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.
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.
> 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.
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.
> 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...
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. Implementing bad ideas 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. In that scenario, there's just no room for AI usage. If you have to be responsible for all the code, then... AI has to either manage it completely autonomously (which even Fable can't) or... humans have to be in the loop which means they still have to understand the code. The best way to understand the code is to write the code yourself. So there's no productivity gain to be had.
I'm pro-AI, but I think we're due for a big crash next year.
I feel like the author is jumping way too fast from "OpenAI is losing money" to "the whole AI economy is broken." A company being in the red during aggressive scaling doesn't automatically mean the unit economics don't work.
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.
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.
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.
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.
> To generate the $309 billion needed to service their debt, the AI industry will need to replace 46.8 million jobs, equivalent to around 27% of the current number of jobs in the US.
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.
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! ;)
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.
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.
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.
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.
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.
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.
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.
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
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
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
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?
> 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
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.
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”.
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.
...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.
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?
> 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?
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.
> 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?
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.
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?
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.
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 ..
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.
> 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.
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.
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. Implementing bad ideas 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. In that scenario, there's just no room for AI usage. If you have to be responsible for all the code, then... AI has to either manage it completely autonomously (which even Fable can't) or... humans have to be in the loop which means they still have to understand the code. The best way to understand the code is to write the code yourself. So there's no productivity gain to be had.
I'm pro-AI, but I think we're due for a big crash next year.
I feel like the author is jumping way too fast from "OpenAI is losing money" to "the whole AI economy is broken." A company being in the red during aggressive scaling doesn't automatically mean the unit economics don't work.
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
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.
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.
How are Anthropic and OpenAI going to compete on price when they're both already deeply unprofitable?
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.
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.
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.
> To generate the $309 billion needed to service their debt, the AI industry will need to replace 46.8 million jobs, equivalent to around 27% of the current number of jobs in the US.
Lump of labour fallacy spotted.
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.
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! ;)
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
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…
> 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?
Likely free tokens to attract customers
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...
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.
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.
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.
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.
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.
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.
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
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
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
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)
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.
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?
Assuming the analysis is right, and most (or all) of these AI companies will default on their debts, what consequences might that have?
> 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
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.
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”.
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.
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...
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".
...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.
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?
> 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?
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.
> 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?
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?
"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.
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?
Can you cite a source? Everything I've read describes the costing as linear with growth.
What?