Numbers are always interesting, assuming they're real, but I just want to comment on the Cursor thing: Zitron has been insisting for 6 months that Anthropic screwed Cursor somehow by raising prices on them but the claim has always been gibberish. It's not that it's false, it's that it's impossible to figure out what Zitron claims happened. He cannot describe (here or in https://www.wheresyoured.at/anthropic-and-openai-have-begun-...) what the bad change actually was. We know everyone moved to more usage-oriented pricing earlier this year. He cannot explain why this was a price increase for Cursor. He is unable to draw a distinction between a price increase for end users (it's not even clear that it was a price increase for the average end user) and a price increase for Cursor.
Ed has constantly done this, and it's a shame because it has taken the air out of the room for real AI criticism. Most of Ed's criticism comes from a place of giving a narrative to people who are wishing for a magic bullet that makes ChatGPT vanish tomorrow rather than actually pressuring companies about the harms this technology can cause. This in part is why his writing so often focuses on perceived financial issues (despite his lack of credentials in financial journalism) rather than the social harms the technologies cause today (slop, delusions, manipulated truth).
Ed's mentioned ARR in previous articles and it's not a "generally accepted accounting principle". They cherry pick the highest monthly revenue number and multiply that by 12, but that's not their actual annual revenue.
"Cherry pick the highest" is misleading. If your revenue is growing 10% a month for a year straight and is not seasonal, picking any other than the most recent month to annualize would make no sense.
If a company's revenue in January is $100 and it grows by 10% every month, the December revenue is $285. The year's revenue would be about $2,138, but ARR in December would be $3,423. That's 1.6x the actual revenue.
ARR could be a useful tool to help predict future revenue, but why not simply report on actual revenue and suggest it might increase in the next year? I have found the most articles to be unclear to the reader about what ARR actually represents.
Why is the calendar year the relevant unit? If you insist on years, then if you consider the year from June to June, $2,138 would be misleading small.
The point of ARR is to give an up to date measure on a rapidly changing number. If you only report projected calendar year revenue, then on January 1 you switch from reporting 2025 annual revenue to 2026 projected revenue, a huge and confusing jump. Why not just report ARR every month? It's basically just a way of reporting monthly revenue — take the number you get and divide it by 12.
I am really skeptical that people are being bamboozled by this in some significant way. Zitron does far more confusing things with numbers in the name of critique.
You're correct, ARRs can be both misleading and for any 12-month period (I just chose a year to simplify), but the problem is AI companies tend to only release their latest ARR, and only selectively, which I believe is misleading in the opposite direction.
Well people like Sam Altman have not been entirely honest and there's a reason they're not sharing their actual revenue numbers. If they could show they were growing 10% every month they would.
Eh, when you have a company that’s growing, picking the highest and annualizing it is sensible. If we had a mature company with highly seasonal revenue it would be dishonest.
I mean I think there are instances where OpenAI's revenue is seasonal. Lots of students using it during the school year and cancelling it during summer.
I think you missed the forest for the trees. I am sure the student population has some dropoff during summer months but the point is that for businesses that a growing month over month which most of these have since creation, you take the highest number (latest) and annualize it.
I am also willing to bet that the student dropoff is not pronounced. I am more thinking of a business that sells beach umbrellas, they make a lot of sales in the summer months and then next to nothing in the winter months. That would be dishonest.
The article Zitron links says Cursor has single-digit millions of cash burn with about $1B in the bank (as of August). Assuming that is true, they are losing money but have a long runway.
That article says "Anysphere runs pretty lean with around 150 employees and has a single digit monthly cash burn, a source tells me." That would be total cash burn, i.e., net losses. If their AWS bill is bigger than that it's because they are making up for part of it with revenue.
TFA claims Cursor "obtains the majority of its compute from Anthropic — with AWS contributing a relatively small amount" and therefore only claims that for Cursor the AWS number indicates a "direction of travel" for compute costs. (Debatable whether it does indicate even that, ofc.)
> through September, Anthropic has spent more than 100% of its estimated revenue (based on reporting in the last year) on Amazon Web Services, spending $2.66 billion on compute on an estimated $2.55 billion in revenue.
Well I don't have to scratch my head any longer and wonder why Amazon hasn't jumped on the AI bandwagon with their own Gemini or whatever. They are sitting pretty and selling shovels and pickaxes to the AI fools. Not a bad strategy for them...
Why do they need to compete? AI at Amazon should be laser-focussed on two things: selling compute time to AI Wannabees, and up-selling stuff to me in the shop.
I agree. Unfortunately AWS leadership is in an all out panic to compete, diverting attention away from core services. I wish they’d just get back to focusing on the foundational stuff!
AWS is not winning selling compute to AI folks. GCP and Azure have both done better selling AI services, which is part of the problem. Amazon is still the king for basic compute, but fell behind on the AI infra.
How are you getting 5? In recent memory, there's only 3: dotcom bubble (2000), GFC (2008), covid (2020). You'd either have to go back in time even more (eg. savings and loan crisis in the 80s/90s), or include regional ones (eg. eurozone crisis) to hit 5.
I don't think anyone is expecting Amazon (or Google or Microsoft for that matter) to be taken out by the AI bubble.
I would expect to see OpenAI, Anthropic, and a lot of the little tool wrappers to get taken out though, or at least acquired for pennies on the dollar when it bursts.
But like the last one, it's going to be us, the tax payers, that are left holding the bag.
Why taxpayers? Where’s the systemic risk in AI labs getting acquired for cents on the dollar? The taxpayers weren’t holding the bag during the dot com crash, just investors.
During the dot com crash, none of the companies were "too big to fail" This is looking more like 2008 than dotcom, the entire market is being propped up by basically Nvidia. US GOV is likely to bail them out for "national security"
this is the latest mantra from people who have missed the boat. i'm like lol do you think that industry didn't learn anything (about financing structures) from the last one?
That doesn't tell us what percentage they own, though. When you invest in a company and the value goes up, your percentage doesn't change (unless there are additional investors)
They do and they've had her for a while, but from what I understand that's "so last Tuesday" in the AI race. It was ok in the race with "Ok Google" and "Siri" but not competing with OpenAI, Anthropic, Gemini.
One day we will westerners will learn why the Chinese are releasing models that are optimized for cost of training n yet good enough to run locally or cheaply.
when the music stops, suddenly a lot of people won't just sit on the ground but plunge into the depths of hell.
Yeah, I'm one of them, using Qwen 3 coder on Cerebras as coding agent through CC. What I keep hearing is (very ballpark anecdata)...
50% of people into coding agents are quite concerned about that last mile in difference with frontier models that they "can't afford to lose" - my experience tells me otherwise, the difference is negligible once you have a good setup going and knows how to tune your model + agent.
The other 50% don't give a damn, they just landed, or got locked, into some deal for a coding agent and are happy with what they got, so why change? These deals arrived from the big model providers and resellers first, so Chinese arrived late and with too little to the party.
Running Chinese models (for coding) requires many things that you need to figure-out yourself. Are you running the model on your hw or through a provider? Are you paying by token or on a plan? Does the model pair well with you agent CLI/IDE of choice? (Zed, Cline, Opencode, etc) Does it even work with your favorite tool? (tool calling is very wobbly) Is it fast (tps)? Is it reliable? How do you do "ultrathink" with a secondary model? How do you do "large context"? Does it include a cache or are you going to eat through the plan in 1hr/day? What context size are you getting? Does it include vision and web search or do you have to get another provider/mcp for that? And, yeah, is it in a territory where you can send your client's code to? A lot to grok.
Cerebras Coder Max is really cool if you want to hack your way through this, but they couldn't care less about your experience: no cache, no tool endpoint fine-tuning, no plans or roadmap on updating models, on increasing context windows, adding vision, or anything really. They just deleted some of the tools they were recommending out of the website (ie Cursor) as they got reports of things that stopped working.
i'm starting a new trend: ask every person that is so certain about the negative outlook how big their short position is. so how big is your short position? please let us know.
> Based on discussions with sources with direct knowledge of their AWS billing, I am able to disclose the amounts that AI firms are spending, specifically Anthropic and AI coding company Cursor, its largest customer.
I think this is a minor speed bump and VC’s believe that cost of inference will decrease over time and this is a gold rush to grab market share while cost of inference declines.
I don’t think they got it right and the market share and usage grew faster than inference dropped, but inference costs will clearly drop and these companies will eventually be very profitable.
Reality is that startups like this assume moore’s law will drop the cost over time and arrange their business around where they expect costs to be and not where costs currently are.
> I think this is a minor speed bump and VC’s believe that cost of inference will decrease over time and this is a gold rush to grab market share while cost of inference declines.
It could also be that you give too much credit to the market. People follow trends because in most cases that makes money. There is no other deeper though involved. Look at the financial crisis, totally irrational.
This. Post-crypto, AI was the obvious next gambit for VC. Their money flows to hype, not product. The second that hype starts to fade and the money dries up, VCs will be running with their Harold Hill trunk full of cash toward the border. Just from the content they publish alone you can tell they're channeling their inner Barnum & Bailey in between Ayahuasca seizures.
Isn't the consensus that the MOE architecture and other optimizations in the newest gen models (GPT-5, Gemini 3.0 to come, etc) will reduce inference costs by 50-75% already?
Kind of. Frontier LLMs aren't going to get cheaper, but that's because the frontier keeps advancing.
Price-performance though? The trend is clear: a given level of LLM capability keeps getting cheaper, and that trend is expected to hold. Improvements in architecture and training make LLMs more capability-dense, and advanced techniques make inference cheaper.
My own usage and the usage of pretty much everyone I know says that as inference costs drop, usage goes up in lockstep, and I’m still nowhere near the ceiling of how many tokens I could use if they were free.
I think if these companies are gambling their future on COGS going down, that’s a gamble they’re going to lose.
They haven't though. On two fronts: 1, the soa models have been pretty constantly priced, and everyone wants the soa models. Likely the only way costs drop is the models get so good that people are like hey, I'm fine with a less useful answer (which is still good enough) and that seems, right now, like a bad bet.
and 2 - we use a lot more tokens now. No more pasting Q&A into a site; now people hammer up chunks of their codebases and would love to push more. More context, more thinking, more everything.
You're describing increased spending while calling it increased cost. These aren't the same thing. A task that cost me $5 to accomplish with GPT-4 last year might cost $1 with Sonnet today, even though I'm now spending $100/month total on AI instead of $20 because I'm doing 100x more tasks. The cost per task dropped 80%. My spending went up 5x. Both statements are true.
Fwiw that's not necessarily true, because if Sonnet ends up using reasoning, then you are using more tokens than GPT-4 would have used for the same task. Same with GPT-5 since it will decide (using an LLM) if it should use the thinking model for it (and you don't have as much control over it).
Right, but if I understand you, the counterargument is dumb since the context in which we are discussing is business viability (vcs investing in businesses where the unit economics require inference cost decreases), so actual dollars out rather than some imaginary cost per token is the metric that matters.
Inference is getting so much cheaper that cursor and zed have had to raise prices.
Why do the unit economics require a decrease in inference spend per user? This is discussed at the end of the post. I think this is based on the very strange assumption that these businesses must charge $20 a month no matter how much inference their customers want to use. This is precisely what the move to usage-based pricing was about. End users want to use more inference because they like it so much, and are knocking down these companies’ doors demanding to be allowed to pay them more money to get more inference.
> so actual dollars out rather than some imaginary cost per token is the metric that matters.
Even if we take this as true, the point is that this is different than "the cost of inference isn't going down." It is going down, it's just that people want more performance, and are willing to pay for it. Spend going up is not the same as cost going up.
I don't disagree that there are a wide variety of things to talk about here, but that means it's extra important to get what you're talking about straight.
Playing word games labeling inference narrowly as the cost per token rather than the per-X $ going to your llm api provider per customer/user/use/whatever is kinda silly?
The cost of inference -- ie $ that go to your llm api provider -- has increased and certainly appears to continue to increasing.
> The cost of inference -- ie $ that go to your llm api provider
This is the crux of it: when talking about "the cost of inference" for the purposes of the unit economics of the business, what's being discussed is not what they charge you. It's about their COGs.
That's not word games. It's about being clear about what's being talked about.
Talking about increased prices is something that could be talked about! But it's a different thing. For example, what you're talking about here is total spend, not about individual pricing going up or down. That's also a third thing!
You can't come to agreement unless you agree on what's being discussed.
Anecdote of 1. Costs for openai on a per token basis have absolutely dropped and that accounts for new sota models over time. I think by now we can all agree that inference costs from providers are largely at or above breakeven. So more tokens is a good problem to have.
Color me skeptical. We're running into the speed of light when it comes to transistor size, and the parallelism that made neural nets take off is running into power demands. Where do the exponential hardware gains come from? Optimizing the software by 2x or 4x happens only once. Then there's the other side: if Moore's Law works too well, local models will be good enough for most tasks, and these companies won't be able to do the SaaS thing.
It seems to me like models' capability scales logarithmically with size and wattage, making them the rare piece of software that can counteract Moore's Law. That doesn't seem like a way to make a trillion dollars.
One improvement is from scraping and stealing better quality IP to train on. And they can just ride Moore's law until they profit then lobby governments to require licenses for fast GPUs because of national security.
“Spend” requires a grain of salt here. AWS “invests” in Anthropic and then Anthropic buys AWS. If you follow the money with marked bills, AWS is buying this compute from itself and then claiming revenue.
There’s a lot of that sort of thing going on at the moment in the AI bubble.
I have sat with these numbers for a great deal of time, and I can’t find any evidence that Anthropic has any path to profitability outside of aggressively increasing the prices on their customers to the point that its services will become untenable for consumers and enterprise customers alike.
This is where he misunderstands. Enterprise companies will absolutely pay 10x the cost for Claude. Meta and Apple are two large customers, you think they won't pay $500 a month per employee? $1000 a month per employee? Neither of those are outrageous to imagine if it increases productivity 10%.
Also spend will drop dramatically if the models level out a bit more. The training is what's compute heavy, and if you aren't having to retrain every month, but able to use things like Skills to stay competitive your costs will drop a lot.
I suppose that's the pessimistic-on-AI side. On the other hand, once you create God little things like money are meaningless.
I've been unable to find it but early in the railroad boom did rail companies undercharge for freight in order to build demand and interest? Given that they were also given land in exchange fro building, it would have been in their interests to do so. They could also amortize rail costs over a long time as rails are useful for longer than GPUs - but I am also curious if there was an unsustainably low entry price for railroads like Ed is implying for AI tools.
Ed Zitron's style isn't for everybody, I understand that. But if these numbers, and the direction they're going in, are correct... to me this points to a significant AI bubble deflation coming soon. It just isn't sustainable, it seems.
if you are worried about AWS spend, i have news for you about non AWS spend.
"coming soon" is also really over simplistic. you would have missed some of the greatest tech companies in the past 20 years if you evaluated startups based on their early-year revenue vs infra spend
like sure i have a dog in this fight but i actually want the criticism to sharpen my thinking, unfortunately yours does not meet that bar.
the entire point is that stressing sustainable business practices here at this point in a startup's life is extremely short sighted and is the kind of shitty analysis that gives HN a bad rep.
pull up Uber's financials leading up to IPO. unsustainable and everyone knew it. they work it out after because they burned money and eventually achieved a sustainable moat. this is why venture exists. HN doesnt like venture which is, well, ironic given the domain we're on.
a better negative argument i'd rather see looks like this - "ive run these aws numbers against the typical spend path of pre IPO startups who then later improved their cost baseline and margin profile and even after accounting for all that, Anthropic ngmi". thats the kind of minimum sophistication you need here to play armchair financial analyst. ed zitron, and everyone involved in this entire thread, incl myself, have not done that because we are lazy and ignorant and dont actually care enough about seeking the truth here. we are as unprepared to analyze this AWS spend as we are to understand their 1b -> 10b revenue ramp in 2025. you havent done the work and yet you sit here and judge it unsustainable based off some shitty "leaks". dont pretend that ed's analysis is at all meaningful particularly because he conveniently stops where it supports his known negative bias.
It would be hilarious if all this was a devilish scheme to burn the excess money that was printed during COVID. Unfortunately, what is more likely is another bailout if the bubble bursts.
Just a reminder, Ed Zitron is neither an AI researcher, nor an Engineer, nor a Financial Analyst, nor an Economist nor an Insider and has ZERO clue in multiple dimensions (technology, investing, unit economics, growth, TAM) to analyze any of this
If Ed is neutral or an expert, he would have had the following analysis
"What is the unit cost of serving a Token? It is the cost of electricity + amortized cost of GPU (GPUs would have been Capex, but because of their fast depreciation rate, you can claim they should be Opex). Given this cost structure, every SOTA labs (Google, Anthropic and OpenAI) are profitable and actually have high unit margins of 50-60%."
High Unit Margins and growth means, these labs can be profitable anytime they choose to
Correct. And yet, he's provided some of the most level headed takes on the current LLM boom, to the point where FT Alphaville link to his analysis of the economics.
The jury is still out there if his analysis are level-headed or not. He says things that *some* people want to hear, but that's not level-headed
E.g
What is the unit cost of serving a Token?
It is the cost of electricity + amortized cost of GPU (GPUs would have been Capex, but because of their fast depreciation rate, you can claim they should be Opex).
Given this cost structure, every SOTA labs (Google, Anthropic and OpenAI) are profitable and actually have high-margins 50-60%.
With this margin and growth, the frontier labs can be profitable anytime they want to. But they are sacrificing profitability for growth (as they should be)
Where is Ed's analysis about this? Either he is disingenuous or clueless. Remember people who voluntarily subscribe to Ed, are coming from wanting to hear what they believe.
If he is level-headed, show me an Ed article that is positive about AI
His argument is not that AI is not useful, it's that it's not financially sustainable at the current prices being charged for it - and additionally AI start-ups have 0 moat. Both of which likely are true.
As the saying goes, "the market can remain irrational longer than you can remain solvent." I hope the next fad is to invest in boring companies that employ people and give steady returns, and they can stop replacing all the stores in my city with Dollar Trees.
Numbers are always interesting, assuming they're real, but I just want to comment on the Cursor thing: Zitron has been insisting for 6 months that Anthropic screwed Cursor somehow by raising prices on them but the claim has always been gibberish. It's not that it's false, it's that it's impossible to figure out what Zitron claims happened. He cannot describe (here or in https://www.wheresyoured.at/anthropic-and-openai-have-begun-...) what the bad change actually was. We know everyone moved to more usage-oriented pricing earlier this year. He cannot explain why this was a price increase for Cursor. He is unable to draw a distinction between a price increase for end users (it's not even clear that it was a price increase for the average end user) and a price increase for Cursor.
Ed has constantly done this, and it's a shame because it has taken the air out of the room for real AI criticism. Most of Ed's criticism comes from a place of giving a narrative to people who are wishing for a magic bullet that makes ChatGPT vanish tomorrow rather than actually pressuring companies about the harms this technology can cause. This in part is why his writing so often focuses on perceived financial issues (despite his lack of credentials in financial journalism) rather than the social harms the technologies cause today (slop, delusions, manipulated truth).
Zitron is too much of a small player to "suck the air off other criticism of ai".
Claiming that a single journalist blog has power to stop others from criticiaing ai for different reasons ia kind of absurd.
The buried lede:
Anthropic: "$2.66 billion on compute on an estimated $2.55 billion in revenue"
Cursor: "bills more than doubled from $6.2 million in May 2025 to $12.6 million in June 2025"
Clickthrough if you want the analysis and caveats
I could find an ARR for Cursor of $500M. Why do they they in this article that Cursor is loosing with this spending number ?
Ed's mentioned ARR in previous articles and it's not a "generally accepted accounting principle". They cherry pick the highest monthly revenue number and multiply that by 12, but that's not their actual annual revenue.
"Cherry pick the highest" is misleading. If your revenue is growing 10% a month for a year straight and is not seasonal, picking any other than the most recent month to annualize would make no sense.
If a company's revenue in January is $100 and it grows by 10% every month, the December revenue is $285. The year's revenue would be about $2,138, but ARR in December would be $3,423. That's 1.6x the actual revenue.
ARR could be a useful tool to help predict future revenue, but why not simply report on actual revenue and suggest it might increase in the next year? I have found the most articles to be unclear to the reader about what ARR actually represents.
Why is the calendar year the relevant unit? If you insist on years, then if you consider the year from June to June, $2,138 would be misleading small.
The point of ARR is to give an up to date measure on a rapidly changing number. If you only report projected calendar year revenue, then on January 1 you switch from reporting 2025 annual revenue to 2026 projected revenue, a huge and confusing jump. Why not just report ARR every month? It's basically just a way of reporting monthly revenue — take the number you get and divide it by 12.
I am really skeptical that people are being bamboozled by this in some significant way. Zitron does far more confusing things with numbers in the name of critique.
You're correct, ARRs can be both misleading and for any 12-month period (I just chose a year to simplify), but the problem is AI companies tend to only release their latest ARR, and only selectively, which I believe is misleading in the opposite direction.
Because that's a part of the generally accepted accounting principles: https://www.rightrev.com/gaap-revenue-vs-arr/
Nobody considers a year from June to June because that would be misleading.
That is an article explaining why ARR is useful and important despite not being the same thing as GAAP revenue.
How can you talk about ARR if you only look at 1 year?
It's useful for financial planning. Less useful for overall financial reporting given how volatile it is.
Well people like Sam Altman have not been entirely honest and there's a reason they're not sharing their actual revenue numbers. If they could show they were growing 10% every month they would.
They are sharing their actual revenue numbers. That's what ARR is. Take the number and divide it by 12 and that's monthly revenue.
It's literally not ARR is the highest monthly revenue times 12. Dividing it by 12 doesn't get you the actual, on the books monthly revenue numbers.
Eh, when you have a company that’s growing, picking the highest and annualizing it is sensible. If we had a mature company with highly seasonal revenue it would be dishonest.
I mean I think there are instances where OpenAI's revenue is seasonal. Lots of students using it during the school year and cancelling it during summer.
The graph that was widely shared to make this claim was from OpenRouter and did not represent ChatGPT usage in any way.
I think you missed the forest for the trees. I am sure the student population has some dropoff during summer months but the point is that for businesses that a growing month over month which most of these have since creation, you take the highest number (latest) and annualize it.
I am also willing to bet that the student dropoff is not pronounced. I am more thinking of a business that sells beach umbrellas, they make a lot of sales in the summer months and then next to nothing in the winter months. That would be dishonest.
Then why aren't AI companies reporting their actual monthly revenues?
They are. That is what ARR is.
The article Zitron links says Cursor has single-digit millions of cash burn with about $1B in the bank (as of August). Assuming that is true, they are losing money but have a long runway.
https://www.newcomer.co/p/cursors-popularity-has-come-at-a
Single-digit cash burn on AWS, which the article says is only a small part of its compute, with the majority coming from Anthropic.
That article says "Anysphere runs pretty lean with around 150 employees and has a single digit monthly cash burn, a source tells me." That would be total cash burn, i.e., net losses. If their AWS bill is bigger than that it's because they are making up for part of it with revenue.
Ah, gotcha. I misunderstood your comment.
For context, AWS's total revenue for 2024 was $107.6B: https://ir.aboutamazon.com/news-release/news-release-details...
specific clarification: That was only Cursor's AWS bill. If they are using other providers, wasn't clear.
TFA claims Cursor "obtains the majority of its compute from Anthropic — with AWS contributing a relatively small amount" and therefore only claims that for Cursor the AWS number indicates a "direction of travel" for compute costs. (Debatable whether it does indicate even that, ofc.)
> through September, Anthropic has spent more than 100% of its estimated revenue (based on reporting in the last year) on Amazon Web Services, spending $2.66 billion on compute on an estimated $2.55 billion in revenue.
Well I don't have to scratch my head any longer and wonder why Amazon hasn't jumped on the AI bandwagon with their own Gemini or whatever. They are sitting pretty and selling shovels and pickaxes to the AI fools. Not a bad strategy for them...
Amazon trained their own models like Nova and has AI coding assistants like Amazon Q, but I don't know anyone outside Amazon who is using them.
Yes, the Amazon AI stuff isn’t great. They don’t really have the right leadership or talent to do anything particularly competitive there.
Why do they need to compete? AI at Amazon should be laser-focussed on two things: selling compute time to AI Wannabees, and up-selling stuff to me in the shop.
Everything else is expense.
I agree. Unfortunately AWS leadership is in an all out panic to compete, diverting attention away from core services. I wish they’d just get back to focusing on the foundational stuff!
AWS is not winning selling compute to AI folks. GCP and Azure have both done better selling AI services, which is part of the problem. Amazon is still the king for basic compute, but fell behind on the AI infra.
Amazon owns 15-19% of Anthropic. So yes and no.
GenAI financing is a flat circle. The bubble bursting is going to have a huge blast radius.
They survived the dot-com bubble, I don't see the AI bubble taking out Amazon.
It might take out your 401k for a decade.
I'm on like my fifth once in a generation financial crisis.
At a certain point you just expect it.
How are you getting 5? In recent memory, there's only 3: dotcom bubble (2000), GFC (2008), covid (2020). You'd either have to go back in time even more (eg. savings and loan crisis in the 80s/90s), or include regional ones (eg. eurozone crisis) to hit 5.
Are you counting eurozone as 2011? I think it counts enough as American if it caused a downgrade in the USG's credit rating.
[1]: https://en.wikipedia.org/wiki/August_2011_stock_markets_fall
Maybe they had strong exposure to Worldcom or Enron?
If you were 100 then you'd be right on pace. Or slightly ahead. Apparently a generation is 25-30 years. I think they should update that
. . . once in a [news media financial analyst] generation financial crisis.
Its like fruit fly generations, not 20-30 year human cohort generations.
I don't think anyone is expecting Amazon (or Google or Microsoft for that matter) to be taken out by the AI bubble.
I would expect to see OpenAI, Anthropic, and a lot of the little tool wrappers to get taken out though, or at least acquired for pennies on the dollar when it bursts.
But like the last one, it's going to be us, the tax payers, that are left holding the bag.
Why taxpayers? Where’s the systemic risk in AI labs getting acquired for cents on the dollar? The taxpayers weren’t holding the bag during the dot com crash, just investors.
During the dot com crash, none of the companies were "too big to fail" This is looking more like 2008 than dotcom, the entire market is being propped up by basically Nvidia. US GOV is likely to bail them out for "national security"
this is the latest mantra from people who have missed the boat. i'm like lol do you think that industry didn't learn anything (about financing structures) from the last one?
All they will learn is how to shift more of the risk on to retail investors and B tier institutional investors.
I guess by this logic we should never have a depression/recession/bubble burst ever again? We always learn from our mistakes!
No way their share is so high. Where did you get these numbers from?
it's a number 2 years out of date. they invested 8b and ant is now worth 183b.
That doesn't tell us what percentage they own, though. When you invest in a company and the value goes up, your percentage doesn't change (unless there are additional investors)
that's really naive im afraid. you have to take pro rata or the percentage goes down. amzn did not take pro rata.
They have their own models, the Nova series, although my experience has been pretty mixed with them.
They actually do have Alexa.
They do and they've had her for a while, but from what I understand that's "so last Tuesday" in the AI race. It was ok in the race with "Ok Google" and "Siri" but not competing with OpenAI, Anthropic, Gemini.
One day we will westerners will learn why the Chinese are releasing models that are optimized for cost of training n yet good enough to run locally or cheaply.
when the music stops, suddenly a lot of people won't just sit on the ground but plunge into the depths of hell.
I keep hearing this but I don’t know of many folks utilizing Chinese models, even those hosted in an agreeable territory.
Yeah, I'm one of them, using Qwen 3 coder on Cerebras as coding agent through CC. What I keep hearing is (very ballpark anecdata)...
50% of people into coding agents are quite concerned about that last mile in difference with frontier models that they "can't afford to lose" - my experience tells me otherwise, the difference is negligible once you have a good setup going and knows how to tune your model + agent.
The other 50% don't give a damn, they just landed, or got locked, into some deal for a coding agent and are happy with what they got, so why change? These deals arrived from the big model providers and resellers first, so Chinese arrived late and with too little to the party.
Running Chinese models (for coding) requires many things that you need to figure-out yourself. Are you running the model on your hw or through a provider? Are you paying by token or on a plan? Does the model pair well with you agent CLI/IDE of choice? (Zed, Cline, Opencode, etc) Does it even work with your favorite tool? (tool calling is very wobbly) Is it fast (tps)? Is it reliable? How do you do "ultrathink" with a secondary model? How do you do "large context"? Does it include a cache or are you going to eat through the plan in 1hr/day? What context size are you getting? Does it include vision and web search or do you have to get another provider/mcp for that? And, yeah, is it in a territory where you can send your client's code to? A lot to grok.
Cerebras Coder Max is really cool if you want to hack your way through this, but they couldn't care less about your experience: no cache, no tool endpoint fine-tuning, no plans or roadmap on updating models, on increasing context windows, adding vision, or anything really. They just deleted some of the tools they were recommending out of the website (ie Cursor) as they got reports of things that stopped working.
i'm starting a new trend: ask every person that is so certain about the negative outlook how big their short position is. so how big is your short position? please let us know.
here's another trend: every time a person is on this hype bandwagon ask them how much are they invested in nvidia/ms/openai/etc
I am not invested in anything except popcorn to watch it burst;)
> Based on discussions with sources with direct knowledge of their AWS billing, I am able to disclose the amounts that AI firms are spending, specifically Anthropic and AI coding company Cursor, its largest customer.
so he got a leaked copy of their AWS bills?
Usually, the reporter inspects the document but does not get to take a copy
I think this is a minor speed bump and VC’s believe that cost of inference will decrease over time and this is a gold rush to grab market share while cost of inference declines.
I don’t think they got it right and the market share and usage grew faster than inference dropped, but inference costs will clearly drop and these companies will eventually be very profitable.
Reality is that startups like this assume moore’s law will drop the cost over time and arrange their business around where they expect costs to be and not where costs currently are.
> I think this is a minor speed bump and VC’s believe that cost of inference will decrease over time and this is a gold rush to grab market share while cost of inference declines.
It could also be that you give too much credit to the market. People follow trends because in most cases that makes money. There is no other deeper though involved. Look at the financial crisis, totally irrational.
This. Post-crypto, AI was the obvious next gambit for VC. Their money flows to hype, not product. The second that hype starts to fade and the money dries up, VCs will be running with their Harold Hill trunk full of cash toward the border. Just from the content they publish alone you can tell they're channeling their inner Barnum & Bailey in between Ayahuasca seizures.
Isn't the consensus that the MOE architecture and other optimizations in the newest gen models (GPT-5, Gemini 3.0 to come, etc) will reduce inference costs by 50-75% already?
Kind of. Frontier LLMs aren't going to get cheaper, but that's because the frontier keeps advancing.
Price-performance though? The trend is clear: a given level of LLM capability keeps getting cheaper, and that trend is expected to hold. Improvements in architecture and training make LLMs more capability-dense, and advanced techniques make inference cheaper.
Sounds interesting, do you have some links with more info about this?
Thanks!
> inference costs will clearly drop and these companies will eventually be very profitable.
Inference costs for old models will drop, but inference costs may stay the same if models continue to improve.
No guarantee that any wrapper for inference will be able to hold on to customers when they stop selling $1.00 for $0.50.
My own usage and the usage of pretty much everyone I know says that as inference costs drop, usage goes up in lockstep, and I’m still nowhere near the ceiling of how many tokens I could use if they were free.
I think if these companies are gambling their future on COGS going down, that’s a gamble they’re going to lose.
> inference costs will clearly drop
They haven't though. On two fronts: 1, the soa models have been pretty constantly priced, and everyone wants the soa models. Likely the only way costs drop is the models get so good that people are like hey, I'm fine with a less useful answer (which is still good enough) and that seems, right now, like a bad bet.
and 2 - we use a lot more tokens now. No more pasting Q&A into a site; now people hammer up chunks of their codebases and would love to push more. More context, more thinking, more everything.
You're describing increased spending while calling it increased cost. These aren't the same thing. A task that cost me $5 to accomplish with GPT-4 last year might cost $1 with Sonnet today, even though I'm now spending $100/month total on AI instead of $20 because I'm doing 100x more tasks. The cost per task dropped 80%. My spending went up 5x. Both statements are true.
Here's an analogy you may understand:
https://crespo.business/posts/cost-of-inference/
Fwiw that's not necessarily true, because if Sonnet ends up using reasoning, then you are using more tokens than GPT-4 would have used for the same task. Same with GPT-5 since it will decide (using an LLM) if it should use the thinking model for it (and you don't have as much control over it).
This is addressed in the post.
Right, but if I understand you, the counterargument is dumb since the context in which we are discussing is business viability (vcs investing in businesses where the unit economics require inference cost decreases), so actual dollars out rather than some imaginary cost per token is the metric that matters.
Inference is getting so much cheaper that cursor and zed have had to raise prices.
Why do the unit economics require a decrease in inference spend per user? This is discussed at the end of the post. I think this is based on the very strange assumption that these businesses must charge $20 a month no matter how much inference their customers want to use. This is precisely what the move to usage-based pricing was about. End users want to use more inference because they like it so much, and are knocking down these companies’ doors demanding to be allowed to pay them more money to get more inference.
> so actual dollars out rather than some imaginary cost per token is the metric that matters.
Even if we take this as true, the point is that this is different than "the cost of inference isn't going down." It is going down, it's just that people want more performance, and are willing to pay for it. Spend going up is not the same as cost going up.
I don't disagree that there are a wide variety of things to talk about here, but that means it's extra important to get what you're talking about straight.
Playing word games labeling inference narrowly as the cost per token rather than the per-X $ going to your llm api provider per customer/user/use/whatever is kinda silly?
The cost of inference -- ie $ that go to your llm api provider -- has increased and certainly appears to continue to increasing.
see also https://ethanding.substack.com/p/ai-subscriptions-get-short-...
> The cost of inference -- ie $ that go to your llm api provider
This is the crux of it: when talking about "the cost of inference" for the purposes of the unit economics of the business, what's being discussed is not what they charge you. It's about their COGs.
That's not word games. It's about being clear about what's being talked about.
Talking about increased prices is something that could be talked about! But it's a different thing. For example, what you're talking about here is total spend, not about individual pricing going up or down. That's also a third thing!
You can't come to agreement unless you agree on what's being discussed.
Anecdote of 1. Costs for openai on a per token basis have absolutely dropped and that accounts for new sota models over time. I think by now we can all agree that inference costs from providers are largely at or above breakeven. So more tokens is a good problem to have.
Point 2 was the analysis I saw. Context size and token cost grow inversely at a rate that keeps prices constant, almost like supply and demand curves.
Color me skeptical. We're running into the speed of light when it comes to transistor size, and the parallelism that made neural nets take off is running into power demands. Where do the exponential hardware gains come from? Optimizing the software by 2x or 4x happens only once. Then there's the other side: if Moore's Law works too well, local models will be good enough for most tasks, and these companies won't be able to do the SaaS thing.
It seems to me like models' capability scales logarithmically with size and wattage, making them the rare piece of software that can counteract Moore's Law. That doesn't seem like a way to make a trillion dollars.
One improvement is from scraping and stealing better quality IP to train on. And they can just ride Moore's law until they profit then lobby governments to require licenses for fast GPUs because of national security.
Given how hard it is to understand AWS billing, especially if you have custom pricing, I doubt his numbers are anywhere near correct.
That said, I hope they're using their prime Visa card so they can get some cash back on that spend.
“Spend” requires a grain of salt here. AWS “invests” in Anthropic and then Anthropic buys AWS. If you follow the money with marked bills, AWS is buying this compute from itself and then claiming revenue.
There’s a lot of that sort of thing going on at the moment in the AI bubble.
Is that just for inference or is that the cost of training the models as well?
There is no breakdown. Assuming they're real, they're just spending numbers.
Also spend will drop dramatically if the models level out a bit more. The training is what's compute heavy, and if you aren't having to retrain every month, but able to use things like Skills to stay competitive your costs will drop a lot.
I suppose that's the pessimistic-on-AI side. On the other hand, once you create God little things like money are meaningless.
I've been unable to find it but early in the railroad boom did rail companies undercharge for freight in order to build demand and interest? Given that they were also given land in exchange fro building, it would have been in their interests to do so. They could also amortize rail costs over a long time as rails are useful for longer than GPUs - but I am also curious if there was an unsustainably low entry price for railroads like Ed is implying for AI tools.
Ed Zitron's style isn't for everybody, I understand that. But if these numbers, and the direction they're going in, are correct... to me this points to a significant AI bubble deflation coming soon. It just isn't sustainable, it seems.
if you are worried about AWS spend, i have news for you about non AWS spend.
"coming soon" is also really over simplistic. you would have missed some of the greatest tech companies in the past 20 years if you evaluated startups based on their early-year revenue vs infra spend
like sure i have a dog in this fight but i actually want the criticism to sharpen my thinking, unfortunately yours does not meet that bar.
are you glossing over the whole "aws bill isnt structural infra spend" thing?!
they spend 104% of revenue on ONE cloud provider and costs scale linearly with revenue growth. assume zitron didn't pull these numbers out of his ass.
educate me how this isnt selling $20 bills for $5. you're a smart dude; i myself aint seeing the "sustainable business practices" here
the entire point is that stressing sustainable business practices here at this point in a startup's life is extremely short sighted and is the kind of shitty analysis that gives HN a bad rep.
pull up Uber's financials leading up to IPO. unsustainable and everyone knew it. they work it out after because they burned money and eventually achieved a sustainable moat. this is why venture exists. HN doesnt like venture which is, well, ironic given the domain we're on.
a better negative argument i'd rather see looks like this - "ive run these aws numbers against the typical spend path of pre IPO startups who then later improved their cost baseline and margin profile and even after accounting for all that, Anthropic ngmi". thats the kind of minimum sophistication you need here to play armchair financial analyst. ed zitron, and everyone involved in this entire thread, incl myself, have not done that because we are lazy and ignorant and dont actually care enough about seeking the truth here. we are as unprepared to analyze this AWS spend as we are to understand their 1b -> 10b revenue ramp in 2025. you havent done the work and yet you sit here and judge it unsustainable based off some shitty "leaks". dont pretend that ed's analysis is at all meaningful particularly because he conveniently stops where it supports his known negative bias.
Afaik Uber numbers were significantly smaller.
Ai can both be a bubble and also the greatest economic value add of this generation at the same time. It doesn’t have to be either or.
All bubbles (dot com, housing, tech, crypto, etc) have a lot of losers and a few big winners.
That is less a reflection on the market of the bubble and more a reflection of the number, skill and risk taking of the prospectors.
It would be hilarious if all this was a devilish scheme to burn the excess money that was printed during COVID. Unfortunately, what is more likely is another bailout if the bubble bursts.
Premium newsletter? Sigh, I’m getting old…
Just a reminder, Ed Zitron is neither an AI researcher, nor an Engineer, nor a Financial Analyst, nor an Economist nor an Insider and has ZERO clue in multiple dimensions (technology, investing, unit economics, growth, TAM) to analyze any of this
>> nor a Financial Analyst, nor an Economist
Those people arent exactly experts or right most of the time either
If Ed is neutral or an expert, he would have had the following analysis
"What is the unit cost of serving a Token? It is the cost of electricity + amortized cost of GPU (GPUs would have been Capex, but because of their fast depreciation rate, you can claim they should be Opex). Given this cost structure, every SOTA labs (Google, Anthropic and OpenAI) are profitable and actually have high unit margins of 50-60%."
High Unit Margins and growth means, these labs can be profitable anytime they choose to
An 18 hour old account, named suspiciously like an ai model company, trying to discredit a prominent AI skeptic.
Here I am, 2.5 year account discrediting him - the confirmation bias of Zitron is so thick, you need diamond tipped cutters to cut through it.
I'm betting big on AI (not a shill if you are alluding to that). Address my points. All other things are irrelevant.
Correct. And yet, he's provided some of the most level headed takes on the current LLM boom, to the point where FT Alphaville link to his analysis of the economics.
The jury is still out there if his analysis are level-headed or not. He says things that *some* people want to hear, but that's not level-headed
E.g
What is the unit cost of serving a Token? It is the cost of electricity + amortized cost of GPU (GPUs would have been Capex, but because of their fast depreciation rate, you can claim they should be Opex). Given this cost structure, every SOTA labs (Google, Anthropic and OpenAI) are profitable and actually have high-margins 50-60%.
With this margin and growth, the frontier labs can be profitable anytime they want to. But they are sacrificing profitability for growth (as they should be)
Where is Ed's analysis about this? Either he is disingenuous or clueless. Remember people who voluntarily subscribe to Ed, are coming from wanting to hear what they believe.
If he is level-headed, show me an Ed article that is positive about AI
> If he is level-headed, show me an Ed article that is positive about AI
Why should those two things go together?
If 1 Billion people voluntarily use a product and many claim to be productive, there must be something good about the product right?
But I guess Ed Zitron has found his audience
His argument is not that AI is not useful, it's that it's not financially sustainable at the current prices being charged for it - and additionally AI start-ups have 0 moat. Both of which likely are true.
FT Alphaville dissed bitcoin and Tesla for 10 years. Not the greatest track record.
As the saying goes, "the market can remain irrational longer than you can remain solvent." I hope the next fad is to invest in boring companies that employ people and give steady returns, and they can stop replacing all the stores in my city with Dollar Trees.