They've got, ballpark, $5t to $10t to make back in the next 5 years, or the hardware buildouts will start getting written down.
This means we're going to need $1t+ per year in spending, per year, on tokens. 200m knowledge workers in the world, 30m developers. We're talking about a world where you need 5% of every knowledge workers salary to go into tokens. 20% if you're a developer.
That's a _huge_ shift. Most people I know cite +20%-40% velocity with these tools, against the actual work their company cares about doing. +20% speed for +20% spend isn't going to motivate a trillion dollars a year in spending.
We're not there yet. This is still the upswing of the hype cycle, and unless we figure out how to make developers 2x, 5x, 10x as productive on stuff that matters, this isn't going to play out well.
I asked claude to generate a frontend and it made the same template. Same san serif and serif fonts together. Same colors. Same typography. Same layout and animations even. It’s wild how similar it is. No not similar it’s the same damn thing.
> Gartner forecasts that large AI companies would need to earn cumulatively close to $7 trillion in AI-driven revenue through 2029, which is close to $2 trillion per year by the end of the period. In order to achieve “historic returns,” the providers would need to earn nearly $8.2 trillion in the same period.
You are making the assumption that the models are only used / paid for by 2.5% of the population (your knowledge workers value). There will be new value created by these models which people are happy to pay for which simply did not exist at all before. It is also naive to say that the hyperscalers are going to be expecting a return on this in 5 years, it will be entirely propped up by investments / IPOs as has been the case with any tech company for decades now to reach scale. The hyperscalers are currently spending ~650m combined annually, which they have the cash for and can sell in future compute instantly.
Source on 200 million knowledge workers worldwide? My understanding is that it's just above 1 billion. I dont think a billion subscriptions at $1000/yr is out of the question but it might take a decade to get roiling
Yeah, it’s nothing, and it’s also not the cost that enterprises are paying. As the article states, the price is $20 per seat per month, PLUS per-token API usage. Enterprises are paying consumption billing, not fixed rate oversubscribed “all you can eat per seat.”
“Tokens” don’t have an intrisic cost or value. Saying that I used $2,180.16 worth of tokens is like relying on the salesperson to convince me I’m getting a billion dollars worth of pots and pans for $19.99.
I think it’s funny how we are throwing critical thinking out the window when it comes to evaluating biased sources of info.
I'm not sure what you're pushing back against here.
I spent $200. If I had been paying API pricing it would have been $2,180.16. The article is about how enterprise customers get charged API pricing, which means if I had been employed by one of those companies I would have cost them $2,180.16.
And "large" just means that AWS will assign an account manager to talk with you. I was at a start-up who spent $300k/year on AWS and that was enough to get special attention and discounts. Enterprise pricing is confusing.
> If I had been paying API pricing it would have been $2,180.16
The point being made above is that API pricing is calculated... somehow... seemingly arbitrarily. Possibly untethered to the infrastructure costs entirely: which would be the basis of any 'value', however that holds the labor theory of value, which isn't accurate either. So how do you accurately price these tokens at all (other than through price-discovery: which is slow, messy and fuzzy)?
API pricing drops DRAMATICALLY in enterprise agreements.
As with pretty much anything priced on volume/usage.
Enterprise deals are negotiated ad-hoc, the listed pricing is simply a jumping off point for the final negotiated discount.
If you’re going to give 20,000 employees Claude code you are not going to be spending $1B per year on Anthropic tokens as if you gave everyone an individual API key.
Tokens do have a clearly calculable intrinsic cost. There's the marginal cost of production (i.e. the inference cost) and the amortized R&D cost that goes into the model producing them.
Yes, value is hard to calculate, but luckily market pricing mechanisms exist exactly for this purpose. There isn't a better number to use than what people are willing to pay for them.
So he's saying that on an enterprise plan, he'd be spending $2,180.16. He's not paying that much, but enterprises are.
Does this analysis factor in potential caching of tokens on the server side? It seems that if they organize things well (as a model provider), they can save quite a lot on that. Looking at my Cursor statistics makes it clear that the token calculations are not at all trivial.
The costs are exorbitant and most software is not produced by companies with such a huge moat. Anthropic made a profit through their recent bait amd switch pricing. There is zero useful insights online to indicate whether this might die due to commoditisation with good enough open models or fail the race to get more people subsidising unsustainable growth with other people’s money. Who knows? In any case they dont seem to be able to drop usage costs so the business model seems based on wishes
With deepseek and xiaomi mimo models slashing their prices 99%, I don't see a great future for openai / antrhopic with regards to their 1T valuations. Maybe 1T valuation will be the whole market, West + East.
They'll still have their dedicated enterprise customers. I think the Chinese providers will pull more of the single users who're paying their own way, than those backed by company budget. And it's a pretty good split as the demand becomes better distributed, resulting in better service (I'll never forgot must how bad access to Claude became until they got access to Colossus) and less potential for lock-in (we really don't want there to be a duopoly, etc on good AI).
So how do openai and anthropic plan to keep customers when GLM-5.1 is just as good and open source and a lot cheaper?
I don't see the business model working. My closest friend actually does automation software for large companies.
He does not use Claude or openai at all. He primarily uses gpt 120b on cerebras and glm-5.1 for heavy thinking work.
And some other small models for various tasks. All open source.
And these systems are extremely useful for the businesses and are able to run fully automated pipelines that are very stable and fast.
We discuss this a lot, and we both think any business doing heavy agentic work on Claude and openai just aren't aware of exactly how good and cheap open source has gotten on the last year.
So... once the legacy businesses and developers catch up, won't Claude and openai be unable to recoup their costs?
For coding you always want to go with the best model in the category, not something that would be the best model if we went 1 year back which GLM 5.1 is, and I'm saying that as a big fan of GLM cause I run a translation site where GLM is good enough for the price.
Most of the money right now is in coding. Openai and Anthropic just have to be 6 months ahead of SOTA open source models and they'll capture most of the enterprise and dev market
Yes I'm an engineer (20 years most in games/graphics industry) and only use it for code. I've been using glm 5.1 this week a lot. I went in expecting another "decent" but not really "up to standard" open source model.
I highly doubt I'll ever use Claude again.
I think you are wrong about Claude being any significant level better
For coding assistance, I have tried OpenCode with several large open models through OpenRouter. All were fairly bad compared to Claude Opus.
Could you provide some hints on how I should be holding these open models so that I might get more value out of them?
I agree with the common trope that open models lag behind by about a year, but something magical happened just around a year ago when the state of the art models became extremely useful. By this reasoning we're about to see open models perform well, but I'm afraid there is more to it than just waiting for another revolution around the sun.
Note, my application is coding assistance. Open models can be great for other purposes.
Great article I know this upsets a lot of people who are used to thinking Anthropic/OpenAI are just lighting cash on fire but they've cornered the market on enterprise who cannot walk away from these $200/month plans
However the valuations are still far far away from actual sanity
It's good enough for personal stuff. It doesn't compare to the latest Opus I use at work. You can certainly argue I don't need Opus for work, but there is clearly a difference.
Also, at least with z.ai, GLM-5.1 is s l o w! After using Claude at work, I get really impatient with GLM-5.1 at home. When doing "true" vibe coding (i.e. not really examining the code), Opus is a ton faster (easily 5x).
But yeah, I'm not willing to personally pay for the frontier models. I won't even renew my annual Z.ai plan - it's become too expensive.
task i am working on right now at work is comparing two verisions of apis and documenting responses in their outputs. i suspect a vast majority of work at entrprise is of similar complexity.
right now everyone is using latest and greatest to do dumb stuff like that. that would change fast if companies start caring about costs.
How is the lack of bad news declaring a victory for AI? I am yet to see any company concretely publish analysis about the ROI from AI. Most companies as far as I know are still treating AI investment as sunk cost with no expectation of returns at the moment. We could very well see a world where companies heavily scale back investment.
If nothing else this blog did give me the idea that I should split my $200 claude max plan into two $100 CC max and $100 codex plan, esp because Claude is now offering 1.5x weekly limits so its the 5x usage is now more like 7.5x usage.
I think the reasons for them going with API pricing will become abundantly clear when the S-1s become available. If they don't have a story covering how they can get revenue closer to expenses, then they're relying on the market to believe the pixie dust version of their profitability story, which I think people increasingly don't.
So it largely sounds like many more people will be able to write software - and will use AI to do it. Existing software engineers will continue to automate their tasks away like they always did, but perhaps at a faster rate.
The impact of AI in other fields seems to be muted.
I think it is applicable to a much wider range of knowledge work, but it's also harder to apply there.
Software development has the huge advantage that mistakes and hallucinations are very easy to spot: the software works or it doesn't.
Spotting errors in a research report or legal brief is a whole lot harder!
But... non-software professionals spend a huge amount of their time on tasks that can be safely automated - reformatting documents, extracting numbers from PDFs, all kinds of flavor of data entry.
Learning how to use a tool like Claude Cowork can take a big dent out of those.
I wonder how a focus on per-token API profits will impact the incentives to improve token efficiency and drive down costs through optimized compute. I suppose as long as a few leading labs are competing, we'll see progress in this regard, but it's certainly less in their interest than it is with a flat subscription pricing model.
Ai has become indispensable but maybe not at all cost. My company just had a company-wide meeting to talk about how they're restricting who can use which models and instructing us the "be more responsible with company's tokens". And it's not an small company by any means.
>Somehow this fragment turned into headlines like Uber’s COO says it’s getting harder to justify the money spent on AI tokenmaxxing, because the market for stories about AI failures remains enormous.
I notice this all over the place. Many people hate AI and want it to fail, and they're willing to invent misinformation if it supports that idea.
They've got, ballpark, $5t to $10t to make back in the next 5 years, or the hardware buildouts will start getting written down.
This means we're going to need $1t+ per year in spending, per year, on tokens. 200m knowledge workers in the world, 30m developers. We're talking about a world where you need 5% of every knowledge workers salary to go into tokens. 20% if you're a developer.
That's a _huge_ shift. Most people I know cite +20%-40% velocity with these tools, against the actual work their company cares about doing. +20% speed for +20% spend isn't going to motivate a trillion dollars a year in spending.
We're not there yet. This is still the upswing of the hype cycle, and unless we figure out how to make developers 2x, 5x, 10x as productive on stuff that matters, this isn't going to play out well.
Here is a serious question.. Can we sell into the hype cycle and on the way down with this: https://safebots.ai/costs.html
I asked claude to generate a frontend and it made the same template. Same san serif and serif fonts together. Same colors. Same typography. Same layout and animations even. It’s wild how similar it is. No not similar it’s the same damn thing.
That ignores depreciation, too
And third-party (i.e. Chinese) or self-hosted models.
> $5t to $10t to make back in the next 5 years
Wait what? They spent 2 order of magnitude less on hardware.
From the verge: https://archive.is/kU4Zg
> Gartner forecasts that large AI companies would need to earn cumulatively close to $7 trillion in AI-driven revenue through 2029, which is close to $2 trillion per year by the end of the period. In order to achieve “historic returns,” the providers would need to earn nearly $8.2 trillion in the same period.
Those numbers don't even track even in the same sentence. If it is $2T/year by the end of 2029, it would be something < $5T cumulative in 3 years.
You are making the assumption that the models are only used / paid for by 2.5% of the population (your knowledge workers value). There will be new value created by these models which people are happy to pay for which simply did not exist at all before. It is also naive to say that the hyperscalers are going to be expecting a return on this in 5 years, it will be entirely propped up by investments / IPOs as has been the case with any tech company for decades now to reach scale. The hyperscalers are currently spending ~650m combined annually, which they have the cash for and can sell in future compute instantly.
Source on 200 million knowledge workers worldwide? My understanding is that it's just above 1 billion. I dont think a billion subscriptions at $1000/yr is out of the question but it might take a decade to get roiling
> We're talking about a world where you need 5% of every knowledge workers salary to go into tokens.
They are assuming ~10% global GDP growth instead of ~3%. You probably don't need the same %s if the pie grows a ton.
I'm highly skeptical we get that growth, but if you aren't, it makes it easier to digest.
200$ per month per seat is nothing .
A single 3D CAD license pack for the guys in our R&D group costs multiple thousands of dollars per seat, per month.
It's about time software seats get some love too.
Yeah, it’s nothing, and it’s also not the cost that enterprises are paying. As the article states, the price is $20 per seat per month, PLUS per-token API usage. Enterprises are paying consumption billing, not fixed rate oversubscribed “all you can eat per seat.”
AutoCAD is $175 per user per month [1].
[1] https://www.autodesk.com/products/autocad/buy
How many guys is that? Every single white color worker is in the AI ICP.
> $2,180.16 worth of tokens for $200
“Tokens” don’t have an intrisic cost or value. Saying that I used $2,180.16 worth of tokens is like relying on the salesperson to convince me I’m getting a billion dollars worth of pots and pans for $19.99.
I think it’s funny how we are throwing critical thinking out the window when it comes to evaluating biased sources of info.
I'm not sure what you're pushing back against here.
I spent $200. If I had been paying API pricing it would have been $2,180.16. The article is about how enterprise customers get charged API pricing, which means if I had been employed by one of those companies I would have cost them $2,180.16.
What am I missing?
Large enterprises make deals and won’t be paying 2,180.16$ either. Just like with AWS
That doesn't seem to be the case. From what I've seen enterprise deals get API pricing now. Have you seen evidence that's not true?
And "large" just means that AWS will assign an account manager to talk with you. I was at a start-up who spent $300k/year on AWS and that was enough to get special attention and discounts. Enterprise pricing is confusing.
The point is that those a real prices real people are paying for real API usage. it's not made up.
your point is large players won't pay those prices at massive volume. ok
> If I had been paying API pricing it would have been $2,180.16
The point being made above is that API pricing is calculated... somehow... seemingly arbitrarily. Possibly untethered to the infrastructure costs entirely: which would be the basis of any 'value', however that holds the labor theory of value, which isn't accurate either. So how do you accurately price these tokens at all (other than through price-discovery: which is slow, messy and fuzzy)?
> So how do you accurately price these tokens at all
Like anything else in the economy: at the point where enough customers can pay you, and not enough will go to the cheaper competition.
API pricing drops DRAMATICALLY in enterprise agreements.
As with pretty much anything priced on volume/usage.
Enterprise deals are negotiated ad-hoc, the listed pricing is simply a jumping off point for the final negotiated discount.
If you’re going to give 20,000 employees Claude code you are not going to be spending $1B per year on Anthropic tokens as if you gave everyone an individual API key.
Do you have any numbers or reports to back that up?
Tokens do have a clearly calculable intrinsic cost. There's the marginal cost of production (i.e. the inference cost) and the amortized R&D cost that goes into the model producing them.
Yes, value is hard to calculate, but luckily market pricing mechanisms exist exactly for this purpose. There isn't a better number to use than what people are willing to pay for them.
So he's saying that on an enterprise plan, he'd be spending $2,180.16. He's not paying that much, but enterprises are.
a little critical thinking led me to read that sentence as $2180 worth of tokens [at current api pricing]
I think it's funnier that you can believe some things have an intrinsic cost and others don't
Does this analysis factor in potential caching of tokens on the server side? It seems that if they organize things well (as a model provider), they can save quite a lot on that. Looking at my Cursor statistics makes it clear that the token calculations are not at all trivial.
I believe the ccusage tool I used takes cached token pricing into account.
The costs are exorbitant and most software is not produced by companies with such a huge moat. Anthropic made a profit through their recent bait amd switch pricing. There is zero useful insights online to indicate whether this might die due to commoditisation with good enough open models or fail the race to get more people subsidising unsustainable growth with other people’s money. Who knows? In any case they dont seem to be able to drop usage costs so the business model seems based on wishes
Usage costs will come down with better hardware. Hardware is improving rapidly each generation.
That trend held true for the past three years, but it doesn't feel as safe to me now.
But memory costs are going way up. And both OpenAI and Anthropic bumped up the price of their frontier models in April.
[delayed]
With deepseek and xiaomi mimo models slashing their prices 99%, I don't see a great future for openai / antrhopic with regards to their 1T valuations. Maybe 1T valuation will be the whole market, West + East.
They'll still have their dedicated enterprise customers. I think the Chinese providers will pull more of the single users who're paying their own way, than those backed by company budget. And it's a pretty good split as the demand becomes better distributed, resulting in better service (I'll never forgot must how bad access to Claude became until they got access to Colossus) and less potential for lock-in (we really don't want there to be a duopoly, etc on good AI).
So how do openai and anthropic plan to keep customers when GLM-5.1 is just as good and open source and a lot cheaper?
I don't see the business model working. My closest friend actually does automation software for large companies.
He does not use Claude or openai at all. He primarily uses gpt 120b on cerebras and glm-5.1 for heavy thinking work. And some other small models for various tasks. All open source.
And these systems are extremely useful for the businesses and are able to run fully automated pipelines that are very stable and fast.
We discuss this a lot, and we both think any business doing heavy agentic work on Claude and openai just aren't aware of exactly how good and cheap open source has gotten on the last year.
So... once the legacy businesses and developers catch up, won't Claude and openai be unable to recoup their costs?
For coding you always want to go with the best model in the category, not something that would be the best model if we went 1 year back which GLM 5.1 is, and I'm saying that as a big fan of GLM cause I run a translation site where GLM is good enough for the price.
Most of the money right now is in coding. Openai and Anthropic just have to be 6 months ahead of SOTA open source models and they'll capture most of the enterprise and dev market
Most work is not coding.
And also, people have it wrong… their models are not the main problem anymore. It’s the RAG
Depending on RAG is a workflow problem, not an AI problem
For coding like for everything else in life cost is a factor.
Yes I'm an engineer (20 years most in games/graphics industry) and only use it for code. I've been using glm 5.1 this week a lot. I went in expecting another "decent" but not really "up to standard" open source model.
I highly doubt I'll ever use Claude again.
I think you are wrong about Claude being any significant level better
For coding assistance, I have tried OpenCode with several large open models through OpenRouter. All were fairly bad compared to Claude Opus. Could you provide some hints on how I should be holding these open models so that I might get more value out of them?
I agree with the common trope that open models lag behind by about a year, but something magical happened just around a year ago when the state of the art models became extremely useful. By this reasoning we're about to see open models perform well, but I'm afraid there is more to it than just waiting for another revolution around the sun.
Note, my application is coding assistance. Open models can be great for other purposes.
Great article I know this upsets a lot of people who are used to thinking Anthropic/OpenAI are just lighting cash on fire but they've cornered the market on enterprise who cannot walk away from these $200/month plans
However the valuations are still far far away from actual sanity
Have you tried the large open source code models?
I use glm-5.1 and occasionally deep seek v4.
They are as good or better than Claude's latest models.
And significantly cheaper. I've converted 3 of my engineer friends as well. All three have dropped their $200 month plans they had with anthropic.
We've all been a bit shocked at just how good these models are now.
If you "have" tried GLM (I specifically find it shockingly good for code). Did you not think it's not competitive to Claude, and why?
I use GLM-5.1.
It's good enough for personal stuff. It doesn't compare to the latest Opus I use at work. You can certainly argue I don't need Opus for work, but there is clearly a difference.
Also, at least with z.ai, GLM-5.1 is s l o w! After using Claude at work, I get really impatient with GLM-5.1 at home. When doing "true" vibe coding (i.e. not really examining the code), Opus is a ton faster (easily 5x).
But yeah, I'm not willing to personally pay for the frontier models. I won't even renew my annual Z.ai plan - it's become too expensive.
task i am working on right now at work is comparing two verisions of apis and documenting responses in their outputs. i suspect a vast majority of work at entrprise is of similar complexity.
right now everyone is using latest and greatest to do dumb stuff like that. that would change fast if companies start caring about costs.
> enterprise who cannot walk away from these $200/month plans
But that's the point of the article. Enterprise plans are starting to get API pricing, not the subsidized subscription pricing.
How is the lack of bad news declaring a victory for AI? I am yet to see any company concretely publish analysis about the ROI from AI. Most companies as far as I know are still treating AI investment as sunk cost with no expectation of returns at the moment. We could very well see a world where companies heavily scale back investment.
If nothing else this blog did give me the idea that I should split my $200 claude max plan into two $100 CC max and $100 codex plan, esp because Claude is now offering 1.5x weekly limits so its the 5x usage is now more like 7.5x usage.
I think the reasons for them going with API pricing will become abundantly clear when the S-1s become available. If they don't have a story covering how they can get revenue closer to expenses, then they're relying on the market to believe the pixie dust version of their profitability story, which I think people increasingly don't.
So it largely sounds like many more people will be able to write software - and will use AI to do it. Existing software engineers will continue to automate their tasks away like they always did, but perhaps at a faster rate.
The impact of AI in other fields seems to be muted.
I think it is applicable to a much wider range of knowledge work, but it's also harder to apply there.
Software development has the huge advantage that mistakes and hallucinations are very easy to spot: the software works or it doesn't.
Spotting errors in a research report or legal brief is a whole lot harder!
But... non-software professionals spend a huge amount of their time on tasks that can be safely automated - reformatting documents, extracting numbers from PDFs, all kinds of flavor of data entry.
Learning how to use a tool like Claude Cowork can take a big dent out of those.
Realistically, OpenAI found product market fit with the OpenAI API playground in 2021. People were using that as ChatGPT at the time.
I wonder how a focus on per-token API profits will impact the incentives to improve token efficiency and drive down costs through optimized compute. I suppose as long as a few leading labs are competing, we'll see progress in this regard, but it's certainly less in their interest than it is with a flat subscription pricing model.
Ai has become indispensable but maybe not at all cost. My company just had a company-wide meeting to talk about how they're restricting who can use which models and instructing us the "be more responsible with company's tokens". And it's not an small company by any means.
>Somehow this fragment turned into headlines like Uber’s COO says it’s getting harder to justify the money spent on AI tokenmaxxing, because the market for stories about AI failures remains enormous.
I notice this all over the place. Many people hate AI and want it to fail, and they're willing to invent misinformation if it supports that idea.
> Anthropic are strongly rumored to be about to have their first profitable quarter.
Is that quarter same as any other quarter in terms of infrastructure costs (e.g. are there any temporary discounts happening coincidentally)?