Many people hate AI and want it to fail, and they're willing to invent misinformation if it supports that idea.
Well, it is a big news when the COO of Uber says it no? Not quite some small consultancy shop here.
But the COO did not say that. The headline was deliberately misrepresenting what he said.
The article was posted on HN and discussed a day or two ago.
https://news.ycombinator.com/item?id=48268871
No, he said exactly that, if you remove the corporate sanitised language designed to not offend the Uber CTO.
I think you're putting way too much weight into what one person said in unprepared remarks at the 27 minute mark in a 32 minute podcast conversation.
Unprepared remarks are the best source of info frankly.
That "one" person is the COO of Uber. And the other one - the one based on whose statement about burning through yearly AI budget in the first few months - the whole discussion sprung up internally at Uber in the first place is the bloody CTO of that huge company. So yes, their words do have A TON OF WEIGHT. Thats why they are in such important positions, arent they? They're not quite the Derek from the pub, casually commenting on how Liverpool will fare this season.
I think the way people reacted to those statements was entirely out of proportion to what was said.
I repeat: a CTO saying that they spent their entire AI budget for 2026 when that budget was clearly set in 2025 before anyone knew what those November models + harnesses were capable of is entirely unsurprising. Any analysis that doesn't also point out the difference between 2025 and 2026 era coding agents is either ignorant or deliberately misleading.
Yes, but that's irrelevant, because the COO uses that to base his core argument - that all that jackshit 1800 code changes per week that the CTO boasts about, mean absolutely nothing in terms of value. It means they are spending a lot on it, to gain as he diplomatically said "perhaps 20% more" - and I wonder 20% of fucking what - it's a ride-sharing app, what could they be possibly building on top of it with all that token crap?
You have to try pretty hard to get to "all that jackshit 1800 code changes per week that the CTO boasts about, mean absolutely nothing in terms of value" from what he said on that podcast.
(We still don't even know what Uber's planned AI budget for 2026 was. They didn't reveal that when asked - in https://www.theinformation.com/newsletters/applied-ai/uber-c... it says "He wouldn’t disclose exact figures of the company’s software budget or what it spends on AI coding tools").
I don't have to try at all - I think anyone who spent as much as an internship, let alone years at a modern tech corp would have no trouble distilling the absolutely clear message - we are spending too much for too little value. And actually wtf am I explaining myself? Every major tech outlet interpreted it like that too. It's not that hard Simon.
I think that both you and the other major tech outlets interpreted that poorly.
Oh really? How about the king of non-tech-outlets and digestible, shallow bites for the not-reading-books-middle-class, the Business Insider?
https://www.businessinsider.com/uber-coo-andrew-macdonald-ai...
You know their business is literally correct interpretation of the C-Suite statements.
I think Business Insider are not a particularly high quality publication - they're prone to clickbait - and that headline is an example of why I think that.
Hah, I just checked their homepage and here they go again amplifying that COO fragment from that podcast:
https://www.businessinsider.com/tokenmaxxing-debate-uber-exe...
> "That link is not there yet, right?" Macdonald said in comments that went viral, racking up over 2 million views on X. "I think maybe implicitly there is more that is getting shipped, but it's very hard to draw a line between one of those stats and, 'OK, now we're actually producing 25% more useful consumer features.'"
Yeah, something going "viral on X" is clearly a sign that it's quality information!
So no one got it right? Myself, tech outlets, non-tech outlets, everyone on twitter, etc?
That's why I wrote a whole section about it. I was deeply disappointed at how little thought people had appeared to put into this before amplifying the misleading headlines.
For someone who cares about media hype - https://hn.algolia.com/?query=author%3Ahansmayer%20hype&type... - you don't seem to be very discerning with regards to this particular story.
Who's to say those enterprises won't churn after XYZ comes out with a decent enough model that costs 10x less to use?
There's a whole bag of clever tricks you can play to juice short term results leading to an IPO that may not work longer term.
I'll believe they've found product-market fit when they have a product. Right now they're selling the infrastructure, in a highly subsidized and undifferentiated way (at least over a sufficient long period of time of, say, a couple of years).
Anthropic for sure. It's a useful professional product that I find many use-cases for in my professional and private life. OpenAI not so much.
$50/month tops is like 60B
I think this was obvious since the birth of ChatGPT
Intelligence is a universal good, it can apply to anything, and no, "human intelligence" is not the only form that is useful nor special. There are limitations to AI but also huge advantages, and its obvious that the advantages are worth paying for, given their revenue.
> Anthropic are strongly rumored to be about to have their first profitable quarter.
Is that quarter same as any other quarter in terms of infrastructure costs (e.g. are there any temporary discounts happening coincidentally)?
Didn't xAI basically donate the compute for that quarter so Anthropic could get to say they turned a profit?
The SpaceX S-1 says they're charging Anthropic $1.25b a month.
It also states that the first few months (this current quarter where Anthropic are reporting profit) are discounted.
Hey man, that discounted rate on Colossus 1 inference is purely coincidental...
That might be an understatement
The article is at least one year too late. Claude Code has been the product-market fit. It's so obvious retrospectively that questioning it at this point would be very silly.
The problem is not whether they have PMF (they do) but how they're going to compete against on-prem and Chinese providers.
Having PMF != printing money forever.
The author claim:
> That’s $2,180.16 worth of tokens for $200
No matter what it means, rebuild the same thing you built with these $2,000 tokens with DeepSeek Pro V4 and let's see if Claude has a chance to survive.
Ed Zitron begs to differ
I think we are overly fixating on Anthropic and OpenAI.
Since there are lots of models that are competitive and have a much better pricing, both OpenAI and Anthropic seem inefficient. I don't get why someone would want to buy shares after IPO apart from fomo and artificially built enthusiasm.
Anthropic and OpenAI may well be the Altavista and the Yahoo of the AI age.
I'm a huge fan of agent coding but kinda dislike this "llm evangelism".
There are still several open points (eg.: code churn, maintainability, subtle bugs human will never do) that everyone with a minimal programming knowledge that seriously used a LLM agent knows about but somehow none of these "big influencers" never mention (or just saying "it's your fault").
Man what a disaster article
The writer is on to something. Musk saw the writing on the wall and bought cursor. AI money is in coding plans
yep, and the issue is, they took investment
HN is the least agi-pilled place on the internet
HN has been on every wrong side of AGI predictions since the founding article on OpenAI...
You mean that HN has been predicting AGI will happen, or you mean the Singularity happened earlier today and I missed it that humanity is already all killed?
can we stop throwing around the term "profitable"? didn't they say "operating profit"? so so different to actually making a profit
I’m happy with spending zero dollars on ai lol..
They've finally forced the system to fake profits. Attacking personal computing, sucking the rest of the industry dry, yeah eventually you can make people pay for your shit. I still pray these vultures will implode without too-big-to-fail safety nets. And big daddy Trump can't help them anymore by fixing the system for them, no matter how much billions they are paying into his personal pockets.
The future are small models, nobody really needs big compute in the long run, that's why big tech is going for our personal hardware. So we won't become their competitor in their rent only economy. True competition is eliminated, natural evolution is being fixated by the government. This is not going to end well for the USA.
Sounds like mostly-baseless praise for the LLM behemoths / capital investment black holes.
> I currently subscribe to the $100/month Max plan from Anthropic and the $100/month Pro plan from OpenAI
... which already indicates a bias.
> If you are a heavy user of coding agents these plans are a fantastic deal... that’s $2,180.16 worth of tokens for $200—not bad at all!
Thank you for the sales pitch. Perhaps go buy a car and tell us how [insert your manufacturer] has "found product-market fit"?
> Anthropic are strongly rumored to be about to have their first profitable quarter.
First, the strong rumor is a claim by Anthropic itself. But even assuming that's true - it's an "operating profit", i.e. disregarding the massive capital expenses for years, and may also disregards ongoing capital expenses, if those happen not to be taken this particular quarter.
> 1 Trillion .. companies spending $200+/month/user will get you there a whole lot faster
First note the use of the first person plural to talk about Anthropic and OpenAI.
But that aside - most companies aren't paying $200 USD/user-month. But even if they were - if we take the 30 Million SW developers mentioned in trjordan's comment as subscribers, that's 2400/user-year * 30 Million = 72 Billion USD / year. And this is already rather optimistic, but - want to double that number of subscribers? Fine, make it 150 Billion / year. Still not there with a rosy outlook and assuming the hype and enthusiasm continue for many years.
And those rosy estimates are more likely than not unrealistic. I am reminded of this review of some empirical research regarding the benefit of LLM/AI use:
https://cmr.berkeley.edu/2025/10/seven-myths-about-ai-and-pr...
The big assumption with all of these sorts of analyses is that things will continue as they are for the foreseeable future.
In hype-driven markets, you cannot be certain of that.
Let's take a view that the author is right: coding agents and their associated harnesses were the inflection point for some degree of profitability and widespread consumption, and that these tools are now yet another SaaS subscription or API bucket expense to bake into every single developer (or developer-adjacent) in the organization alongside your collab suite, HR seat, CRM seat, design seat, etc. To be fair I honestly think that's a safe assumption to make for highly technical firms whose image is derived from remaining on the cutting edge of things.
That begs the following questions, which we won't know until IPOs start happening:
* Are subscriptions profitable, or just API consumption?
* What's the run rate when we just consider subscription-based usage like Claude Code and Codex? What about API calls?
* Is there any profitable pathway forward at which enterprises can get unlimited usage but at fixed rates via subscription?
* What does customer churn look like for subscription users versus API users?
We also have a number of questions for customers that I suspect we'll start seeing receipts for in the coming months, at least from the early adopters:
* What was the net gain (loss) from leveraging coding agents?
* What's the cost of a developer with or without access to a coding agent + harness? Is it cheaper to hire an outsourced worker with a coding agent subscription, or a domestic worker without one?
* At what point does further AI spend result in diminishing returns, i.e. where's the 'sweet spot' for spend?
* Did AI boost actual revenue and outcomes, or did it just gamify KPIs?
* What roles or work did AI actually replace, versus merely displace during the hype cycle?
Not to mention the questions regarding the technology itself:
* Will we develop the means to run foundational/frontier models at edge using less resources through some existing (e.g. distillation) or new technology, thus cutting off the profit centers of these firms?
* When the market mismatch between supply and demand is resolved, won't it be more affordable for consumers and companies to operate their own AI infrastructure rather than support further centralized buildouts?
* Will coding agents improve to the point of being able to bootstrap and self-orchestrate on edge/consumer hardware without substantial technical expertise, or at least improve to the point that traditional IT teams can securely operate them internally without an expensive subscription or API token bucket?
All of these will influence the long tail of this bubble, because it is a bubble at this point. Even if these companies are indeed profitable thanks to the coding agent inflection point, there's still so many unanswered questions about utility beyond coding that it's impossible to extrapolate a future. If coding agents are indeed the extent of utility for profitability, then there's no possible way these entities will recoup the investment already sunk into their infrastructure buildouts. Even if more profitable uses are discovered, does this offset or replace the firms disappearing due to AI speculation and their associated contributions to the economy as a whole (RE: the consumer compute industry at present, higher energy costs due to datacenter builds, opportunity cost from harms to local infrastructure from haphazard builds, etc)? Should these firms indeed be runaway successes and immensely profitable to the point of paying off their investors and growing the larger economy, does this end up stifling innovation in a world where most new ideas are fed into LLMs for R&D that are then controlled by only a handful of companies and immensely wealthy people, via systems that are easily surveilled and stolen from without recourse?
So many, many questions yet to be answered. Betting the farm because of coding agents is one hell of a gamble.
Wait till people figure out they can swap Claude code for DS4 Pro and spend a fraction in API billing (actually, significantly less than $100) while barely noticing a difference.
> Anthropic are strongly rumored to be about to have their first profitable quarter
No, its more like their own leak to WSJ and according to Ed Zitron -> seems to be heavily engineered via non-GAAP practices such as counting potential, but not realised revenue as actual revenue - the stuff for which I would be arrested if I did it at my company.
Also it appears according to Ed's analysis - strangely they seem to be projecting only that one quarter as profitable - potentially to calm the investors ahead of the IPO. Investor fraud anyone?
Also it was but a few months ago that their CFO said, in a court filing, that Anthropic's revenue across the entire lifetime of the company "exceeds $5 billion". Pretty strange.
https://www.reuters.com/commentary/breakingviews/anthropic-g...
How is it strange? The "exceeds $5B" quote was from December 2025. Anthropic has seen tremendous growth since then, ever since Claude Code with Opus 4.5 got really good at coding.
If you've ever been at a startup, this is exactly what it looks like when you go from not having product-market fit to having it (though with a few extra zeros on the end compared to most).
Ah yes, December 2025...such a long, long time ago...
Your comment is not a serious one. Their revenue has quadrupled in just a few months. So yes, December 2025 is a long time ago now.
You've had a couple lobotomies too many if you think their revenue has quadrupled in just a few months.
Hell, say it did, how would you possibly know?
Dec 3rd 2025: https://www.anthropic.com/news/anthropic-acquires-bun-as-cla... - "In November, Claude Code achieved a significant milestone: just six months after becoming available to the public, it reached $1 billion in run-rate revenue."
Feb 12th 2026: https://www.anthropic.com/news/anthropic-raises-30-billion-s... - "Today, our run-rate revenue is $14 billion, with this figure growing over 10x annually in each of those past three years."
Apr 6th 2026: https://www.anthropic.com/news/google-broadcom-partnership-c... - "Demand from Claude customers has accelerated in 2026. Our run-rate revenue has now surpassed $30 billion—up from approximately $9 billion at the end of 2025."
All three of those are official releases from Anthropic. You can choose not to believe the if you like, but since they plan to IPO this year it's in their interest not to get caught lying to potential investors.
As a non-public company they can use whatever non-GAAP black magic accounting to claim anything while still not technically lying. It just doesn't correlate to anything we'd actually call revenue.
It's still notable that, by whatever black magic accounting they are using, their number was $9bn in December and $30bn in April.
Not at all, because the magic can be applied differently at different times. They're undergoing a funding run now so they've got a massive incentive to come up with all kinds of revenue.
They've also signed a deal for billions in compute with xAI for april-may so they're certainly using that to fake billions in revenue using non-GAAP bullshit. It just seems a tad more likely than them legitimately increasing actual revenue by 233% in four months out of the blue.
Sorry man I hate to say this but you and many others need to stop commenting on accounting, finance and economics as its clearly way out of your realm of expertise.
Do you know what revenue recognition is? Do you know what accrual accounting is? Do you know of the phenomenon that is 'managed earnings'?
The only true objective number in finance is cash flows.
Do you have a source for that?
> Their revenue has quadrupled in just a few months
Maybe, maybe not. We haven't seen that S-1 yet. All we have is the 5B in lifetime so far. PLUS - revenue quadrupled or not, it only matters if their costs did not expand at the same rate or more. Revenue is not profit.
OK, so when S-1 comes out you will finally allow yourself to be wrong? Your prior is, a 1T company plans to IPO and their leader has been loudly committing an insane amount of fraud? I mean this of course is possible but that is quite the conspiracy. The scrutiny of an IPO would be a crazy thing to do if you were committing fraud at the scale you're suggesting.
Revenue is not profit yet the discussion in this particular thread is about revenue.
> 1T company plans to IPO and their leader has been loudly committing an insane amount of fraud?
Ever heard of Enron, Theranos, SBX ? They were all hiding in plain sight - who could've thought they were frauds?
That’s why I said it’s possible but it’s a very improbable and weird prior assumption to make
> weird prior assumption to make
No, at this level of capital involved, and so much opacity around the company financials, it's a perfectly reasonable assumption.
No, it's not. It's a stupid thing to say. Perfectly stupid assumption. There are 1000s of multi billion $ revenue companies operating and as a % the number that are fraudulent is close to zero, especially those public or looking to go public. There is always the possibility, but it's extremely naive to think it's likely.
Ed is a smart guy, but you or anyone basing your opinion on what one eloquent journalist says is ultimately a risky bet, no matter how much his reporting hits your particular dopamine receptors.
Please don't forget that Ed's entire brand identity is now 1:1 with exposing "AI" as a giant, unmitigated failure.
That's a very specific flow chart to hook your caboose to when none of this is even remotely close to endgame.
Pretty much. Ed does a lot of great work in digging through all this stuff but his conclusions always feel far too doomer oriented. OpenAL should have closed 5 times by now if you have been following his assertations from the start.
There will be big parts of what he says are true once the rubble settles but it will not be anywhere near what he is predicting. How that will shape out may not be great for the average person, what money shuffling tricks will be used? But it won't be a total wreck.
> It won't be a total wreck.
Honestly, I think it's very short-sighted to assume that all of this will be seen as any kind of wreck in the long term.
Normies are still catching up and reacting to chat-based LLMs.
HN types are further ahead of the curve, but still catching up and reacting to agentic coding and design workflows.
What often gets completely ignored is that entirely new modalities for how the underlying tech can be applied will continue to be demonstrated, and those will once again cause new ripples of excitement and disgust.
There are companies building world models and systems for protein discovery. Comparatively speaking, these approaches are barely in the zeitgeist today.
Deciding that we already have the data points we need to extrapolate how all of this plays out is like someone in 1974 deciding that microprocessors are just for accounting and inventory. Don't be that someone.
I think the big issue isn't so much a technology thing, I mean that will improve for a long while yet, but it is an economic one. My whole concern over the rapid expansion of LLM's is the massive build out on a technology that hasn't found its feet in a big enough range of markets yet that are willing to pay top dollar. Yes, world models for protein discover is very cool stuff and I kind of hate it gets lumped in with these other companies efforts because it has a very clear path forward that doesn't rely on massive IPO's just to keep the lights on.
This stuff is here to stay but I'm not sure how many of the current front runners will be able to stay solvent if they cannot turn these things in to massive money spinners. Revenue is fine-ish but spending is out of control. I see the debt in hundred of billions of dollars and start to wonder "Who is going to pay for this?" and "Will the people be willing to pay that much?". It just all feels forced rather than organic growth.
This is why I think Google may end up being one of the leaders in this field. They have their custom TPU's that seem to be fairly efficient at these tasks, they are slowly but surely improving their training and inference tech using their massive data set and most importantly, other parts of the business can subsidize this stuff for a decade if needed until it is genuinely profitable.
I am not against the industry but I do worry that many are rushing in with no means of genuine sustainability other than jump out for a golden parachute and let someone else clean up the mess.
I do hope I am wrong.
I think that is a reasoned and well-articulated reaction, thanks.
No question that there are some players who need a good weekend in Vegas to make it across the chasm of sorrow.
I mean optimistically I hope there is a big reduction of hardware requirements in future. One of those technological jumps that nobody sees in advance but looks obvious in hindsight. If the need to the amount of compute goes down enough,that is a brilliant path through.
It could happen, I hope it does.
We don't have to take Ed's word for it. Anybody who's capable of doing grade school math can see that the numbers simply don't work. These companies are literally spending orders of magnitude more money than they're actually bringing in. Cursor, who've been renting Claude, estimated just recently that a $200-per-month Claude Code subscription could use up to $2,000 in compute. https://www.forbes.com/sites/annatong/2026/03/05/cursor-goes...
Interesting story. Here's what it says:
> According to a person familiar with the company’s internal analysis, Cursor estimated last year that a $200-per-month Claude Code subscription could use up to $2,000 in compute, suggesting significant subsidization by Anthropic. Today, that subsidization appears to be even more aggressive, with that $200 plan able to consume about $5,000 in compute, according to a different person who has seen analyses on the company’s compute spend patterns.
The load-bearing detail here is if that means $2,000 of internal server+electricity costs, or $2,000 if they were to charge at their API pricing instead of the subscription cost.
The latter is how I understand these things to work right now. If it's the former then yeah, Anthropic are losing a TON of money on those subscriptions.
That's the big question, and nobody really knows what the operating costs actually are right now.
Frankly, everyone in the industry knows. When people make these statements without additional clarity they always talk about API prices. You can look at the NVL72 specs and make estimates for electricity and ownership costs rather easily. Inference at data-center scale is dirt cheap, even with public codes using dynamo and sglang. The mystery is why the early misconceptions about inefficient inference persisted even after NVIDIA was very open about everything they did to help reduce costs dramatically in the last two years.
Also, if I understand correctly, they are rumored to have a profitable EBITDA.
It's a funny metric considering Depreciation is a huge cost for them.
"We are profitable when we don't count our expenses"
There's a good reason to look at it separately: if inference is profitable then they make money (or at least lose less money) when they get more customers, because any fixed costs are spread across more usage.
Assuming that there are infinite suckers with cash to spend. It's entirely possible (if unlikely) that the market is not big enough to cover the training costs. Especially for multiple companies all burning insane amount of money on the regular.
Depreciation is part of the cost of inference. Inference happens in GPUs that have a relatively short lifespan.
Those GPUs are very expensive.
Inference is expensive because a GPU can only process a certain amount of requests in a given timeframe. Remember that Anthropic is constrained in compute.
If they are constrained, it means that those GPUs are not idle. If they have more customers, they will need more GPUs.
If they have to play silly games using EBITDA to be "profitable", then it means that they need to ramp up prices a lot more than they already did.
Which is why in these discussions I always say that inference is also extremely expensive. Too many people like to pretend without any evidence that inference is cheap.
Anthropic and OpenAI don't own data centers. Since they're renting GPU's, that's not depreciation. Paying rent is an operating cost.
Language models don't wear out the same way; upgrading is a choice.
Not a real choice.
You can "just not update an LLM" in theory. But if your competition updates LLMs, and gets more capable, more efficient LLMs, and you don't? They get more capable "expensive tiers", and cheaper "cheap tiers" of LLMs. What are you going to do then? Bleed userbase and die?
Sure, that's the competitive arms race aspect of it. But there's still some control over timing.
If they’ve entered into contractual agreements then actually it is debt.
I think the key thing that depreciates is all their models. You train one at crazy cost and 6 months later it’s worth $0. If you ignore that depreciation you look much more profitable.
Model inference compute outweights training compute by 10:1 and more for frontier LLMs. "LLM depreciation" is an expense, but not a dealbreaker.
Bubble popped when they increased prices. IPO may help cover some of the costs, but AI is very elastic and can be swapped out for any other company second to second. Which is why I think they bought up ram and disks like they did, to starve out competition and local models.
Correct. That’s exactly right.
The move to buy up ram is straight out of a industrial organisation textbook.
It was the root concession that scale will not solve AGI
AI companies/users are filled with liars and grifters, so any numbers/outlook they report should be highly suspect.
I must admit that I am going to find it fascinating when we hit the point where it becomes nearly impossible to deny the efficacy of these tools. I have straight up had people, even in real life, suggest that I'm lying about my productivity gains or what I'm able to accomplish with them.
Like, I understand the reasonable arguments against (I even agree with a few), but it's clear that some people have fully inserted their head into the sand and just don't want to believe any of this could be true. Which will be harsh, since I think getting hit with this train all at once in the future is going to be a rougher ride than a slower coming-to-terms-with, even if the result is one we're unhappy with.
I don't deny their efficacy, I'm saying that there's a massive crop of grifters and liars building them and using them.
In the meanwhile, Google AI search still says the next year after 2026 will be 2028.
Ok? Then don't use AI to do arithmetic. It's not their strong suit.
Oh, basic counting is now arithmetic? But I was told they were superintelligent and were going to cause an apocalypse because they can do pretty much everything ? Somehow because they can excrement a lot of text, we were told they can do everything else too?
> Oh, basic counting is now arithmetic?
Ye- yes? It's addition by 1.
Google's "AI overviews" have been utter junk pretty much consistently since they launched the feature.
There's a saying "the fish stinks from the head".
Yeah I'll believe it when I see it. Revenue is increasing but so are their costs.
Back in 2024 their CEO claimed training costs would rise to $10-100B in the next years.
https://www.tomshardware.com/tech-industry/artificial-intell...
thats not that far off. Costs like $100Ms to train a frontier coding agent model today, billions if you count the full pipeline. Combine that with the infra we're building out, the fact that you have multiple labs building similar scaled models, the industry-wide costs of training frontier models could easily surpass 10B/yr in 2027
Yes, when he made that claim back in 2024 they were spending like $100M to train a model.
Their CEO claims a lot of wild shit. He claimed in January this year, that in about 2-3 weeks from this moment, i.e. "in 6 months" that AI will be doing all of SWE work. Lets hold these people accountable for a change!
> "in 6 months" that AI will be doing all of SWE work
I assume this is the quote you're referring to from Davos?
"I have engineers within Anthropic who say I don’t write any code anymore. I just let the model write the code, I edit it. I do the things around it… we might be six to twelve months away from when the model is doing most, maybe all of what SWEs do end to end."
that was in Jan, he said "might" and he said 6-12 months. Yes! Let's hold him accountable for saying reasonable things!
Reasonable things? He said the same shit over and over over the last several years. Yes, lets hold him accountable - you don't make such "oopsies" accidentally, several times in a row.
Seems pretty reasonable to me. Timescales are hard for anyone to predict. He is forced to do these predictions to know how much compute to buy in advance. Surprisingly, he underbought compute and now has to scramble to secure it from xAI or wherever he can. So he was overly conservative...
> Timescales are hard for anyone to predict
Indeed. That's why serious people are very careful, even if they are not running a company supposedly worth 1T USD
> He is forced to do these predictions to know how much compute to buy in advance
Ah well, that explains it. For my companies next quarter, I'll just pull some random numbers out of my ass so we can make plans with material impact to company business based on that.
> That's why serious people are very careful, even if they are not running a company supposedly worth 1T USD
10x revenue growth per year, even more this year...his predictions about when agents will claim SWE e2e work are his speculations, relevant because people care about what he thinks as he is closer than anyone to the leading edge of the technology. It's also important for him to be as accurate as he can about this because he has to put his money where his mouth is. He has to sign the right amount of compute otherwise he screws himself. He got it wrong in the opposite direction that you're implying, so at this point it sounds like you are more interested in your axe to grind than the truth on the ground.
You think enterprises are adopting CC because they think "oh this will replace my SWEs I can fire them"? That's not happening at major companies. They buy CC because it's useful and the writing is so clearly on the wall in so many data points that to suggest otherwise is a bit silly at this point.
> For my companies next quarter, I'll just pull some random numbers out of my ass so we can make plans with material impact to company business based on that.
You, as a leader of a company, don't have to make predictions? Don't have to make bets about what the best thing for you to do next year? That must be incredibly nice.
Amodei and everyone else need to plan compute and plan their products and roadmap. You want him to....not do that?
> 10x revenue growth per year
To the stunning tune of 5B in the lifetime .
> You think enterprises are adopting CC because they think "oh this will replace my SWEs I can fire them"?
Yeah, that's actually Darios main talking point
> They buy CC because it's useful and the writing is so clearly on the wall in so many data points that to suggest otherwise is a bit silly at this point
Right, really sound arguments - writing is "clearly on the wall" and there are "so many data points". I'd be keen to use those immediately, but I am kind of missing the key of the "many data points" - namely, what did you build with it and how much ARR is it generating?
> You, as a leader of a company, don't have to make predictions
I have to make predictions, but not confabulations, lies and idiocies.
> Amodei and everyone else need to plan compute
FOR WHAT? Again, what was built with their shitty product in various companies and how much ARR did it generate? Uber seems to get no value out of it.
Anthropic has generated far more than 5B in revenue, I don’t know what sort of computer you have but it evidently does have the Internet, I would recommend using that unless the Internet CEOs are also in trouble for hyping that one up.
> Right, really sound arguments - writing is "clearly on the wall" and there are "so many data points".
Thank you for recognizing this. Don’t read Ed and think you understand anything about AI is all I’ll say. Read epoch capability index paper and look at the dashboard chart or the METR time horizon chart and methodology and then return with what I imagine from historical comments will be another ferocious and impressive act of mental gymnastics.
> I have to make predictions, but not confabulations, lies and idiocies.
Idk you’ve been misquoting and aggressively against addressing any facts you are presented with and yet bring no facts of your own (hint: if you know what you’re talking about typically you can calmly discuss with actual facts). That feels pretty similar to confabulations, I won’t say idiocy I’m sure you are not an idiot but you seem to have a lot in common with your caricatures of tech CEOs.
> FOR WHAT?
Their product.
> Anthropic has generated far more than 5B in revenue
A sworn affidavit by the Anthropic CFO from Dec. 2025 is what you need to look up mate.
I work in big tech and probably 90% of code over the last month has been written by AI. And I suspect it's probably higher within Anthropic, which is probably what he's basing his opinion on.
So, he's closer to correct than not.
That said, your recollection is also flawed. It was in mid-March, and here's the relevant quotes:
>I think we’ll be there in three to six months—where AI is writing 90 percent of the code. And then in twelve months, we may be in a world where AI is writing essentially all of the code.
[...]
>But the programmer still needs to specify, you know, what are—what are the conditions of what you’re doing, what—you know, what is the overall app you’re trying to make, what’s the overall design decision? How do we collaborate with other code that’s been written? You know, how do we have some common sense on whether this is a secure design or an insecure design?
[...]
>So as long as there are these small pieces that a programmer, a human programmer, needs to do, the AI isn’t good at, I think human productivity will actually be enhanced. But on the other hand, I think that eventually all those little islands will get picked off by AI systems.
With another 3-4 months left on the clock, his prediction seems remarkably on point for at least certain organizations and domains.
I welcome you to also hold yourself accountable in the coming months if this trend continues. ;)
> And I suspect it's probably higher within Anthropic
That probably explains why their uptime and reliability are so bad.
Written, but was it reviewed? Do you need to edit code written by LLM?
I agree that most of the things are written by AI but writting code was never the bottleneck in big tech.
Yep! We have a review process where we have a few agents, each tuned to a particular domain of expertise (security, code quality, etc) which iterate until the feedback meets a certain threshold, at which point it goes over to humans for (hopefully) final review.
That said, I generally agree that you're correct: writing code in many ways has not been the biggest bottleneck. However, by removing much of that writing, it frees up engineers to work on the uniquely human things that are larger bottlenecks.
I had a few comments in a thread here touching on where I think most of the value has come from for us (which is largely search/understanding of our dependencies and making away team work far more viable, which aids with cutting through bureaucracy and the tendency for teams to push back on work): https://news.ycombinator.com/item?id=48298731
Haven't you heard - these days they just throw slop generated by LLM agents over to other LLM agents which cosplay as internal QA. They know it works because they write really strict .MD files where they instruct agents in English language to 'never do this' and 'always do that'.
This is really what you think happens at large tech companies? You don't think it's possible this is maybe even slightly overly simplifying what the relevant processes are?
Read the other comment in the thread. Your buddy literally confirmed exactly what I wrote.
Comment does indicate you don’t really seek to know how things work with respect to this and seem to not be able to imagine that the Occam’s razor is: agents are more useful than you think they are.
>Read the other comment in the thread. Your buddy literally confirmed exactly what I wrote.
Please engage in good faith. I commented that humans are the final step of the review process.
> I welcome you to also hold yourself accountable in the coming months if this trend continues. ;)
My company did not swallow hundreds of billions in shady investment deals and is not publicly traded. We work with real money, and the revenue on our books is the revenue that is actually booked, not fake revenue we plan in 2 years time to maybe happen. So no, I am not going to hold myself accountable. But people who work with other people's money should be absolutely held accountable when their wild imaginations don't come true, repeatedly, quarter after quarter, year after year!
I think he means hold yourself accountable when it turns out your predictions and pessimism don't age well.
Mate, for 5 years I've been hearing that crap. I am not predicting anything / on the contrary the AI boosting bunch is. When are your predictions coming true?
AFAIK, most predictions from several years ago were for...approximately now to within the next few years. Can you be more specific?
You criticized a very specific (and fake/misquoted) prediction, ignored the correction, and are now criticizing vague hand-wavey "predictions" that you have left unspecified.
Can you please stop with the angry/ranty replies and actually have a real conversation grounded in actual facts?
Now, having said all of the above...I'll also point out that these are predictions, not promises/guarantees. These people are being asked to forecast and are doing so. I hardly think they should be held responsible for not being literal oracles, but even so--please, at least quote them correctly/at all.
In short: be better than the hallucinations you're seen to call out from the models.
What predictions, sorry?
I will note that you have essentially not responded to anything specific in my comment, nor at least acknowledged that you misstated Dario Amodei's actual prediction.
Elon playbook
>according to Ed Zitron
So, unsourced vibes from a shady guy whose entire empire is built on being against AI?
I genuinely don't know how folks can continuously buy into anything he has to say after that Wired piece. The credibility there is seriously lacking.
Please, continue to be skeptical of the labs. But people need to stop talking about this dude as if he's the Holy Grail of the anti-AI movement. It's going to blow up in y'alls faces.
Ed actually provides sources and goes into an incredible amount of detail as to how he came to his conclusions. The average AI booster just goes "I totally built ten businesses off vibe coding but I can't tell you anything because it's a SECRET!". And the mainstream tech media is so in the pocket of big tech and AI corporations that they might as well just publish their PR emails at this point. Yeah, I'll listen to Ed thank you very much.
I think it's telling that most critics don't address his actual points, but instead his credibility because he's a "hater".
Ed actually seems to make some really serious errors in his work. Tim Lee called out a particularly egregious one here, though it's one of many: https://x.com/binarybits/status/2050562429709377986
That said, I really mean it when I say that I don't actually think Ed is a good choice for the anti-AI movement. I think an actual opposition is useful, but he ain't it.
I really recommend you read the Wired profile if you haven't yet and form your own opinion: https://www.wired.com/story/ai-pr-ed-zitron-profile/
I read the profile and didn't see anything really wrong. Why would PR companies have to believe in their clients? Why does he have to be held to higher moral standards than Sam Altman who’s a total lying snake?
The error you call out is hardly “serious”, as the whole argument is uninteresting. It is a stupid indefensible error but the argument about revenue being 20% or 30% lower than reported isn’t that central to his overall thesis. Stuff that matters is inference cost, profitability, actual training costs.
> So, unsourced vibes from a shady guy whose entire empire is built on being against AI?
Actually he provides sources when he analyses stuff and imho much better than the usual corporate "Sam Altman says we should ask ChatGPT how to raise babies" crap. Also, I don't know many 'shady' guys who have built entire "empires", nor does he seem to actually have an empire. Usually being shady means you are kind of unknown and all. I am not glorifying Ed, don't even know him personally. I am not even impressed with his writing style much to be honest. But he brings important facts and information to light, which otherwise would have been lost in the cacophony of corporate media light treatment of these con-men. Holy Grail? Blowing up in our faces? WTF are you talking about?
>Actually he provides sources when he analyses stuff and imho much better than the usual corporate
You said it was likely an internal leak to the WSJ "according to Ed Zitron". Did Ed have a source for that, or was it just vibes?
The source was the article in the WSJ itself, which then referred to their source at the Anthropic. Which kind of is a textbook definition of "leak". Because otherwise Anthropic would have their lawyers hunting both the employee breaking their stringent NDA and the WSJ as well...
Fair enough, but I have to admit I'm puzzled about why you felt the need to then attribute it to Zitron?
Why puzzled ? I literally said "According to Ed Zitron", implying that's where I stumbled upon the article. I've no time to read corporate media, at least not regularly.
>more like their own leak to WSJ and according to Ed Zitron
^ Apologies, the above read to me like you were saying that Ed himself was claiming that Anthropic leaked to the WSJ.
> I currently subscribe to the $100/month Max plan from Anthropic and the $100/month Pro plan from OpenAI. If you are a heavy user of coding agents these plans are a fantastic deal.
Agreed. But its only a great deal because it is heavily subsidized, as you said yourself. Enjoy while it lasts, but in my book, product-market fit means something along the lines of "product which enjoys a loyal customer base, sold at a price perceived fair by the customers, and generating profit. How many of these does your definition of product-market fit hit here?
How will they stay profitable if every business lays off engineers because of AI and there are no engineers to use it? /s
[flagged]
We detached this comment from https://news.ycombinator.com/item?id=48297777 and marked it off topic.
The guidelines specifically ask us to avoid nasty comments like this. These lines in particular make it clear why your comment is out of line:
Be kind. Don't be snarky. Converse curiously; don't cross-examine. Edit out swipes.
When disagreeing, please reply to the argument instead of calling names. "That is idiotic; 1 + 1 is 2, not 3" can be shortened to "1 + 1 is 2, not 3."
Please don't fulminate. Please don't sneer, including at the rest of the community.
Please respond to the strongest plausible interpretation of what someone says, not a weaker one that's easier to criticize. Assume good faith.
Eschew flamebait. Avoid generic tangents.
https://news.ycombinator.com/newsguidelines.html
Fun fact: the $20/month subscription fee for ChatGPT Pro - which set the standard for at least a couple of years - really was an arbitrary decision made based on a Google form: https://simonwillison.net/2025/Aug/12/nick-turley/
I wonder how Ed Zitron will shift goal posts this time, and how long it will take for that article, when published, to reach HN front page.
Simon Willison just hit the "Publish to top of HN" button.
Wish I'd hit that one the other day on this one, which I cared a lot more about: https://news.ycombinator.com/item?id=48228321