It's not an affordability crisis, it's a financial crisis. The models get cheaper super fast. By this time next year Fable 5 will cost less than Sonnet does today. That's not the problem. The problem is that many companies are going to realize that they don't get any ROI from AI. Generating code faster != more profit. Most of the Fortune 500 will likely realize this and then the token budgets will come crashing down. Most of their ideas are _bad_ ideas. Moving faster, won't lead to more profit. Sure, you can use AI to potentially replace software engineers, but the F500 are also terrified of not having accountability or making mistakes. They won't be firing any engineers.
I think the biggest problem is not necessarily the cost to develop & serve the models, but how quickly user behavior changed with token based pricing.
I know a lot of people at companies where the marching orders changed on a dime end of Q1/start of Q2. These are shops that were fully on the "use AI or die (because we will fire you)" train.
Now there's monitoring, reporting, alerting not just on overall cost but on "over-use" of best/priciest models based on total-or-percent tokens/dollars, etc. All of this comes with direct developer engagement & standardized management escalation for holding it wrong.
To me this customer behavior does not smell like a product you can 10x the pricing on to get profitable. We have exited the exploration phase and now ROI matters.
I can give you some additional anecdotal evidence to support your comment.
I work at a Fortune 200 company. At first, it was the Wild West. Need an LLM? You got it. Need to or want to build an army of agents? Done and done. We literally had everything at the tips of fingers for about 3 months. Teams were building their own internal tools, the team I work on canceled contracts with several software vendors because teams were building the same tools for what they thought was nothing.
Then they signed contracts with Anthropic and Google because I would assume they saw the token usage was through the roof. One month later? They completely cut off access to everybody for both Claude and Gemini. If you wanted access? Suddenly it was several forms, along with several approvals and a rock solid business case why you needed it. And before you got to the forms? You were added to a waiting list that was thousands of people long.
The entire company is now in damage control after trying to get the genie back in the bottle. I'm guessing someone saw how much we would be paying for the tokens we'd been using and decided to shut the party down so to speak.
AFAIK nobody was collecting analytics. The one team I was working on had put out a goal of "30% more efficient" using AI tools. Its about as subject as you can get. We never got around to what exactly that meant before everything got shut down.
Myself and several other devs were laughing about the whole thing. The company was so amped about what AI could do they never even bothered collecting any analytics that would affirm or deny any of this had a positive impact. Even some of my team members were talking about the placebo effect AI has had on a lot of C-Suite folks.
I can second this. Our company and department was all-in on AI. And since the token-based pricing came in, we got an email from IT that tried to explain that most developers don't know how to choose models and that the cheap models should be good enough for most of our work ..
Over the last month I have seen companies scrambling to measure deliverables against cost. Most of the back room talk is to the affect of giving devs a small allowance ($500 a month) and then making them prove their own productivity increases (again, based on deliverables, not LoC) before they either take it away or give them more.
Obviously this won’t be on an individual basis but some kind of unit.
Either way, with how much I see these companies cutting back I have no idea how the big AI companies are going to be profitable.
I would imagine it only gets worse in the face of good-enough open/chinese/local models too right?
Microsoft adding Deepseek support already as I recall?
That is - for any definition of "they are behind X months" then eventually they get to the point Claude was in January when the world freaked out, but at 1/10th the cost. A lot of firms are going to mandate that is good enough for their developers.
I believe this hasn't been confirmed yet but I think it speaks to a bigger problem for the AI companies which is, if you give capable developers a good reasoning LLM, they can make it work like it was a really expensive model.
I believe we are 100% at the stage of good enough for the vast majority of tech companines. Fable and others will be more valuable for non-traditional tech companies.
I read somewhere that the chinese AI companies are sharing knowledge and it would not surprise me if the government is applying pressure by saying work together or else. If they work together, they can truly commoditize LLMs and with China ramping up hardware support for AI, I see the future being inference speed and hardware being the moat.
If hardware becomes the moat, the US frontier labs are screwed. We have AWS, Azure, GCP. All three have or are making inference silicon. LLMs become just another service in the public cloud's large service catalog, and open weight wins.
Which makes sense to me. Selling a chatbot interface/model access to the general public was never going to be a viable long term play. You still need developers to wrap the models into specialized tools. Queue the Jobs quote "It's a feature, not a product."
The big thing is, the western world has moved so much of the manufacturing to China and think a lot of people will not forgive Samsung and others, so I can see China owning a good portion of the supply chain.
> The big thing is, the western world has moved so much of the manufacturing to China
I built my career on Solaris and it got rugpulled by Linux.
That wasn’t because of software, it was because of hardware. Linux’s cost advantage existed because Sun hardware had huge margins, because their software was basically free.
AI will probably be a repeat of this. Whoever can come up with the hardware solution that minimizes the cost per token will win.
I believe the 5090 still holds this crown, but someone certainly knows better than I do.
but not all tokens are equal and vertical integration is the name of the game. Solaris did not lose to Linux, it lost to the LAMP stack on commodity x86 hardware. without the "AMP" part, Linux would've been dead in the water.
100%. There will be strict quotas on the expensive models and day to day work will be done on the cheap models that are "good enough" with escalation to the metered models when the cheaper options are spinning their wheels. Eventually the US frontier lab APIs will only get the most heavily triaged work that multiple tiers of cheaper Chinese open weight models have failed on.
And of course the C-suite will have unlimited access to Mythos tier models, which they'll use to summarize reports, while passing down mandates to rank and file to increase usage of less expensive models.
Yup, we are in the process of getting access to US hosted Chinese models. I've been petitioning Google and our rep, we will see but I suspect they will cave eventually. Gemini sucks and if they don't sell what their customers want, we go shopping around.
"Elastic" in economics happens to refers to how elastic the supply/demand is when the price changes (not vice versa, as you're describing). So e.g. an inelastic demand means the quantity demanded changes very little when the price doubles.
Elastic demand means buyers are highly sensitive; a price hike causes a massive drop in purchases. Inelastic demand means buyers aren’t very sensitive; they keep buying regardless of price
> Zitron's numbers don't tell us the real cost of generating tokens but, subject to the assumption that the platforms are not subsidizing the token price, that means Anthropic is subsidizing their enterprise customers by up to 40 times, and OpenAI up to 70 times
Neither Anthropic nor OpenAI are subsidizing enterprise customers. Neither Anthropic nor OpenAI allow Business nor Enterprise customers access to the high value $200/mo plan. Both organizations have moved to a "cheaper plan per user + API Pricing after that" (e.g. $20/mo + usage). The $100/$200/mo plans are for individuals only (of course, many individuals use these plans at work, but that's beside the point; they aren't selling this plan to enterprises).
> SemiAnalysis also analyzed the platform's gross margins, implausibly assuming that tokens were priced at 4 times the cost of generating them and: With the current subsidies, all it takes for a user to have a gross margin of at best negative 25% is for them to use as little as 25% of their rate limit.
The article's source for this claim is not SemiAnalysis; its Zitron. But once you dig through his article, Zitron links to a SemiAnalysis tweet [1] where they, as the paragraph states, implausibly assume gross margins of 75% to come up with their weird analysis of the subscription plans. Citing this for anything is weird, because afaik that 75% number is a total shot in the dark. We have no clue what their margins are. My take is that the only reason that 75% number is implausible is because it may underestimate the inference margins of Ant/OAI's API pricing.
> it may underestimate the inference margins of Ant/OAI's API pricing.
If true then why are neither Anthropic or OpenAI dropping their API pricing to gain market share when both are clearly doing all sorts of political and PR maneuvering to compete in a cutthroat market?
Since they aren't dropping the API usage prices (and are in fact raising them in a lot of subtle ways) then one of these options almost has to be true: they are still subsidizing inference, training costs are so ridiculously high that they need to make huge profits off inference or collapse in on themselves, or they are price fixing.
The training costs are very likely the reason. Dario has talked about how each individual model is profitable, but how the expenditure training the next generation of models makes it look like they're not profitable at any given moment in time, and I believe he's being honest about that.
The market for open weight model hosting gives you an idea of the profitable price floor, it's pretty clear there's markup baked into OAI/Anthropic's APIs.
> Neither Anthropic nor OpenAI allow Business nor Enterprise customers access to the high value $200/mo plan.
they may not "allow" it, but i've seen first hand enterprises encourage employees to use these accounts personally and get reimbursed later to avoid pay-as-you-go w/limits pricing for users who do tokenmaxing as a cost control measure...
Given my experience with hosting these models at scale, working and optimizing load, I don't think the margins are nearly as high as 75% if the models are as big as people often claim.
Only reason deepseek is so cheap is because well I don't know, but actual pricing should be around their initial price which was 4x, at that price you have a healthy 25-50% margin based on occupancy, given the deepseek v4 is a very sparse moe model.
GLM 5.2 for example doesn't have more than 30-50% margins that's assuming old pricing for GPUs, current inflated GPU pricing well I am certain the margins must be lower.
Ofc you can host for cheaper with quantization, and if you have very consistent capacity/utilization, which is not the norm with AI workloads.
Overall for large models like GPT 5.5 or Opus there must be healthier margins of around 50-70% assuming GPU pricing didn't increase for these companies. Even if it did 30-40% margin should be possible, even in worst case assuming all GPU they had saw a jump in pricing.
For smaller models it's hard to say, I would guess 20% but these models might be much smaller than I suspect, then it might be double that.
Note the issue is less intelligent tokens don't linearly scale down in memory usage, which is the biggest pain point of serving models. Context sizes have fucked us all.
Also anyone claiming OAI makes less margins on APIs or stuff might be wrong given they are on much lower context size, 1M context definitely is a lot more expensive to serve especially with smaller models like sonnet.
Serving the API is profitable. They are unprofitable because of R&D (and maybe subscription costs?). If they can continue to find access to R&D capital, there is space to reduce API costs.
There is no moat until a company achieves RSI and/or AGI, and the one that does succeed in moat-making will do so by hacking into and destroying their competitor's infrastructure.
Once moat is achieved, you don't have to compete on price. Of course it'll be academic because the AI will probably destroy all of us.
They may not be able to! It's pretty widely acknowledged, for example, that if there's some surprising plateau hiding around the corner they're both going to fail. But that could mean that they're overcharging for AI usage to get research money and sustainable rates are lower rather than higher.
I think that for coding we're past the plateau issue. The frontier models of today are good enough and very valuable. The expensiveness in running them will eventually be solved by cheaper faster hardware.
I do hope that a day will come where you can buy the nvidia spark thingy for 5k that can run the equivalent of Opus 4.6 or 4.5 locally and that would be a massive thing.
> The expensiveness in running them will eventually be solved by cheaper faster hardware.
How?
* Moores Law is almost over. The 5090 improves over the 4090 mostly because of quant improvements.
* even if the hardware improves, there’s a huge incentive to slow roll the next generation. Nobody wants to end up like Sun Microsystems. Sun’s used hardware was faster than its new hardware, once you considered price. Sun ended up competing with its own used equipment.
The most obvious place for improvement is RAM, network and storage.
If someone can bring more RAM onto the market, that will unstick things.
GPUs are not really the ideal architecture for running neural networks; they are heavily bottlenecked by memory bandwidth and struggle to keep all their tensor cores supplied with data.
There is significant room to make more specialized neural network accelerators with new compute-in-memory architectures.
If the brain can run 86 billion neurons on 30W it must be possible.
The whole hidden plateau hypothesis is kinda bunk, because we're already pretty far in a plateau for general knowledge/question answering, but there are many subdomains where we can push model capabilities, and as we saturate one subdomain we can just shift to another economically valuable one.
There isn't one AI intelligence S curve, there are thousands of them, and they're mostly invisible in the major benchmarks, but for someone trying to do work in that specific area of capability, the progress is transformative.
I'm skeptical of a hidden plateau, but I really think it's overconfident to assume there's not one. Remember that it doesn't even have to be a technical plateau; the effective plateau of e.g. car speeds is determined by regulations and road conditions, and far below what "frontier cars" are capable of on a controlled racetrack.
1 player is enough to create tension on price when "don't buy it at all" is a comptetative option. By most accounts, Anthropic and OpenAI both lose to "just don't buy" when they try charging at cost.
Chinese models are dropping in price thanks to ridiculous levels of state subsidy where companies are forced into aggressive price wars to survive and grab market share. I am guessing this will also blow up sometime next year or in 2029 at the maximum.
Btw, some Chinese corporates have already seen this and increased their price. Zhipu AI & Tencent for example. Alibaba, Baidu, and Tencent also announced multiple price increases for their AI services.
China has the benefit of vast solar power and rapidly increasing battery capacity. Yes, that's subsidized, but it pays for itself in the long run.
And, even with the price increases, Z.ai and Tencent are still much cheaper than Anthropic or OpenAI models. I think there's an efficiency focus among the Chinese models that is absent at OpenAI and Anthropic, and in the end I suspect efficiency will be the winning feature. Google seems to understand that. Gemini 3.5 Flash is pretty competitive with the big guys, and it's small enough for Google to run it profitably (I assume) for a price that's much less than the frontier models. Gemma 4 models are showing off a bunch of efficiency techniques (MTP, QAT, the 12B encoder-less vision model that soundly outperforms much larger vision models, DiffusionGemma), and I assume they have several more techniques that aren't published.
Chinese companies like Deepseek are operating on shoestring budgets (allegedly less than 300 employees at Chinese wages). It’s not that self evident there is anything that needs subsidized besides compute (due to limited manufacturing capacity and access to Western chips in China)
Shouldn't we know a better answer to these questions once Anthropic's IPO materials surface publicly? I understand, and maybe even expect, SpaceX's materials to be all over the place and skate on by any discussion of unit economics, but the nerds over at Anthropic might just be forthright enough to just tell us what their margin is on tokens as part of their IPO.
To be honest, making sense of finances of fully public companies is often hard, because in practice, accounting is hard. How you account for depreciacion, cost, investment, fixed vs marginal costs is in practice fluid, companies have an incentive to make it look attractive, while also optimising for tax and shifting revenue around to narrowly beat analyst recommendations.
Here's a concrete example. Does some random AI company make operating profit on inference? I.e. if you only kept marginal costs, would you make a profit?
Well, depends what you account as your costs. If you're using hand-me-down hardware from previous generation's training, how much do you charge yourself internally for it? Maybe you show less, so investors take solace in profitable inference, even if you're losing money overall. How exactly are you accounting for electricity costs between training and inference? Is your army of SREs mostly servicing training new models (R&D expenditure) or inference (operating cost)?
This even has a name, and is called the "big bath" approach. If investors expect one part of your business to be a fiscal black hole, just shove all your costs there. They are accepting of it, and you make the rest of the business look better.
I'm not accusing AI companies of cooking the books, rather I'm trying to highlight you could see all the cash flows and still not know how much money is made or lost where.
I saw some commentary that their free cash flow is misleading because it doesn't subtract the stock compensation they are paying to attract / keep top AI talent. Their point was also that deciphering financial statements is hard
Well it probably doesn't help that Dario is going around on podcasts saying things like "frontier labs need $1T of revenue or they will go bankrupt" lol.
Dario’s company may be creating super intelligence that will kill us all in the near future, but at least he seems to be brutally honest about all of it.
My point is that $1T of revenue is A LOT. Apple & Google each only did $400B revenue in 2025. Facebook did $200B. Think of how many decades it took the 3 to get there.
So depending on how literally we interpret Darios comment, OpenAI & Anthropic need to get to Apple+Google+Meta revenue numbers in like single digit years?
The estimate that AI companies need to replace 27% of jobs to service their debt is interesting. But at least Anthropic and Meta seem to have their eyes on replacing software engineers.
There are ~1.6M software engineers on the US [0], earning a bit under 150k/year on average [1]. If AI companies captured all of that spend, that amounts to about 250B/year. The article assumed that they need around 300B/year to keep up with their debt.
At least based on Meta's recent behavior, forcing 30-50% of developers to switch to data labeling, it looks like that is actually their game plan.
Once your dependent, they can drive up the price just because. It doesn't need to be for existential reasons.
This is the crisis point for vibe-coders. A developer can go back to writing code by hand, as horrible as that might sound. Someone who hasn't learned to code but builds with AI can't go back. They either pay or they stop. That will be an painful choice whichever way you fall.
There are already open weight models out there that are capable and cheap enough for a lot of coding tasks. Not as good as Claude but not far from it. There's no going back to pre-AI coding.
I can't speak for everyone, but for most of my coding tasks, Claude is just barely good enough. There's no going all the way back, and perhaps open weight models will keep improving, but at least 50% of my work would be better done by hand than by a worse-than-Claude agent.
I consider Opus 4.5 the crossover point where coding with agents got more efficient than not coding with agents. They were too stupid before that, and wasted more time than they saved for anything beyond a basic CRUD app or HTML page.
Certainly, the best models have gotten better since then, but I wouldn't consider DeepSeek V4 Pro or GLM 5.2 to be a big enough downgrade to be worse than coding by hand. I'm willing to spend a premium for the best model for coding because it wastes less of my time with dumb stuff, so I've got a Claude subscription. But, there is a limit to how much of a premium I'll pay. 10x over Chinese models? OK, fine. Opus saves me enough time to make it worth a couple hundred bucks a month. But, 100x, or more? Nah. I'll go a little slower, review the PRs a little more carefully.
And, open weights models do keep improving. DeepSeek V4 Pro is a notable improvement over earlier DeepSeek models, and the first DeepSeek model to cross the "better to work with it than without it" threshold into Opus 4.5 (or better) territory. GLM 5.2 is somewhere in the ballpark of Opus 4.6 (though without vision, a notable limitation for anything that requires a UI).
All of the silent, hidden model routing OpenAI does strongly suggests that the unit economics are not just fine, at least not yet.
If apparently the only way you can make money with your product this early is to dilute and adulterate it behind the scenes, it strongly suggests you want the customer to continue to believe they are getting value that you can't afford to supply.
More prosaically: if either of these firms could prove that they were even really close to profitable on inference, they would have bloomin' said so while they were trying to raise more money.
The dependent idea is questionable- when your boss tells you to not use the most expesive models-you just dont
I would assume when price hikes happen either
1) less non technical people would vibecode as it doesnt impact the work that much
2) people use the cheaper chinese models
3)we're jamming ai into everything because were exploring. We will just niche down into use cases that provide high roi
AI is a worker for me. That i pay for. Basically i am in the same game now to reduce the prizes i have to pay for my workers. Just like the employers are, that seek to reduce costs for employees, as we are simply too expensive. We need more competition among the workers. Let's introduce more chinese workforce! ;)
If you had a choice of maybe 3-4 contracting firms to hire workers from and you weren't large enough to negotiate on price I think you'd be in a pet bad spot as a business?
I'm finding it challenging to believe they wouldn't just cannibalize anything dependent on them in that way or at minimum launch a directly competing product.
Once locals get to Opus levels I think it we may see a phase change because that + a reasonably competent programmer is going to be a very powerful combination for most practical programming problems.
Frontier models may eventually achieve super-intelligence (no opinion beyond mild skepticism) but super-intelligence isn't necessary for most practical day-to-day programming. The problems, as always, become communication, understanding what users really need, etc. that is, softer skills.
A superintelligence is one that exceeds human intelligence in all areas. Which roughly translates to learning, adapting, and performing more quickly and efficiently than even the best humans. This is closely related to "the singularity", which is when technological growth becomes uncontrollable by humanity.
That sounds like a definition designed for goal shifting. This AI is better than human at 99.9% of things, but humans are better at pooping. Therefore we don't yet have a superintelligence.
If that AI were given an identical human body (and interface to that body) to someone who had not yet learned how to do that, and it outperformed them in figuring it out, then that would settle it.
Otherwise I don't see the comparison.
If I'm intelligent enough to use a tool, but I don't have the tool, that doesn't mean anyone who does have the tool is automatically more intelligent than me.
Likewise, comparing my performance without the tool against someone's performance with the tool wouldn't be benchmarking their performance, only benchmarking them with the tool's performance. The fairer comparison would be against me also with the tool.
Lol I feel like no one has any attention span here. Tech shit is expensive in the beginning when it's new. It gets cheaper with time. This is a tech forum, don't we know this? Of course people overreact in both directions on both sides of the issue. It's a very fast technology, wait for things to settle before making grand declarations.
Yeah, but in the short-term there's $600B/yr of debt-financed depreciating capital investments waiting to financially blow up.
If you zoom out to the year 2100, it becomes a little pimple on the economy that is ready to pop, but in the here and now it can cause a lot of damage to real people's wages and finances over the next 3 years.
> Lol I feel like no one has any attention span here. Tech shit is expensive in the beginning when it's new. It gets cheaper with time.
The funniest comment here. Have you seen the prices of the technical shit for the past two years? Dang, GPUs are not getting any cheaper, but more expensive with each year.
brother its been 1 year since claude code released. how fast are you expecting these things to happen? the physical world and hardware are still constraints. someone has to dig shit out of the ground to build these things.
That's an artificially inflated market. OpenAI and xAI bought everything for like two years into the future, partly to inflate the AI bubble, partly to lock-in a monopoly on the kinds of compute you need for AI, and partly to scale up actual operations. They can't realistically keep buying all the RAM in the world forever, the money has to run out eventually (though the market can remain irrational for quite a long time and can keep giving OpenAI and Apartheid Clyde money well past the point of reason).
I don't have a crystal ball, but based on similar historical scenarios, I think that one or two of these companies will win--probably because of some unique application, delivery or trade secret that will drive 80% of their revenue.
I'm sure investors thought one or two of the ISPs laying all that fiber would be collecting fat rents on them until the sun burnt out. I'm glad they got so much in the ground before there was a reckoning. I hope this industry ships more very expensive models, ASAP.
The US govt is going to ban foreign models and foreign providers, and frontier labs are still cooked, because US companies will RLwash Chinese models to try and get in on the captive market. The frontier labs have already lost the war for coding, their next play is custom models for specific domains... Anthropic Galen for biomedical research, Anthropic Locke for legal analysis, etc, and you won't see _ANY_ intermediate work on the model, you will put in query, maybe get some questions fired back during work, and get a "final report."
Eventually the frontier labs will try to cut out the middle man once these models prove themselves and start doing partnerships with big firms in the domains, so they can take a % of the profits in perpetuity rather than just taking a one time payment. For example, after Anthropic Galen, they'll do a partnership with Pfizer to generate Ozempic-Superjacked and take 20% royalties on global sales.
>The frontier labs have already lost the war for coding
This is a delusional take. Sorry, but anyone claiming this hasn't used Fable and compared it to the current best open source models. I see a lot of hype posting about GLM5.2. I see absolutely ZERO people using it in production compared to GPT 5.5 or Opus 4.8.
The people have a right to make and use whatever models they want, protected by the constitution. At a minimum, the models are described in research papers that are unquestionably protected speech. Skilled devs turn those into programs, also protected speech.
Second this, I am also running qwen 3.6 35b Q8 on a 5090 liquid getting around 250 tokens / second and it is plenty capable. I actually haven't even looked at models recently because I am happy with what I have.
And.. now I feel the need to look again. Darn, there goes my afternoon
You can run most open models on cloud hardware. Google Cloud gives you a click to deploy, but then you have saturation / ROI considerations, versus Google serving them up multi-tenant, per-token.
The US government isn’t supposed to be allowed to constrain speech, but they do have the power to constrain commerce, and they can ban the sale of AI services and AI-capable hardware if they choose.
I'm looking at how market/human forces are going to make the game play out when extended to its logical conclusion, not the score on the scoreboard RIGHT now.
I'd guess Anthropic will probably win, and LLMs will probably still be with us and be much better in 10 years time.
But next year we could be in the middle of a massive $600B/yr capital-spending bubble deflating hard with unemployment accelerating towards 10% (or higher).
The internet never failed, but the telcom/dotcom collapse still happened in 2001.
We're seeing the first 20 years of the dot-com cycle, but compressed into two years, and trying hard not to fall into the tar pit of ad-supported services.
Wandering around pretending to be researchers who are only just figuring out how to make money is, for the short term, an incredibly good way to attract a load of naïve money; not all sharks are smart.
And it does, nowadays, give you a bit of a veneer of mere curiosity when you're being accused of massive theft.
Even without doing that the Chinese are already going to impact our labs presence everywhere else in the world. With Fable getting pulled, any model coming out of the US is now unreliable and untrusted. No one in any other country would in their right mind choose OpenAI or Anthropic for anything.
The big push for regulation and export controls is only going to ensure OpenAI & Anthropic are more like the automakers. Only in business because of protectionism, left to screw over US consumers meanwhile the rest of the world gets to enjoy cheap EVs
The issue is the cost is not going to be a hindrance for companies that have gone all in on the AI development. They may still find it cheaper than hiring engineers and if needed they will layoff a few more.
The companies that did not yet jump on this bandwagon and are still evaluating will have a decision to make.
No matter what the AI companies are going to change their pricing strategy and it’s going to become a lot lot more expensive to use. I am just hoping the price stays like this until I am done with my big chunk of work
This article seems to be struggling with telling apart the difference between R&D and operating expenses? The fact that AI companies are extremely unprofitable doesn’t mean they are subsidizing token costs, they still can have very decent gross margins on them
I think a lot of the cost comparisons to employees are off by a factor of 2 or more. AI is the ultimate contractor. Available instantly. Doesn't charge during idle periods. Pre-vetted and pre-trained. No contract negotiations or complex accounting.
That is worth a small multiple of the fully-loaded employee cost. So AI might be easily worth more than $200 per human-equivalent hour. With high utilization, that might be $8000-10000 a month.
With that kind of spend, AI provider financials looks less frightening.
I can't wrap my head around how revenue > COGS but at the same time AI is being subsidized and the real cost is not affordable.
You don't price based on cost, you price based on willingness-to-pay.
So maybe labs are "overcharging" enterprises on interference (because, up til now, enterprises have seemingly had unlimited budget for tokens) and "undercharging" individuals and SMBs (because they don't have an unlimited budget).
I can't go back to a life without AI, and I don't want to. But if AI were billed by token instead of subscription, my monthly cost would probably be ten times what it is now. I could switch to a Chinese model, but I'm not sure how things will look by then.
What makes AI so convenient is how good it is at doing red-team code reviews on my work. I used to need all this unnecessary communication just to get a review, but now I only have to reach out to the people I actually want to talk to.
> Anthropic is subsidizing their enterprise customers by up to 40 times, and OpenAI up to 70 times
might as well be the other way around with non subscribed token being 50x overpriced, or any combination thereof
also uber was non profitable for the longest time, raking up 31b in losses, on the bet of capturing the market worldwide. scale here is different, but it's also 10 years later, with a lot more volatility and floating cash in the market (voo grew 327% over that period, not unreasonable that round size grew on the same trajectory)
The coming AI enshittification is going to be epic. For those of us who have been on the web for more than five minutes, we can see this a mile away.
If you think search ads are annoying, pre-roll YouTube ads are annoying, streaming ads are annoying, or basically ads-on-any-screen-anywhere-at-any-time are annoying, just wait until every stupid thing is powered by AI and is subtly trying to manipulate you to buy/watch/believe some crap all the time.
Jeopardizing a $200/month subscription in return for $1/month in ad revenue seems insane. Using ads on a $20/month subscription to entice you into a $200/month one, OTOH...
Is it not also possible that some of the shift is a consequence of increase of use? While we can be extremely cynical at the finances at play, the lock down and increase of token pricing might be demonstrating a burgeoning demand, which would be a positive indicator.
Yes. If we spend more on building AI infrastructure then current total global gross software sales, the only way the math works is if we create and sell much more software or if we start charging more for it.
The math doesn’t add up and the wheels are starting to come off the bus.
The conversation in a lot of wealth management offices has shifted dramatically in the last few month from “how do I get in on this AI thing?” to “how do I protect my assets when this AI stuff blows up.”
There’s little question now if this will all implode, just when and who’s going to lose their shirt and be left without chairs when the music stops.
What’s playing out now is the scene from The Big Short where the banks wouldn’t mark down the value of bonds until they secured a short position. Once the big money has their helmets on it will stop providing fuel for the bubble and then look out below!
With these confident comments I would appreciate some kind of origin of the information. Not even necessarily a source accessible to me, just: are you in any wealth management offices? Or are you reporting other people's opinions? Or does it just sound right given the spirit of our time?
If that happens the AI companies will first try to negotiate with their creditors and after that likely declare bankruptcy with the creditors taking over what’s left of the assets. Shareholders will be wiped out and employees will be left with nothing. Various franken-companies will emerge from the bankruptcy ashes and the world will move on with AI sans the present irrational exuberance.
> If that happens the AI companies will first try to negotiate with their creditors and after that likely declare bankruptcy with the creditors taking over what’s left of the assets.
Due to the fact that we’ve already done this before (Enron, Global Crossing) -
I’m willing to bet that there are contracts in place ALREADY, that define what happens in the event of a default.
In particular, I’ll bet that the buildings, the GPUs, the patents, etc…
All of these have probably been accounted for.
I worked at a data center that closed during the WorldCom era, and when they put the padlocks on the door, there were still websites “hosted” from the building.
I don’t know if they killed the power or what. I’d cleared out my desk long before they locked it all up. I wouldn’t be surprised to learn that these websites couldn’t get their own servers, since ownership was tied up in the courts.
In the Bay Area during that time, there were row upon row of empty office buildings.
Hence the IPO. Push the risk on to retail and index funds, away from private credit. Plus Microsoft, Google, and Amazon will also be holding the bag and have huge balance sheet write downs, the compute commitments have not yet been paid for it's all just promises.
The banks aren't has exposed this time, as in 2008, most of it is tied up in private credit, its more akin to the fiber buildout in the 90s.
Yes, and private credit investors are rushing for the exits but can’t get out because of withdraw limits. It’s starting to get ugly. Folks owning something they think is going to tank and they can’t sell.
Yep, if they make it to IPO, as SpaceX has, and if they manage to get into several indexes (as SpaceX is already doing, I assume it's already in the Russell indexes, and will soon be in the Nasdaq 100 index), it'll be a bunch of working class people's retirement accounts holding the bag. And, those same companies might be deemed Too Big To Fail, and they'll get even more working class folks money in the form of tax-funded bailouts.
A wealth transfer from the working class to a handful of billionaires bigger than any the world has ever seen (and the world has seen a lot of wealth transfer from the working class to billionaires).
> OpenAI Had $13.07 Billion In Revenue, $34 Billion In Costs and Expenses, and $20.92 Billion In Losses, with a net loss attributable to the company of $38.53 Billion
This is going to be the new most misquoted/misunderstood data of the year, isn't it? The cost is mostly from a one-time accounting situation due to their pivot from a non-profit organization.[0] If we trust the leak [1] OpenAI is likely turning profitable this year.
I don't see any real point being made in (or point of) the article. The author sort of just...dumped a bunch of links with the noise that is so incredibly mainstream at the moment that I doubt any of it is news to anyone even somewhat tracking the AI cycle. Most of it (except for maybe the BLS[1] stat) is just regurgitation.
[1]: And this too is incorrect, should be " the number of jobs displaced would be around 32.5M"
(the post says 32.5K)
How can you lock in when the harnesses are basically thin clients around the APIs and you can replicate them using agents in a short period of time? I haven't seen a compelling thesis yet for how you achieve vendor lock in for LLMs. Claude Code is a bit sticky, but if we're being honest its just because Codex doesn't have all the same features yet.
Luckily the industry is much too wise, after a couple of decades of cloud infrastructure, to willingly opt to make itself entirely dependent on one of two platforms with opaque and complicated pricing. We've learned our lessons, oh yes
This summarizes half of the entire AI scene as these guys generate content to paint the entire world the way like to: US equity markets are facing three IPOs .. each led by a world-class bullshitter”.
The willingness to throw capital at AI is definitely doing some crazy things, but this article has some bad takes on the data.
> [Ratio of per-token cost to subscription cost] means Anthropic is subsidizing their enterprise customers by up to 40 times, and OpenAI up to 70 times
Actually, they could be subsidizing by more (if they are taking a loss on API), or not at all (if they are soaking API customers by a massive margin).
Separately, these subscriptions get sold to large groups with varying usage, so it's crazy to model assuming every subscription is maxed out. Banks, gyms, and many other businesses work this way, offering consumers flexible access to services that they will realistically use in bursts. It's not always worth the complexity to prevent overuse by a small minority. You can feel like this kind of business model isn't as transparent, but it's silly to pretend it can't work.
> OpenAI spent 44% of their revenue [$5.3B] on sales and marketing! The hype needed to keep the AI bubble inflated is incredibly expensive.
Over that same period (2025), OpenAI added $10B in realized revenue and $14B in run-rate. Sounds like they're getting >2X return within 12 months of those go-to-market dollars. Compare that to like, any other business.
> Thus in recent weeks the idea that Generative AI (LLMs for short) is too expensive has been all over mainstream business media.
Would it be smarter for these companies never to test customers' price tolerance? The quotes following this make it seem like the companies are getting important information about the nature of that price tolerance, and preparing to react. This is the work markets do on both sides to understand the value of a new product.
There are lots of good arguments about AI overinflation, but in order for them to be useful, they have to be rigorous and targeted.
I know a lot of level-headed engineers here may not side with me, but I say let the companies who abandoned their people at the drop of a hat, with CEOs who waved their flag around on social media, proudly declaring how they'd now run their companies with 75% fewer employees wither and die. If I had been let go, there's no way I'd go back to a company like that, and there should be a black list of CEOs who acted this way established and kept public. These CEOs are not holistic thinkers, and are too susceptible to mass hysteria and too irresponsible to real people and their lives to be trusted with the vision for any company ever again.
GM just did this in the last 30 days [1], and their sales are likely going to be just fine. In fact the auto industry has repeatedly automated jobs over the last 100 years, and they still make decent sales numbers.
If you decided to boycott every company that replaced staff with automation, you would be forced to exit the economy. Every company does this to some degree and the customers who vote with their wallet do not seem to care about a reduction in force.
Robots that replace auto industry factory workers exist; the CEO of GM didn't imagine them as part of some sort of business media induced psychotic episode.
The same is not true for the software industry execs.
> GM just did this in the last 30 days [1], and their sales are likely going to be just fine. In fact the auto industry has repeatedly automated jobs over the last 100 years, and they still make decent sales numbers.
I worked at Verizon during their layoffs last year. Biggest layoffs in the USA.
As someone who’s been laid off before, I knew that it generally boosts the stock price.
I bought VZ because of that. It’s up 15% since the layoffs.
Microsoft, an AI stock, is down 30% in the same timeframe.
This is true, and I'm sure AI cuts will continue, but it's obvious that the ones who went "all in" at AI's mass introduction were drinking a special kind of Kool-Aid reserved for the truly sycophantic Wall Street lap dogs, not the CEOs who think about risk and are cautious about betting the farm on a relatively new and mostly untested technology. GM is over 100 years old, and no doubt released improvements that were well-tested and predictable, because you don't take massive chances with a company that well established. It was a couple years into the mass AI deployment that studies on the minimal overall productivity gains of AI even started to come out(!) This was "get on the bandwagon" thinking at a massive scale, which shows you how many CEOs are not independent thinkers at all, but are really just followers. Yes, use AI, but do it responsibly, never forgetting that your investors aren't your only stakeholders - so are your people.
Is deleting a letter after an LLM generated the article an insurmountable task? These quaint signals only screen out the lowest of effort slop writers. Better than absolutely nothing, but barely.
It does remind me of the time a chef told me when he puts lemon juice over a dish, he would intentionally not remove any seeds that went on it because it was a signal of quality. I wonder if future slop chefs will intentionally place seeds on dishes that came from a box...
...and "maiinstream" -- seeing glaring typos (easily caught by spellcheck) now makes me wonder: did they decide to leave them in (or add them explicitly) to signal they didn't use AI to write, or (the more paranoid option) did they tell the LLM to add a few typos...
I didn't get the sense this was LLM-written, but typo-signalling is... I donno a bit weird. Firefox is underlining some of the words as I write. I'm leaving "donno" unchanged even though it's flagging it as a misspelling but I suppose I'd still opt to fix something like "maiinstream" even at the risk of potentially seeming more LLM-ish!
It's funny when you watch the doomscroll all these anthropic guys talking about how you should be writing self-improving loops and that's all they do. Of course that's all they do, they don't have to pay for their tokens.
Can confirm, my experience in “loop engineering” was “this is neat” for 45 minutes until a daily ration of tokens was evaporated. The quadratic cost trap is prohibitive to experimentation.
As a localLLM evangelist, I am hopeful this will bring more attention to the joys of rolling your own sovereign AI.
Yeah, i'm hoping that gets smoother. I've been experimenting with omlx and opencode on my m5x64gb and keep running into issues w/ Qwen3.6-35B-A3B-MLX-8bit exceeding it's memory limit at the most inopportune times. Playing with 12B gemma4 (8bit) more today.
Maybe I should be aiming for something targeting 48gb of memory?
I really can’t stand when writers point to the difference in price per token on the api and subscription and use that as evidence that inference loses money. This author even says it’s implausible that the api charges 4x marginal cost when I think it’s very likely even higher than that. The entire rest of the post sits on this faulty assumption. Fixed costs don’t matter when marginal revenue is profitable and growing rapidly. The ai labs only have 2 questions. Can they prevent users from switching to open source models? Can they scale the number of users on enterprise plans the way they did for coding but in a more general way for all knowledge jobs?
agreed, this doesnt even account for prompt caching or the fact that anthropic has substantial proprietary efficiencies on their inference stack specific to their models and scale.
> Can they scale the number of users on enterprise plans the way they did for coding but in a more general way for all knowledge jobs?
Do these knowledge jobs have a significant corpus of not only knowledge but discussion and problem solving, all conveniently labelled for the AI to train on? Probably not. Coding has stack overflow, what does, say, advertising use?
I agree this is a hard problem for the labs. I would be hesitant about “probably not” though. There is just as much marketing copy floating around as there is coding training data. I struggle a bit in this question because I’ve only ever worked as a software engineer, so I can’t exactly make claims about all the work other jobs do. But, one example is I was talking to a doctor friend of mine the other day. He was talking about how he had to take his recertification exam recently and put the questions into chatGPT and thought it gave answers that were generally more thoughtful and correct than his own. Does that mean doctors are done? Of course not, but he’s now pushing hard for more ai tool use in his practice.
> Coding has stack overflow, what does, say, advertising use?
Advertising has centuries of print ads, 100 years of radio advertising, 70 years of TV commercials, etc. And modern AI does not necessarily need labeling.
The article fails to mention DeepSeek, Alibaba, Qwen, Xiaomi, MiMo, z.ai, or GLM. It's hard to take such an article seriously that doesn't do this. (Our monthly total spend is around $180 with a team of 6, about half technical; our biggest line items are for American models or subscriptions which we probably will be planning to get rid of.)
And then remarks like this:
Anthropic, OpenAI and Microsoft have all now transitioned customers from subscriptions to token-based pricing.
Huh? I use OpenAI via a subscription, as is anyone else using GPT-5.5-Pro who isn't a multimillionaire.
They're referring to Enterprise customers, though should have been clear about it. Enterprise plans on Claude for example no longer include any baseline tokens. It's 100% usage based pricing.
True, but my friends in Enterprise still just purchase Claude Code subs and expense them. They basically get an allowance of $500 or so per month to buy various tools, and of course are banned from Chinese models. (Claude, Codex, Antigravity allowed, basically.)
> Our monthly total spend is around $180 with a team of 6, about half technical; our biggest line items are for American models or subscriptions which we probably will be planning to get rid of.)
Please tell more :). Do you pay per token from bedrock / openrouter / somewhere else? How many tokens you use over the month, and how many for each task? Which harnesses?
Pay for DeepSeek directly. One developer insists on having his own account and in theory expenses it, but he forgets to turn in $10 expense reports. (Total spend in last two months = about $45.)
Pay for OpenAI Pro directly, but I’m the only guy that uses Codex. $100 a month. My nontechnical partner likes to talk to ChatGPT 5.5 Pro for image related tasks (think generating interior decorating pics).
The nontechnical staff use a Gemini account on a Google family AI Pro sub. I use Antigravity when working on Android or Google Cloud API codebases.
Everyone gets OpenCode Go. The cost is trivial. $10 a month per person.
Pay for MiMo directly. We use it during Chinese off peak hours though. Total spend so far $25 in last month.
We run a few Qwen models locally and pretty much have them pegged all day. RTX 5090 on a PC and a Mac Studio.
There’s also Grok which is used for Imagine for artistic / graphic design related work. I also use the subscription for a vision model in my oh-my-pi harness.
We’re having discussions about how to pull in GLM-5.2 cost effectively. We compete with third world development shops so we can’t really pass on inference costs, but we can benefit from getting jobs done for customers faster. But ⅔ of our work is either internal or open source projects we can’t bill for.
Not the GP, but I use Opus for planning, Deepseek for actual coding (implementing the plan) and GPT for review. GPT is inexhaustible on the $20/mo plan, Deepseek is dirt cheap (maybe $10/mo) and Claude is Claude.
Subsidizing makes no sense when there's no - possibility of a - moat. Although it's very possible that China in general subsidizes Chinese labs in some way so they maintain pressure on US labs. But you only have to look at proxies such as OpenRouter to see that the individuals aren't doing any subsidizing on per token costs.
Most of the "affordability" and "pricing" discussion is pointless because we don't have any real numbers on their margins per token. So, yes, they are subsidizing their subscription plans compared to the API prices, but the API prices could already be stupidly inflated, so the relative price comparison is a nothing burger.
Until we know (or at least get a hint) on their margins on API prices, any pricing discussion is pointless.
We have a pretty good idea of how much it costs to serve these models. You can pencil out the economics and guess at the model sizes and we know pretty decently how expensive the hardware is.
This like claiming it's meaningless to guess the margins of a restaurant without going into their books and seeing the exact recipets and recipes.
They ain't doing dark arts in the back. You can guess at what goes into the food based on similar recipies and how much that costs based on what you pay at the grocery store.
This doesn’t solve the problem because (tautologically) the more AI prices go down the less money the companies make. If right now today the companies are operating at a profit and a price war causes the API costs to sink 90% next year, and their capex amortization costs stay fixed.
AI prices going down means the models are improving, particularly from the efficiency angle (which is inevitable, given the nature of tech). That means all they have to do is maintain a large enough customer base at a rate high enough to ensure loss decreases continuously over time, until eventually the pass the point where they're just gaining. Healthy competition ensures that improvement savings are actually passed on to users in a measured manner, so they don't become too greedy in trying to get to and increase gains.
But now you’re describing a commodity, and the competition will erode profits, and their valuations are bananas, unless someone can find a business model that truly differentiate and creates a moat.
Is this really for comparable models though? Will folks at scale continue to choose Anthropic frontier AI model if OpenAI releases a similar generation at a 90% discount with comparable capabilities? It feels like the fungibility assumes delineation by capability _and_ cost. No one is choosing sonnet over opus at similar price points.
This doesn't really tell you anything useful. AI companies have both built huge datacenters and raised a colossal amount of money. Include caching, quantization and etc. All of those would allow them to undercut on price considerably, even more so if you count in all the users who don't actually cap out their plans. Prices going down doesn't really tell you anything about the production cost, especially in a market where every major participant is happy to burn money just for the marketshare.
There are many research avenues which are open which reduces cost dramatically. Smaller task specific/ language specific/ domain specific models, in fact they could even be better. The earlier computers were the size of a building. So prediction based on current state into the unknown future possibilites is wrong. The hardware will be all the more valuable if cheaper ways to run become possible. The hardware gets cornered in a sense.
Because of it's unpredictability and massive dependence on the training data, when LLMs start hallucinating most of the time the only fix these "engineers" have is to feed it another LLM... The genius was the transformer architecture, and evidently none of us have a damn clue how it works
Every 6-12 months or so we get an increase in one or more of things like: compute power, compute efficiency, GPU power, GPU efficiency, network bandwidth increase, memory speed increase, component density increase in the same form factor, etc.
For awhile it was every 2-3 years you'd start a hardware refresh. As companies moved into more and more training, this timeframe started to shrink. It went from 36 months to 24 months. From 24 months to around 16-18 months. Last I checked last year, it was at 12 months. I think things may have slowed because of component availability, but otherwise whole data centers would be 6-12 months into full operations before they would start a refresh cycle.
Not to mention the massive increase in power density demand and cooling demand per rack that entails.
So no, "AI costs" have not gone down, in fact they are more expensive on training AND inference than ever.
This is why many are concerned about the heroin drip of api costs into orgs. For the companies that are public, look into their financials. It's gonna hit companies and high volume users like a ton of bricks.
I'm no economist but if true don't you have the opposite problem? How do you get people to need X many tokens per day such that you can sell enough to make money? Wouldn't you need an absence of competition for that to be ok?
the demand for intelligence is infinite. you sound like someone in 1960 wondering what the hell we would even do with the functionally infinite cpu cycles we have available to us now.
The quality of what you can get from DeepSeek V4 Pro for $10 is light years ahead of what you could get for $20 a year ago.
Likewise, the quality of what I can get from a local model like Qwen 3.6 on an RTX 5090 is light years ahead of what I could get a year ago on the same hardware.
This is just hand waving on the obvious consensus that cost per capability is going down. There’s no doubt about it. Hell you can run a Gemma 4 model on your laptop that mogs GPT 4. But yeah you can use fuzziness as an excuse and ignore the trend.
It's not an affordability crisis, it's a financial crisis. The models get cheaper super fast. By this time next year Fable 5 will cost less than Sonnet does today. That's not the problem. The problem is that many companies are going to realize that they don't get any ROI from AI. Generating code faster != more profit. Most of the Fortune 500 will likely realize this and then the token budgets will come crashing down. Most of their ideas are _bad_ ideas. Moving faster, won't lead to more profit. Sure, you can use AI to potentially replace software engineers, but the F500 are also terrified of not having accountability or making mistakes. They won't be firing any engineers.
I think the biggest problem is not necessarily the cost to develop & serve the models, but how quickly user behavior changed with token based pricing.
I know a lot of people at companies where the marching orders changed on a dime end of Q1/start of Q2. These are shops that were fully on the "use AI or die (because we will fire you)" train.
Now there's monitoring, reporting, alerting not just on overall cost but on "over-use" of best/priciest models based on total-or-percent tokens/dollars, etc. All of this comes with direct developer engagement & standardized management escalation for holding it wrong.
To me this customer behavior does not smell like a product you can 10x the pricing on to get profitable. We have exited the exploration phase and now ROI matters.
I can give you some additional anecdotal evidence to support your comment.
I work at a Fortune 200 company. At first, it was the Wild West. Need an LLM? You got it. Need to or want to build an army of agents? Done and done. We literally had everything at the tips of fingers for about 3 months. Teams were building their own internal tools, the team I work on canceled contracts with several software vendors because teams were building the same tools for what they thought was nothing.
Then they signed contracts with Anthropic and Google because I would assume they saw the token usage was through the roof. One month later? They completely cut off access to everybody for both Claude and Gemini. If you wanted access? Suddenly it was several forms, along with several approvals and a rock solid business case why you needed it. And before you got to the forms? You were added to a waiting list that was thousands of people long.
The entire company is now in damage control after trying to get the genie back in the bottle. I'm guessing someone saw how much we would be paying for the tokens we'd been using and decided to shut the party down so to speak.
Was there at least performance gains to be measured?
AFAIK nobody was collecting analytics. The one team I was working on had put out a goal of "30% more efficient" using AI tools. Its about as subject as you can get. We never got around to what exactly that meant before everything got shut down.
Myself and several other devs were laughing about the whole thing. The company was so amped about what AI could do they never even bothered collecting any analytics that would affirm or deny any of this had a positive impact. Even some of my team members were talking about the placebo effect AI has had on a lot of C-Suite folks.
I can second this. Our company and department was all-in on AI. And since the token-based pricing came in, we got an email from IT that tried to explain that most developers don't know how to choose models and that the cheap models should be good enough for most of our work ..
Have they built an internal ai enablement team?
Yes :D
I do a lot of client work for fortune 100’s.
Over the last month I have seen companies scrambling to measure deliverables against cost. Most of the back room talk is to the affect of giving devs a small allowance ($500 a month) and then making them prove their own productivity increases (again, based on deliverables, not LoC) before they either take it away or give them more.
Obviously this won’t be on an individual basis but some kind of unit.
Either way, with how much I see these companies cutting back I have no idea how the big AI companies are going to be profitable.
I.e., the demand for programming tokens turns out to be quite elastic.
I would imagine it only gets worse in the face of good-enough open/chinese/local models too right?
Microsoft adding Deepseek support already as I recall?
That is - for any definition of "they are behind X months" then eventually they get to the point Claude was in January when the world freaked out, but at 1/10th the cost. A lot of firms are going to mandate that is good enough for their developers.
> Microsoft adding Deepseek support
I believe this hasn't been confirmed yet but I think it speaks to a bigger problem for the AI companies which is, if you give capable developers a good reasoning LLM, they can make it work like it was a really expensive model.
I believe we are 100% at the stage of good enough for the vast majority of tech companines. Fable and others will be more valuable for non-traditional tech companies.
I read somewhere that the chinese AI companies are sharing knowledge and it would not surprise me if the government is applying pressure by saying work together or else. If they work together, they can truly commoditize LLMs and with China ramping up hardware support for AI, I see the future being inference speed and hardware being the moat.
If hardware becomes the moat, the US frontier labs are screwed. We have AWS, Azure, GCP. All three have or are making inference silicon. LLMs become just another service in the public cloud's large service catalog, and open weight wins.
Which makes sense to me. Selling a chatbot interface/model access to the general public was never going to be a viable long term play. You still need developers to wrap the models into specialized tools. Queue the Jobs quote "It's a feature, not a product."
The big thing is, the western world has moved so much of the manufacturing to China and think a lot of people will not forgive Samsung and others, so I can see China owning a good portion of the supply chain.
> The big thing is, the western world has moved so much of the manufacturing to China
I built my career on Solaris and it got rugpulled by Linux.
That wasn’t because of software, it was because of hardware. Linux’s cost advantage existed because Sun hardware had huge margins, because their software was basically free.
AI will probably be a repeat of this. Whoever can come up with the hardware solution that minimizes the cost per token will win.
I believe the 5090 still holds this crown, but someone certainly knows better than I do.
but not all tokens are equal and vertical integration is the name of the game. Solaris did not lose to Linux, it lost to the LAMP stack on commodity x86 hardware. without the "AMP" part, Linux would've been dead in the water.
100%. There will be strict quotas on the expensive models and day to day work will be done on the cheap models that are "good enough" with escalation to the metered models when the cheaper options are spinning their wheels. Eventually the US frontier lab APIs will only get the most heavily triaged work that multiple tiers of cheaper Chinese open weight models have failed on.
And of course the C-suite will have unlimited access to Mythos tier models, which they'll use to summarize reports, while passing down mandates to rank and file to increase usage of less expensive models.
Yup, we are in the process of getting access to US hosted Chinese models. I've been petitioning Google and our rep, we will see but I suspect they will cave eventually. Gemini sucks and if they don't sell what their customers want, we go shopping around.
If folks won't pay a higher price, doesn't that mean it's inelastic?
"Elastic" in economics happens to refers to how elastic the supply/demand is when the price changes (not vice versa, as you're describing). So e.g. an inelastic demand means the quantity demanded changes very little when the price doubles.
Elastic demand means buyers are highly sensitive; a price hike causes a massive drop in purchases. Inelastic demand means buyers aren’t very sensitive; they keep buying regardless of price
Ah alright I have it backwards then.
> Zitron's numbers don't tell us the real cost of generating tokens but, subject to the assumption that the platforms are not subsidizing the token price, that means Anthropic is subsidizing their enterprise customers by up to 40 times, and OpenAI up to 70 times
Neither Anthropic nor OpenAI are subsidizing enterprise customers. Neither Anthropic nor OpenAI allow Business nor Enterprise customers access to the high value $200/mo plan. Both organizations have moved to a "cheaper plan per user + API Pricing after that" (e.g. $20/mo + usage). The $100/$200/mo plans are for individuals only (of course, many individuals use these plans at work, but that's beside the point; they aren't selling this plan to enterprises).
> SemiAnalysis also analyzed the platform's gross margins, implausibly assuming that tokens were priced at 4 times the cost of generating them and: With the current subsidies, all it takes for a user to have a gross margin of at best negative 25% is for them to use as little as 25% of their rate limit.
The article's source for this claim is not SemiAnalysis; its Zitron. But once you dig through his article, Zitron links to a SemiAnalysis tweet [1] where they, as the paragraph states, implausibly assume gross margins of 75% to come up with their weird analysis of the subscription plans. Citing this for anything is weird, because afaik that 75% number is a total shot in the dark. We have no clue what their margins are. My take is that the only reason that 75% number is implausible is because it may underestimate the inference margins of Ant/OAI's API pricing.
[1] https://x.com/SemiAnalysis_/status/2064815045767213400?ref=w...
> it may underestimate the inference margins of Ant/OAI's API pricing.
If true then why are neither Anthropic or OpenAI dropping their API pricing to gain market share when both are clearly doing all sorts of political and PR maneuvering to compete in a cutthroat market?
Since they aren't dropping the API usage prices (and are in fact raising them in a lot of subtle ways) then one of these options almost has to be true: they are still subsidizing inference, training costs are so ridiculously high that they need to make huge profits off inference or collapse in on themselves, or they are price fixing.
The training costs are very likely the reason. Dario has talked about how each individual model is profitable, but how the expenditure training the next generation of models makes it look like they're not profitable at any given moment in time, and I believe he's being honest about that.
The market for open weight model hosting gives you an idea of the profitable price floor, it's pretty clear there's markup baked into OAI/Anthropic's APIs.
If true then why are neither Anthropic or OpenAI dropping their API pricing
They are? In the before times of 2025, Opus 4.1 was $75 per million tokens. Opus 4.8 is $25, and Fable is/was $50.
Given my experience with hosting these models at scale, working and optimizing load, I don't think the margins are nearly as high as 75% if the models are as big as people often claim.
Only reason deepseek is so cheap is because well I don't know, but actual pricing should be around their initial price which was 4x, at that price you have a healthy 25-50% margin based on occupancy, given the deepseek v4 is a very sparse moe model.
GLM 5.2 for example doesn't have more than 30-50% margins that's assuming old pricing for GPUs, current inflated GPU pricing well I am certain the margins must be lower. Ofc you can host for cheaper with quantization, and if you have very consistent capacity/utilization, which is not the norm with AI workloads.
Overall for large models like GPT 5.5 or Opus there must be healthier margins of around 50-70% assuming GPU pricing didn't increase for these companies. Even if it did 30-40% margin should be possible, even in worst case assuming all GPU they had saw a jump in pricing.
For smaller models it's hard to say, I would guess 20% but these models might be much smaller than I suspect, then it might be double that.
Note the issue is less intelligent tokens don't linearly scale down in memory usage, which is the biggest pain point of serving models. Context sizes have fucked us all.
Also anyone claiming OAI makes less margins on APIs or stuff might be wrong given they are on much lower context size, 1M context definitely is a lot more expensive to serve especially with smaller models like sonnet.
My take is that Anthropic and OpenAI simply are NOT competing on price. 2 big players are often not enough to create tension on price.
Chinese models and open model providers are, indeed, competing on price, and the difference shows.
How are Anthropic and OpenAI going to compete on price when they're both already deeply unprofitable?
Anthropic just announced it's on track to have its first profitable quarter: https://www.wsj.com/tech/ai/mind-blowing-growth-is-about-to-...
Response: https://www.wheresyoured.at/anthropics-profitability-swindle...
Serving the API is profitable. They are unprofitable because of R&D (and maybe subscription costs?). If they can continue to find access to R&D capital, there is space to reduce API costs.
Nuclear energy is really cheap too... as long as you ignore CapEx, would you like to invest?
how do you have access to their financials? are you an insider?
I'm curious why you didn't pose this question to the grandparent commenter, who first asserted the opposite?
There is no moat until a company achieves RSI and/or AGI, and the one that does succeed in moat-making will do so by hacking into and destroying their competitor's infrastructure.
Once moat is achieved, you don't have to compete on price. Of course it'll be academic because the AI will probably destroy all of us.
They may not be able to! It's pretty widely acknowledged, for example, that if there's some surprising plateau hiding around the corner they're both going to fail. But that could mean that they're overcharging for AI usage to get research money and sustainable rates are lower rather than higher.
I think that for coding we're past the plateau issue. The frontier models of today are good enough and very valuable. The expensiveness in running them will eventually be solved by cheaper faster hardware.
I do hope that a day will come where you can buy the nvidia spark thingy for 5k that can run the equivalent of Opus 4.6 or 4.5 locally and that would be a massive thing.
> The expensiveness in running them will eventually be solved by cheaper faster hardware.
How?
* Moores Law is almost over. The 5090 improves over the 4090 mostly because of quant improvements.
* even if the hardware improves, there’s a huge incentive to slow roll the next generation. Nobody wants to end up like Sun Microsystems. Sun’s used hardware was faster than its new hardware, once you considered price. Sun ended up competing with its own used equipment.
The most obvious place for improvement is RAM, network and storage.
If someone can bring more RAM onto the market, that will unstick things.
GPUs are not really the ideal architecture for running neural networks; they are heavily bottlenecked by memory bandwidth and struggle to keep all their tensor cores supplied with data.
There is significant room to make more specialized neural network accelerators with new compute-in-memory architectures.
If the brain can run 86 billion neurons on 30W it must be possible.
The whole hidden plateau hypothesis is kinda bunk, because we're already pretty far in a plateau for general knowledge/question answering, but there are many subdomains where we can push model capabilities, and as we saturate one subdomain we can just shift to another economically valuable one.
There isn't one AI intelligence S curve, there are thousands of them, and they're mostly invisible in the major benchmarks, but for someone trying to do work in that specific area of capability, the progress is transformative.
I'm skeptical of a hidden plateau, but I really think it's overconfident to assume there's not one. Remember that it doesn't even have to be a technical plateau; the effective plateau of e.g. car speeds is determined by regulations and road conditions, and far below what "frontier cars" are capable of on a controlled racetrack.
That’s the scenario where we’ll all be using Chinese models
1 player is enough to create tension on price when "don't buy it at all" is a comptetative option. By most accounts, Anthropic and OpenAI both lose to "just don't buy" when they try charging at cost.
Chinese models are dropping in price thanks to ridiculous levels of state subsidy where companies are forced into aggressive price wars to survive and grab market share. I am guessing this will also blow up sometime next year or in 2029 at the maximum.
Btw, some Chinese corporates have already seen this and increased their price. Zhipu AI & Tencent for example. Alibaba, Baidu, and Tencent also announced multiple price increases for their AI services.
This is in contrast to American models which receive _ridiculous_ levels of private subsidy.
China has the benefit of vast solar power and rapidly increasing battery capacity. Yes, that's subsidized, but it pays for itself in the long run.
And, even with the price increases, Z.ai and Tencent are still much cheaper than Anthropic or OpenAI models. I think there's an efficiency focus among the Chinese models that is absent at OpenAI and Anthropic, and in the end I suspect efficiency will be the winning feature. Google seems to understand that. Gemini 3.5 Flash is pretty competitive with the big guys, and it's small enough for Google to run it profitably (I assume) for a price that's much less than the frontier models. Gemma 4 models are showing off a bunch of efficiency techniques (MTP, QAT, the 12B encoder-less vision model that soundly outperforms much larger vision models, DiffusionGemma), and I assume they have several more techniques that aren't published.
Chinese companies like Deepseek are operating on shoestring budgets (allegedly less than 300 employees at Chinese wages). It’s not that self evident there is anything that needs subsidized besides compute (due to limited manufacturing capacity and access to Western chips in China)
Shouldn't we know a better answer to these questions once Anthropic's IPO materials surface publicly? I understand, and maybe even expect, SpaceX's materials to be all over the place and skate on by any discussion of unit economics, but the nerds over at Anthropic might just be forthright enough to just tell us what their margin is on tokens as part of their IPO.
To be honest, making sense of finances of fully public companies is often hard, because in practice, accounting is hard. How you account for depreciacion, cost, investment, fixed vs marginal costs is in practice fluid, companies have an incentive to make it look attractive, while also optimising for tax and shifting revenue around to narrowly beat analyst recommendations.
Here's a concrete example. Does some random AI company make operating profit on inference? I.e. if you only kept marginal costs, would you make a profit?
Well, depends what you account as your costs. If you're using hand-me-down hardware from previous generation's training, how much do you charge yourself internally for it? Maybe you show less, so investors take solace in profitable inference, even if you're losing money overall. How exactly are you accounting for electricity costs between training and inference? Is your army of SREs mostly servicing training new models (R&D expenditure) or inference (operating cost)?
This even has a name, and is called the "big bath" approach. If investors expect one part of your business to be a fiscal black hole, just shove all your costs there. They are accepting of it, and you make the rest of the business look better.
I'm not accusing AI companies of cooking the books, rather I'm trying to highlight you could see all the cash flows and still not know how much money is made or lost where.
I saw some commentary that their free cash flow is misleading because it doesn't subtract the stock compensation they are paying to attract / keep top AI talent. Their point was also that deciphering financial statements is hard
Why would it? Stock compensation doesn't affect cash flow, it just dilutes the shareholders.
Except that's the thing, they do stock buybacks so they do not dilute existing shareholders or lower stock prices.
This is the video I watched that explained the shenanigans (from the guests' perspective, not illegal, obfuscated)
https://www.youtube.com/watch?v=YrJzjC4kKCY
Well it probably doesn't help that Dario is going around on podcasts saying things like "frontier labs need $1T of revenue or they will go bankrupt" lol.
Dario’s company may be creating super intelligence that will kill us all in the near future, but at least he seems to be brutally honest about all of it.
The irony in AI triggering societal collapse due to gross economic malfeasance is just fun to think about.
If AI was around in the early 2000s Countrywide.ai would have been a thing.
Which is just a flashy way to say "we have low margins and lots of overhead".
Considering how much they spend on sales, marketing and R&D that doesn't sound that absurd
My point is that $1T of revenue is A LOT. Apple & Google each only did $400B revenue in 2025. Facebook did $200B. Think of how many decades it took the 3 to get there.
So depending on how literally we interpret Darios comment, OpenAI & Anthropic need to get to Apple+Google+Meta revenue numbers in like single digit years?
The estimate that AI companies need to replace 27% of jobs to service their debt is interesting. But at least Anthropic and Meta seem to have their eyes on replacing software engineers.
There are ~1.6M software engineers on the US [0], earning a bit under 150k/year on average [1]. If AI companies captured all of that spend, that amounts to about 250B/year. The article assumed that they need around 300B/year to keep up with their debt.
At least based on Meta's recent behavior, forcing 30-50% of developers to switch to data labeling, it looks like that is actually their game plan.
[0] https://en.wikipedia.org/wiki/Software_engineering_demograph...
[1] https://www.indeed.com/career/software-engineer/salaries
The unit economics might be just fine. We'll know more after IPO.
The drug dealer analogy has a darker side to it, however.
Once your dependent, they can drive up the price just because. It doesn't need to be for existential reasons.
Once your dependent, they can drive up the price just because. It doesn't need to be for existential reasons.
This is the crisis point for vibe-coders. A developer can go back to writing code by hand, as horrible as that might sound. Someone who hasn't learned to code but builds with AI can't go back. They either pay or they stop. That will be an painful choice whichever way you fall.
There are already open weight models out there that are capable and cheap enough for a lot of coding tasks. Not as good as Claude but not far from it. There's no going back to pre-AI coding.
I can't speak for everyone, but for most of my coding tasks, Claude is just barely good enough. There's no going all the way back, and perhaps open weight models will keep improving, but at least 50% of my work would be better done by hand than by a worse-than-Claude agent.
I consider Opus 4.5 the crossover point where coding with agents got more efficient than not coding with agents. They were too stupid before that, and wasted more time than they saved for anything beyond a basic CRUD app or HTML page.
Certainly, the best models have gotten better since then, but I wouldn't consider DeepSeek V4 Pro or GLM 5.2 to be a big enough downgrade to be worse than coding by hand. I'm willing to spend a premium for the best model for coding because it wastes less of my time with dumb stuff, so I've got a Claude subscription. But, there is a limit to how much of a premium I'll pay. 10x over Chinese models? OK, fine. Opus saves me enough time to make it worth a couple hundred bucks a month. But, 100x, or more? Nah. I'll go a little slower, review the PRs a little more carefully.
And, open weights models do keep improving. DeepSeek V4 Pro is a notable improvement over earlier DeepSeek models, and the first DeepSeek model to cross the "better to work with it than without it" threshold into Opus 4.5 (or better) territory. GLM 5.2 is somewhere in the ballpark of Opus 4.6 (though without vision, a notable limitation for anything that requires a UI).
There's a secret third option: learn. At one point, all of us were "nontechnical", but we learned. The trick is to never stop.
Is it? Learning is one thing. But owning a large codebase, you see for the first time, is a completely different level.
Yeah giving up is totally a viable choice, but it isn't the only option.
All of the silent, hidden model routing OpenAI does strongly suggests that the unit economics are not just fine, at least not yet.
If apparently the only way you can make money with your product this early is to dilute and adulterate it behind the scenes, it strongly suggests you want the customer to continue to believe they are getting value that you can't afford to supply.
More prosaically: if either of these firms could prove that they were even really close to profitable on inference, they would have bloomin' said so while they were trying to raise more money.
The dependent idea is questionable- when your boss tells you to not use the most expesive models-you just dont
I would assume when price hikes happen either 1) less non technical people would vibecode as it doesnt impact the work that much 2) people use the cheaper chinese models 3)we're jamming ai into everything because were exploring. We will just niche down into use cases that provide high roi
AI is a worker for me. That i pay for. Basically i am in the same game now to reduce the prizes i have to pay for my workers. Just like the employers are, that seek to reduce costs for employees, as we are simply too expensive. We need more competition among the workers. Let's introduce more chinese workforce! ;)
If you had a choice of maybe 3-4 contracting firms to hire workers from and you weren't large enough to negotiate on price I think you'd be in a pet bad spot as a business?
I'm finding it challenging to believe they wouldn't just cannibalize anything dependent on them in that way or at minimum launch a directly competing product.
It's a really different market, though. New entrants can easily undercut them if they price too high
> Sales and Marketing: $5.73 billion .. That is, OpenAI spent 44% of their revenue on sales and marketing!
Anyone know what they are spending this on? Can't remember seeing one OpenAI ad.. Is it just pr and influencers? Ads in the US?
Once locals get to Opus levels I think it we may see a phase change because that + a reasonably competent programmer is going to be a very powerful combination for most practical programming problems.
Frontier models may eventually achieve super-intelligence (no opinion beyond mild skepticism) but super-intelligence isn't necessary for most practical day-to-day programming. The problems, as always, become communication, understanding what users really need, etc. that is, softer skills.
> Frontier models may eventually achieve super-intelligence but super-intelligence isn't necessary for most practical day-to-day programming
I think you forgot what super-intelligence means…
tbh, not sure i ever understood it
A superintelligence is one that exceeds human intelligence in all areas. Which roughly translates to learning, adapting, and performing more quickly and efficiently than even the best humans. This is closely related to "the singularity", which is when technological growth becomes uncontrollable by humanity.
That sounds like a definition designed for goal shifting. This AI is better than human at 99.9% of things, but humans are better at pooping. Therefore we don't yet have a superintelligence.
If that AI were given an identical human body (and interface to that body) to someone who had not yet learned how to do that, and it outperformed them in figuring it out, then that would settle it.
Otherwise I don't see the comparison.
If I'm intelligent enough to use a tool, but I don't have the tool, that doesn't mean anyone who does have the tool is automatically more intelligent than me.
Likewise, comparing my performance without the tool against someone's performance with the tool wouldn't be benchmarking their performance, only benchmarking them with the tool's performance. The fairer comparison would be against me also with the tool.
Lol I feel like no one has any attention span here. Tech shit is expensive in the beginning when it's new. It gets cheaper with time. This is a tech forum, don't we know this? Of course people overreact in both directions on both sides of the issue. It's a very fast technology, wait for things to settle before making grand declarations.
Yeah, but in the short-term there's $600B/yr of debt-financed depreciating capital investments waiting to financially blow up.
If you zoom out to the year 2100, it becomes a little pimple on the economy that is ready to pop, but in the here and now it can cause a lot of damage to real people's wages and finances over the next 3 years.
Lots of stuff in the zirp era was cheap when it was new and increased in price over time though. Look at grubhub fees or etc.
> Lol I feel like no one has any attention span here. Tech shit is expensive in the beginning when it's new. It gets cheaper with time.
The funniest comment here. Have you seen the prices of the technical shit for the past two years? Dang, GPUs are not getting any cheaper, but more expensive with each year.
brother its been 1 year since claude code released. how fast are you expecting these things to happen? the physical world and hardware are still constraints. someone has to dig shit out of the ground to build these things.
That's an artificially inflated market. OpenAI and xAI bought everything for like two years into the future, partly to inflate the AI bubble, partly to lock-in a monopoly on the kinds of compute you need for AI, and partly to scale up actual operations. They can't realistically keep buying all the RAM in the world forever, the money has to run out eventually (though the market can remain irrational for quite a long time and can keep giving OpenAI and Apartheid Clyde money well past the point of reason).
I don't have a crystal ball, but based on similar historical scenarios, I think that one or two of these companies will win--probably because of some unique application, delivery or trade secret that will drive 80% of their revenue.
Consider Google, Apple, Amazon, etc.
It's still early days...
I'm sure investors thought one or two of the ISPs laying all that fiber would be collecting fat rents on them until the sun burnt out. I'm glad they got so much in the ground before there was a reckoning. I hope this industry ships more very expensive models, ASAP.
The US govt is going to ban foreign models and foreign providers, and frontier labs are still cooked, because US companies will RLwash Chinese models to try and get in on the captive market. The frontier labs have already lost the war for coding, their next play is custom models for specific domains... Anthropic Galen for biomedical research, Anthropic Locke for legal analysis, etc, and you won't see _ANY_ intermediate work on the model, you will put in query, maybe get some questions fired back during work, and get a "final report."
Eventually the frontier labs will try to cut out the middle man once these models prove themselves and start doing partnerships with big firms in the domains, so they can take a % of the profits in perpetuity rather than just taking a one time payment. For example, after Anthropic Galen, they'll do a partnership with Pfizer to generate Ozempic-Superjacked and take 20% royalties on global sales.
>The frontier labs have already lost the war for coding
This is a delusional take. Sorry, but anyone claiming this hasn't used Fable and compared it to the current best open source models. I see a lot of hype posting about GLM5.2. I see absolutely ZERO people using it in production compared to GPT 5.5 or Opus 4.8.
> The US govt is going to ban foreign models
The people have a right to make and use whatever models they want, protected by the constitution. At a minimum, the models are described in research papers that are unquestionably protected speech. Skilled devs turn those into programs, also protected speech.
How could Trump ban tiktok then? And Fable for that matter.
Maybe you're somehow legally allowed to distribute and download the weights, but most of us can't run GLM 5.2 at home.
You won't need a frontier size model for most tasks before long. Qwen 3.6 (small) punches way above its weight. I run it at home @8bit on an OEM Spark
Second this, I am also running qwen 3.6 35b Q8 on a 5090 liquid getting around 250 tokens / second and it is plenty capable. I actually haven't even looked at models recently because I am happy with what I have.
And.. now I feel the need to look again. Darn, there goes my afternoon
And corporations could run DeepSeek models on cloud hardware.
You can run most open models on cloud hardware. Google Cloud gives you a click to deploy, but then you have saturation / ROI considerations, versus Google serving them up multi-tenant, per-token.
The current administration has repeatedly demonstrated they do not feel constrained by laws or the Constitution.
The US government isn’t supposed to be allowed to constrain speech, but they do have the power to constrain commerce, and they can ban the sale of AI services and AI-capable hardware if they choose.
Yes our passionate defense of Academia will surely survive Techno Oligarchs desire for a 20th vacation home
> The frontier labs have already lost the war for coding
You are way too deep in the HN bubble.
I'm looking at how market/human forces are going to make the game play out when extended to its logical conclusion, not the score on the scoreboard RIGHT now.
I'd guess Anthropic will probably win, and LLMs will probably still be with us and be much better in 10 years time.
But next year we could be in the middle of a massive $600B/yr capital-spending bubble deflating hard with unemployment accelerating towards 10% (or higher).
The internet never failed, but the telcom/dotcom collapse still happened in 2001.
We're seeing the first 20 years of the dot-com cycle, but compressed into two years, and trying hard not to fall into the tar pit of ad-supported services.
So long as Chinese labs keep writing white papers, trade secrets aren't going to win the day.
Having growth up in the 90s, it is weird seeing companies share their technology secrets publicly.
Wandering around pretending to be researchers who are only just figuring out how to make money is, for the short term, an incredibly good way to attract a load of naïve money; not all sharks are smart.
And it does, nowadays, give you a bit of a veneer of mere curiosity when you're being accused of massive theft.
Deepseek is 90% cheaper, and nearly as good for coding tasks as claude/codex, and as good given the right plan.
The only moat OpenAI and Anthropic have is regulation. If the Chinese really eant to hammer us, they could realse the full training data and pipeline.
Even without doing that the Chinese are already going to impact our labs presence everywhere else in the world. With Fable getting pulled, any model coming out of the US is now unreliable and untrusted. No one in any other country would in their right mind choose OpenAI or Anthropic for anything.
The big push for regulation and export controls is only going to ensure OpenAI & Anthropic are more like the automakers. Only in business because of protectionism, left to screw over US consumers meanwhile the rest of the world gets to enjoy cheap EVs
The issue is the cost is not going to be a hindrance for companies that have gone all in on the AI development. They may still find it cheaper than hiring engineers and if needed they will layoff a few more.
The companies that did not yet jump on this bandwagon and are still evaluating will have a decision to make.
No matter what the AI companies are going to change their pricing strategy and it’s going to become a lot lot more expensive to use. I am just hoping the price stays like this until I am done with my big chunk of work
This article seems to be struggling with telling apart the difference between R&D and operating expenses? The fact that AI companies are extremely unprofitable doesn’t mean they are subsidizing token costs, they still can have very decent gross margins on them
This article gave me an amusing thought: the only jobs with a high enough salary to be profitably replaced by AI might be software engineers.
I think a lot of the cost comparisons to employees are off by a factor of 2 or more. AI is the ultimate contractor. Available instantly. Doesn't charge during idle periods. Pre-vetted and pre-trained. No contract negotiations or complex accounting.
That is worth a small multiple of the fully-loaded employee cost. So AI might be easily worth more than $200 per human-equivalent hour. With high utilization, that might be $8000-10000 a month.
With that kind of spend, AI provider financials looks less frightening.
I can't wrap my head around how revenue > COGS but at the same time AI is being subsidized and the real cost is not affordable.
You don't price based on cost, you price based on willingness-to-pay.
So maybe labs are "overcharging" enterprises on interference (because, up til now, enterprises have seemingly had unlimited budget for tokens) and "undercharging" individuals and SMBs (because they don't have an unlimited budget).
I can't go back to a life without AI, and I don't want to. But if AI were billed by token instead of subscription, my monthly cost would probably be ten times what it is now. I could switch to a Chinese model, but I'm not sure how things will look by then.
What makes AI so convenient is how good it is at doing red-team code reviews on my work. I used to need all this unnecessary communication just to get a review, but now I only have to reach out to the people I actually want to talk to.
> Anthropic is subsidizing their enterprise customers by up to 40 times, and OpenAI up to 70 times
might as well be the other way around with non subscribed token being 50x overpriced, or any combination thereof
also uber was non profitable for the longest time, raking up 31b in losses, on the bet of capturing the market worldwide. scale here is different, but it's also 10 years later, with a lot more volatility and floating cash in the market (voo grew 327% over that period, not unreasonable that round size grew on the same trajectory)
The coming AI enshittification is going to be epic. For those of us who have been on the web for more than five minutes, we can see this a mile away.
If you think search ads are annoying, pre-roll YouTube ads are annoying, streaming ads are annoying, or basically ads-on-any-screen-anywhere-at-any-time are annoying, just wait until every stupid thing is powered by AI and is subtly trying to manipulate you to buy/watch/believe some crap all the time.
Jeopardizing a $200/month subscription in return for $1/month in ad revenue seems insane. Using ads on a $20/month subscription to entice you into a $200/month one, OTOH...
They will almost certainly get more than $1/month in ad revenue for someone interacting with the AI for hours a day.
Is it not also possible that some of the shift is a consequence of increase of use? While we can be extremely cynical at the finances at play, the lock down and increase of token pricing might be demonstrating a burgeoning demand, which would be a positive indicator.
Yes. If we spend more on building AI infrastructure then current total global gross software sales, the only way the math works is if we create and sell much more software or if we start charging more for it.
The math doesn’t add up and the wheels are starting to come off the bus.
The conversation in a lot of wealth management offices has shifted dramatically in the last few month from “how do I get in on this AI thing?” to “how do I protect my assets when this AI stuff blows up.”
There’s little question now if this will all implode, just when and who’s going to lose their shirt and be left without chairs when the music stops.
What’s playing out now is the scene from The Big Short where the banks wouldn’t mark down the value of bonds until they secured a short position. Once the big money has their helmets on it will stop providing fuel for the bubble and then look out below!
With these confident comments I would appreciate some kind of origin of the information. Not even necessarily a source accessible to me, just: are you in any wealth management offices? Or are you reporting other people's opinions? Or does it just sound right given the spirit of our time?
Assuming the analysis is right, and most (or all) of these AI companies will default on their debts, what consequences might that have?
If that happens the AI companies will first try to negotiate with their creditors and after that likely declare bankruptcy with the creditors taking over what’s left of the assets. Shareholders will be wiped out and employees will be left with nothing. Various franken-companies will emerge from the bankruptcy ashes and the world will move on with AI sans the present irrational exuberance.
> If that happens the AI companies will first try to negotiate with their creditors and after that likely declare bankruptcy with the creditors taking over what’s left of the assets.
Due to the fact that we’ve already done this before (Enron, Global Crossing) -
I’m willing to bet that there are contracts in place ALREADY, that define what happens in the event of a default.
In particular, I’ll bet that the buildings, the GPUs, the patents, etc…
All of these have probably been accounted for.
I worked at a data center that closed during the WorldCom era, and when they put the padlocks on the door, there were still websites “hosted” from the building.
I don’t know if they killed the power or what. I’d cleared out my desk long before they locked it all up. I wouldn’t be surprised to learn that these websites couldn’t get their own servers, since ownership was tied up in the courts.
In the Bay Area during that time, there were row upon row of empty office buildings.
And then the US government will say that these company's futures are in the national interest, and they will be bailed out with taxpayer dollars.
Usually (but not always) such bailouts wipe out the shareholders.
That could happen, but the shareholders still get wiped out.
All of those poor agents will be laid off from their support chat jobs and their roles will get outsourced to India and Philippines.
>what consequences might that have?
All depends on who is holding the bag, and how big the bag is.
Hence the IPO. Push the risk on to retail and index funds, away from private credit. Plus Microsoft, Google, and Amazon will also be holding the bag and have huge balance sheet write downs, the compute commitments have not yet been paid for it's all just promises.
The banks aren't has exposed this time, as in 2008, most of it is tied up in private credit, its more akin to the fiber buildout in the 90s.
Yes, and private credit investors are rushing for the exits but can’t get out because of withdraw limits. It’s starting to get ugly. Folks owning something they think is going to tank and they can’t sell.
Yep, if they make it to IPO, as SpaceX has, and if they manage to get into several indexes (as SpaceX is already doing, I assume it's already in the Russell indexes, and will soon be in the Nasdaq 100 index), it'll be a bunch of working class people's retirement accounts holding the bag. And, those same companies might be deemed Too Big To Fail, and they'll get even more working class folks money in the form of tax-funded bailouts.
A wealth transfer from the working class to a handful of billionaires bigger than any the world has ever seen (and the world has seen a lot of wealth transfer from the working class to billionaires).
Probably 401 (k) plans for the most part.
> OpenAI Had $13.07 Billion In Revenue, $34 Billion In Costs and Expenses, and $20.92 Billion In Losses, with a net loss attributable to the company of $38.53 Billion
This is going to be the new most misquoted/misunderstood data of the year, isn't it? The cost is mostly from a one-time accounting situation due to their pivot from a non-profit organization.[0] If we trust the leak [1] OpenAI is likely turning profitable this year.
[0]: $30Bn of it is the one-time cost. https://www.ft.com/content/e15b0d7e-ff6b-4f16-ba7a-4068feddb...
[1]: I suspect OpenAI itself leaked that financial report. It's almost unbelievably healthy.
These companies biggest source of revenue is per-token pricing though, not subscriptions. On tokens they make a good margin.
I don't see any real point being made in (or point of) the article. The author sort of just...dumped a bunch of links with the noise that is so incredibly mainstream at the moment that I doubt any of it is news to anyone even somewhat tracking the AI cycle. Most of it (except for maybe the BLS[1] stat) is just regurgitation.
[1]: And this too is incorrect, should be " the number of jobs displaced would be around 32.5M" (the post says 32.5K)
Affordability is not the current goal.
Vendor lock-in is the current goal. Consumer prices are a drop in the bucket comparatively.
How can you lock in when the harnesses are basically thin clients around the APIs and you can replicate them using agents in a short period of time? I haven't seen a compelling thesis yet for how you achieve vendor lock in for LLMs. Claude Code is a bit sticky, but if we're being honest its just because Codex doesn't have all the same features yet.
Luckily the industry is much too wise, after a couple of decades of cloud infrastructure, to willingly opt to make itself entirely dependent on one of two platforms with opaque and complicated pricing. We've learned our lessons, oh yes
Maybe they just need the competition to run out of funding first?
That’s an impossible goal; it’s too easy to switch models.
And Microsoft forced M365 subscriptions to include AI for +$30/license.
Cheap, but gave them a massive user base they can claim is using AI
This summarizes half of the entire AI scene as these guys generate content to paint the entire world the way like to: US equity markets are facing three IPOs .. each led by a world-class bullshitter”.
The willingness to throw capital at AI is definitely doing some crazy things, but this article has some bad takes on the data.
> [Ratio of per-token cost to subscription cost] means Anthropic is subsidizing their enterprise customers by up to 40 times, and OpenAI up to 70 times
Actually, they could be subsidizing by more (if they are taking a loss on API), or not at all (if they are soaking API customers by a massive margin).
Separately, these subscriptions get sold to large groups with varying usage, so it's crazy to model assuming every subscription is maxed out. Banks, gyms, and many other businesses work this way, offering consumers flexible access to services that they will realistically use in bursts. It's not always worth the complexity to prevent overuse by a small minority. You can feel like this kind of business model isn't as transparent, but it's silly to pretend it can't work.
> OpenAI spent 44% of their revenue [$5.3B] on sales and marketing! The hype needed to keep the AI bubble inflated is incredibly expensive.
Over that same period (2025), OpenAI added $10B in realized revenue and $14B in run-rate. Sounds like they're getting >2X return within 12 months of those go-to-market dollars. Compare that to like, any other business.
> Thus in recent weeks the idea that Generative AI (LLMs for short) is too expensive has been all over mainstream business media.
Would it be smarter for these companies never to test customers' price tolerance? The quotes following this make it seem like the companies are getting important information about the nature of that price tolerance, and preparing to react. This is the work markets do on both sides to understand the value of a new product.
There are lots of good arguments about AI overinflation, but in order for them to be useful, they have to be rigorous and targeted.
I know a lot of level-headed engineers here may not side with me, but I say let the companies who abandoned their people at the drop of a hat, with CEOs who waved their flag around on social media, proudly declaring how they'd now run their companies with 75% fewer employees wither and die. If I had been let go, there's no way I'd go back to a company like that, and there should be a black list of CEOs who acted this way established and kept public. These CEOs are not holistic thinkers, and are too susceptible to mass hysteria and too irresponsible to real people and their lives to be trusted with the vision for any company ever again.
Someone should keep track of a public database of CEOs who cut workforce while making huge profits. Name, context, situation and all.
Unfortunately, maintaining an opposite list would probably be easier.
That's basically what you'll see if you open a newspaper to the stock page. That's the idea behind business. It's why you have what you have.
Depends on if it's shortsighted by giving up your intellectual capital for short term profit.
GM just did this in the last 30 days [1], and their sales are likely going to be just fine. In fact the auto industry has repeatedly automated jobs over the last 100 years, and they still make decent sales numbers.
If you decided to boycott every company that replaced staff with automation, you would be forced to exit the economy. Every company does this to some degree and the customers who vote with their wallet do not seem to care about a reduction in force.
[1]: https://arstechnica.com/ai/2026/06/gm-installs-robots-at-fla...
Robots that replace auto industry factory workers exist; the CEO of GM didn't imagine them as part of some sort of business media induced psychotic episode.
The same is not true for the software industry execs.
GM is running 0% interest, no payments until n deals right now.
That’s usually a sign that sales are not “just fine”.
They always are.
The above comment, to which you responded, wrote about CEOs who responded to mass hysteria, not those who automated anything.
> GM just did this in the last 30 days [1], and their sales are likely going to be just fine. In fact the auto industry has repeatedly automated jobs over the last 100 years, and they still make decent sales numbers.
I worked at Verizon during their layoffs last year. Biggest layoffs in the USA.
As someone who’s been laid off before, I knew that it generally boosts the stock price.
I bought VZ because of that. It’s up 15% since the layoffs.
Microsoft, an AI stock, is down 30% in the same timeframe.
This is true, and I'm sure AI cuts will continue, but it's obvious that the ones who went "all in" at AI's mass introduction were drinking a special kind of Kool-Aid reserved for the truly sycophantic Wall Street lap dogs, not the CEOs who think about risk and are cautious about betting the farm on a relatively new and mostly untested technology. GM is over 100 years old, and no doubt released improvements that were well-tested and predictable, because you don't take massive chances with a company that well established. It was a couple years into the mass AI deployment that studies on the minimal overall productivity gains of AI even started to come out(!) This was "get on the bandwagon" thinking at a massive scale, which shows you how many CEOs are not independent thinkers at all, but are really just followers. Yes, use AI, but do it responsibly, never forgetting that your investors aren't your only stakeholders - so are your people.
I'll believe it when I see it, but I would love to see it.
Spelling mistake:
"a return on these invetment"
It's Proof of (human) Work. Much more useful than having a sticker saying "Done by a Human".
Is deleting a letter after an LLM generated the article an insurmountable task? These quaint signals only screen out the lowest of effort slop writers. Better than absolutely nothing, but barely.
It does remind me of the time a chef told me when he puts lemon juice over a dish, he would intentionally not remove any seeds that went on it because it was a signal of quality. I wonder if future slop chefs will intentionally place seeds on dishes that came from a box...
Ask your LLM to 'write like a phishing email' to have it seem more human.
I'm actually curious if this works, haven't tried but I assume it would.
...and "maiinstream" -- seeing glaring typos (easily caught by spellcheck) now makes me wonder: did they decide to leave them in (or add them explicitly) to signal they didn't use AI to write, or (the more paranoid option) did they tell the LLM to add a few typos...
I didn't get the sense this was LLM-written, but typo-signalling is... I donno a bit weird. Firefox is underlining some of the words as I write. I'm leaving "donno" unchanged even though it's flagging it as a misspelling but I suppose I'd still opt to fix something like "maiinstream" even at the risk of potentially seeming more LLM-ish!
It's funny when you watch the doomscroll all these anthropic guys talking about how you should be writing self-improving loops and that's all they do. Of course that's all they do, they don't have to pay for their tokens.
Can confirm, my experience in “loop engineering” was “this is neat” for 45 minutes until a daily ration of tokens was evaporated. The quadratic cost trap is prohibitive to experimentation.
As a localLLM evangelist, I am hopeful this will bring more attention to the joys of rolling your own sovereign AI.
Yeah, i'm hoping that gets smoother. I've been experimenting with omlx and opencode on my m5x64gb and keep running into issues w/ Qwen3.6-35B-A3B-MLX-8bit exceeding it's memory limit at the most inopportune times. Playing with 12B gemma4 (8bit) more today.
Maybe I should be aiming for something targeting 48gb of memory?
I really can’t stand when writers point to the difference in price per token on the api and subscription and use that as evidence that inference loses money. This author even says it’s implausible that the api charges 4x marginal cost when I think it’s very likely even higher than that. The entire rest of the post sits on this faulty assumption. Fixed costs don’t matter when marginal revenue is profitable and growing rapidly. The ai labs only have 2 questions. Can they prevent users from switching to open source models? Can they scale the number of users on enterprise plans the way they did for coding but in a more general way for all knowledge jobs?
Then what are the real costs?
Wrote this a while back. https://martinalderson.com/posts/no-it-doesnt-cost-anthropic...
OpenRouter is the best guide to real costs.
agreed, this doesnt even account for prompt caching or the fact that anthropic has substantial proprietary efficiencies on their inference stack specific to their models and scale.
Thanks, that’s exactly what I was looking for!
And much more informative than the speculation and guessing in the article.
> Can they scale the number of users on enterprise plans the way they did for coding but in a more general way for all knowledge jobs?
Do these knowledge jobs have a significant corpus of not only knowledge but discussion and problem solving, all conveniently labelled for the AI to train on? Probably not. Coding has stack overflow, what does, say, advertising use?
I agree this is a hard problem for the labs. I would be hesitant about “probably not” though. There is just as much marketing copy floating around as there is coding training data. I struggle a bit in this question because I’ve only ever worked as a software engineer, so I can’t exactly make claims about all the work other jobs do. But, one example is I was talking to a doctor friend of mine the other day. He was talking about how he had to take his recertification exam recently and put the questions into chatGPT and thought it gave answers that were generally more thoughtful and correct than his own. Does that mean doctors are done? Of course not, but he’s now pushing hard for more ai tool use in his practice.
> Coding has stack overflow, what does, say, advertising use?
Advertising has centuries of print ads, 100 years of radio advertising, 70 years of TV commercials, etc. And modern AI does not necessarily need labeling.
The article fails to mention DeepSeek, Alibaba, Qwen, Xiaomi, MiMo, z.ai, or GLM. It's hard to take such an article seriously that doesn't do this. (Our monthly total spend is around $180 with a team of 6, about half technical; our biggest line items are for American models or subscriptions which we probably will be planning to get rid of.)
And then remarks like this:
Huh? I use OpenAI via a subscription, as is anyone else using GPT-5.5-Pro who isn't a multimillionaire.They're referring to Enterprise customers, though should have been clear about it. Enterprise plans on Claude for example no longer include any baseline tokens. It's 100% usage based pricing.
True, but my friends in Enterprise still just purchase Claude Code subs and expense them. They basically get an allowance of $500 or so per month to buy various tools, and of course are banned from Chinese models. (Claude, Codex, Antigravity allowed, basically.)
> Our monthly total spend is around $180 with a team of 6, about half technical; our biggest line items are for American models or subscriptions which we probably will be planning to get rid of.)
Please tell more :). Do you pay per token from bedrock / openrouter / somewhere else? How many tokens you use over the month, and how many for each task? Which harnesses?
Pay for DeepSeek directly. One developer insists on having his own account and in theory expenses it, but he forgets to turn in $10 expense reports. (Total spend in last two months = about $45.)
Pay for OpenAI Pro directly, but I’m the only guy that uses Codex. $100 a month. My nontechnical partner likes to talk to ChatGPT 5.5 Pro for image related tasks (think generating interior decorating pics).
The nontechnical staff use a Gemini account on a Google family AI Pro sub. I use Antigravity when working on Android or Google Cloud API codebases.
Everyone gets OpenCode Go. The cost is trivial. $10 a month per person.
Pay for MiMo directly. We use it during Chinese off peak hours though. Total spend so far $25 in last month.
We run a few Qwen models locally and pretty much have them pegged all day. RTX 5090 on a PC and a Mac Studio.
There’s also Grok which is used for Imagine for artistic / graphic design related work. I also use the subscription for a vision model in my oh-my-pi harness.
We’re having discussions about how to pull in GLM-5.2 cost effectively. We compete with third world development shops so we can’t really pass on inference costs, but we can benefit from getting jobs done for customers faster. But ⅔ of our work is either internal or open source projects we can’t bill for.
Not the GP, but I use Opus for planning, Deepseek for actual coding (implementing the plan) and GPT for review. GPT is inexhaustible on the $20/mo plan, Deepseek is dirt cheap (maybe $10/mo) and Claude is Claude.
GP is talking about API / token-based prices, that's why I asked.
I don't know, he said "subscriptions" in the line items, but eg I use Deepseek via the API.
Ah maybe you're right.
I can manage this budget with the chinese models in AWS BedRock. However, in my experience, they aren't as good as claude today.
I think the author is referring to enterprise customers. You aren't the "customer" in this case; you're the bait.
How do you know that the other models you are referring to aren't subsidized?
Subsidizing makes no sense when there's no - possibility of a - moat. Although it's very possible that China in general subsidizes Chinese labs in some way so they maintain pressure on US labs. But you only have to look at proxies such as OpenRouter to see that the individuals aren't doing any subsidizing on per token costs.
"Crisis"
The Token Tension :)
Most of the "affordability" and "pricing" discussion is pointless because we don't have any real numbers on their margins per token. So, yes, they are subsidizing their subscription plans compared to the API prices, but the API prices could already be stupidly inflated, so the relative price comparison is a nothing burger. Until we know (or at least get a hint) on their margins on API prices, any pricing discussion is pointless.
I don't understand this line of reasoning at all.
We have a pretty good idea of how much it costs to serve these models. You can pencil out the economics and guess at the model sizes and we know pretty decently how expensive the hardware is.
This like claiming it's meaningless to guess the margins of a restaurant without going into their books and seeing the exact recipets and recipes.
They ain't doing dark arts in the back. You can guess at what goes into the food based on similar recipies and how much that costs based on what you pay at the grocery store.
This is basically bunk because AI costs have gone down by 50x or more (api costs) since 3 years.
This doesn’t solve the problem because (tautologically) the more AI prices go down the less money the companies make. If right now today the companies are operating at a profit and a price war causes the API costs to sink 90% next year, and their capex amortization costs stay fixed.
The math doesn’t math.
AI prices going down means the models are improving, particularly from the efficiency angle (which is inevitable, given the nature of tech). That means all they have to do is maintain a large enough customer base at a rate high enough to ensure loss decreases continuously over time, until eventually the pass the point where they're just gaining. Healthy competition ensures that improvement savings are actually passed on to users in a measured manner, so they don't become too greedy in trying to get to and increase gains.
But now you’re describing a commodity, and the competition will erode profits, and their valuations are bananas, unless someone can find a business model that truly differentiate and creates a moat.
Models are not commodities and are famously non fungible. Each model has its quirks and strengths, weaknesses and idiosyncrasies.
I know because I see how people went over the 4o model. I can see opus behaving clearly differently enough that I pick it for certain tasks.
Is this really for comparable models though? Will folks at scale continue to choose Anthropic frontier AI model if OpenAI releases a similar generation at a 90% discount with comparable capabilities? It feels like the fungibility assumes delineation by capability _and_ cost. No one is choosing sonnet over opus at similar price points.
This doesn't really tell you anything useful. AI companies have both built huge datacenters and raised a colossal amount of money. Include caching, quantization and etc. All of those would allow them to undercut on price considerably, even more so if you count in all the users who don't actually cap out their plans. Prices going down doesn't really tell you anything about the production cost, especially in a market where every major participant is happy to burn money just for the marketshare.
There are many research avenues which are open which reduces cost dramatically. Smaller task specific/ language specific/ domain specific models, in fact they could even be better. The earlier computers were the size of a building. So prediction based on current state into the unknown future possibilites is wrong. The hardware will be all the more valuable if cheaper ways to run become possible. The hardware gets cornered in a sense.
Because of it's unpredictability and massive dependence on the training data, when LLMs start hallucinating most of the time the only fix these "engineers" have is to feed it another LLM... The genius was the transformer architecture, and evidently none of us have a damn clue how it works
Every 6-12 months or so we get an increase in one or more of things like: compute power, compute efficiency, GPU power, GPU efficiency, network bandwidth increase, memory speed increase, component density increase in the same form factor, etc.
For awhile it was every 2-3 years you'd start a hardware refresh. As companies moved into more and more training, this timeframe started to shrink. It went from 36 months to 24 months. From 24 months to around 16-18 months. Last I checked last year, it was at 12 months. I think things may have slowed because of component availability, but otherwise whole data centers would be 6-12 months into full operations before they would start a refresh cycle.
Not to mention the massive increase in power density demand and cooling demand per rack that entails.
So no, "AI costs" have not gone down, in fact they are more expensive on training AND inference than ever.
This is why many are concerned about the heroin drip of api costs into orgs. For the companies that are public, look into their financials. It's gonna hit companies and high volume users like a ton of bricks.
I'm no economist but if true don't you have the opposite problem? How do you get people to need X many tokens per day such that you can sell enough to make money? Wouldn't you need an absence of competition for that to be ok?
the demand for intelligence is infinite. you sound like someone in 1960 wondering what the hell we would even do with the functionally infinite cpu cycles we have available to us now.
If you are an AI bear you have multiple techniques with you
- if AI costs go down you can ask how the companies will make profit and then suggest the bubble popping
- if AI costs go up you can ask how people will afford it and then suggest the bubble popping
- if companies actually do make profit then you can say the companies are getting too big and powerful so it’s a bad thing for consumers
Essentially you have left zero to a small narrow path where you are happy with the outcomes.
Can you cite a source? Everything I've read describes the costing as linear with growth.
The quality of what you can get from DeepSeek V4 Pro for $10 is light years ahead of what you could get for $20 a year ago.
Likewise, the quality of what I can get from a local model like Qwen 3.6 on an RTX 5090 is light years ahead of what I could get a year ago on the same hardware.
https://simianwords.bearblog.dev/conclusive-proofs-that-llm-...
That article seems a bit bogus. Cost per capability is a soft, non-predictive model unlike cost per token which has been trending up.
This is just hand waving on the obvious consensus that cost per capability is going down. There’s no doubt about it. Hell you can run a Gemma 4 model on your laptop that mogs GPT 4. But yeah you can use fuzziness as an excuse and ignore the trend.
What?
He's saying output for 1M tokens on the latest models is $50 now when it used to be $2500.
so how are these labs going to recoup the insane training costs at those prices? even if there is still a fat margin leftover afterwards
They also have to continuously train, forever, to avoid model drift. It's not a one and done thing as far as I'm aware.