I want to clarify that I never claimed that OpenAI is a profitable business, or even has a good or sustainable business model, but simply that this data falsifies a big claim that Zitron has repeatedly made. He seems to have moved the goal posts to claiming that training costs should count toward the gross margins, but that is a totally different argument.
If you make money from your customers on average, you can grow your way to recouping even very high fixed costs. If you instead lose money from the typical customer, as Zitron repeatedly contended, then growing more will make you run out of money faster. This is an incredibly important distinction, and getting it wrong for so long (in the face of reporting from the Information, which Zitron favorably cites elsewhere, as well as independent estimates) is frankly disqualifying for a commentator on AI industry economics.
I’m really not. I’m interpreting the line items in the most straightforward way possible. If you have evidence that the cost of revenue isn’t covering the inference costs, please let me know.
Moreover, this lines up with reporting in the information and independent estimates from epoch ai, both of which I cite.
Investors are able to see OpenAIs books with more detail than this, and they would not be investing if the unit economics were negative.
Finally, if they were losing money per token, why are the models so damn verbose?
Both of the Chinese labs which filed to go public obfuscate their inference costs by putting some of it in the marketing category (since they heavily subsidize their compute). It is likely OpenAi is doing the same types of financial gymnastics.
"To be fair to Cursor, Anthropic also doesn’t count inference costs for non-paying users as COGS, though it breaks out the costs separately so investors can include them in their COGS calculations if they want. The inference costs are a separate line item in the operating expenses part of the income statement.
OpenAI, however, includes inference costs for non-paying users in its COGS by default. Both companies had around 40%gross margins in 2025."
So your argument is that OpenAI spent more on standard marketing in 2025 than the CocaCola Corporation? I find that difficult to believe. It seems much more likely they’re including subsidies and free tokens in this cost
As someone who's been in startups for more than a decade this tale is hilarious. LTV/CAC is everything... And you are just completely disregarding sales and marketing. I know a lot of companies that were amazing businesses if you just looked at LTV compared to AWS payments. Another question is where GPU depreciation is hiding.
It’s also been discussed rather thoroughly that OpenAI is likely putting free tokens given to companies during their roll out under marketing and the training costs - which functionally are a core business expense - under R&D. Which completely skews the numbers versus what would be traditionally reported.
With how many SPV’s, PIK’s, and other financial chicanery going on right now I would hope articles like this would be more critical and honest about what they are writing but alas here we are.
Newer nvidia GPUs generations keep coming out, and the efficiency gains of them makes the previous generations not worthwhile to run. Another problem is they are run quite hot, and they don't live forever under those thermal loads.
As Zitron has pointed out many times, R&D costs are part of the ongoing costs of delivering frontier LLMs. They're broken out as R&D but they're really COGS, unless the frontier LLM play is just going to suddenly go away, in which case OpenAI and Anthropic do not have moats (not to mention the vast data center buildout that this entire bubble is built on would be mostly unnecessary).
The inference profitability is certainly only the case for API usage paying by the token. We know that subscriptions have been incredibly subsidized, and that is likely being written down under sales and marketing which is why that number is astronomical. The question isn’t whether or not inference is profitable at token rates, the question is whether or not the industry is willing to fully shift to per token spending. Signs have pointed to that being too high, or needing to cut back dramatically on token use.
There is also the question of whether the entities who are buying that API usage are, themselves, making a profit on what they are doing with it. A portion of that is coming from AI startups that are still on an investor runway, and selling to THEIR customers at a loss. So even if OpenAI is selling inference at a profit to them, that doesn't mean the whole system is sustainable all the way down. Unsustainable subsidy can enter a market at more than one level.
Now that's two comments which demolish the article :)
I remember reading (from Zitron actually) that Anthropic makes 85% of its revenue through its API. For OpenAI the % must be lower but still substantial.
Whenever someone uses Cursor, Perplexity, etc that generates revenue at Anthropic and OpenAI. But the companies making these API calls are themselves making massive losses, so this doesn't look very sustainable.
Surprised that you don't mention the caveats associated with Anthropic's profitability report when you cite it. There is probably more hallucination in Substack posts about AI than AI output itself.
This is what I wrote: Indeed, these companies are famously unprofitable (well, before Anthropic’s unprecedented revenue tear in recent months)...
I think it's fair to say that Anthropic being on track to being profitable for any quarter this early, caveats included, is counter to the "famously unprofitable" description (which is otherwise a fair description of the genAI industry to date).
Depends on how you judge the caveat. I very much appreciate your overarching point about not relying on a bust to solve the societal problems the tech companies poses, but I still think this passage is not objectively informative.
It is similar but also opposite; one thing which really strikes me is that 25 years ago the debate with respect to Amazon was "GAAP losses don't matter because it is generating so much cash" and here we are with Anthropic saying "cash burn doesn't matter because it has achieved GAAP profitability". Doesn't necessarily mean anything definitively one way or the other but it is a funny way for history to rhyme
Yeah and Amazon and a bunch of other tech stocks got destroyed by the 2000 bubble. Doesn't mean a lot of them didn't become hugely successful companies but market valuations and company profitability are two separate things.
The margin numbers answer the wrong question. R&D here doesn’t behave like infrastructure—it’s ripening fruit, a few months from worthless once a competitor (or OpenAI itself) ships the next one. Fold that into cost of goods rather than R&D and the picture becomes deeply negative. The trend is real: gross margin and R&D-to-revenue are both moving the right way. But improving isn’t closing—not soon, especially with token prices still falling underneath it. The interesting metric is economic half-life of model advantage.
Your reliance on Pangram is beside the question. If R&D is a decaying asset rather than a stock, the margin numbers fold into cost of goods and the picture changes. Does the reframing hold up, or not?
If unit economics are positive (as is overwhelmingly clear by now), then revenue has to grow faster than fixed costs (incl R&D), which is what OpenAI's figures show (and Anthropic's to a much greater degree), and you'll eventually reach profitability. These companies will likely have high R&D costs for a long time, and maybe in perpetuity, but if revenue is high enough, they'll be profitable.
That's not guaranteed to happen ofc, but it's what the cos are projecting and raising money off of.
You’ve restated ‘more revenue than costs = profit’ without showing why revenue should outrun a cost you’ve just admitted never goes away.
Start with the half-life: how fast does a model’s competitive advantage decay? You can’t answer the revenue question without it. Pick a number, then tell me what growth rate beats it.
A lot depends on that "sales and marketing expense". It seems to have made the difference between positive and negative in 2024, made the difference between positive and negative in 2025 and grown at roughly the same rate as revenue. There are plenty of businesses which get into bad trouble because they look at their big gross margin and say that they can grow into profitability, ignoring the cost of getting the customers through the door.
I'd like to know a lot more about what actually goes into that line - are there sales commissions, how do they account for compute swaps and discounts etc? Basically what is the story whereby that line stops growing but revenue continues to rise.
One, don't be that hard on Ed. He's awesome. It isn't that he's wrong, it's just impossible to predict the future. He just points out that the path to success is not rosy and certain. I call it a parlay bet that needs many things to happen correctly for the money numbers to make sense. But to say he's wrong in thinking that they don't make sense is unfair. They do not make sense in any traditional framework. I think like a simple mind....anything that requires this much money to create better not subsidize the product forever. When $200 buys you $4,000 worth of tokens, 60% margin ain't enough. If it doesn't, you will see the customers go elsewhere, since switching costs are low. You can't pick winners and losers, and you can't forecast when the reckoning will happen. But not everyone will win. In that sense, I defend Ed.
He's not awesome, he's dogmatic in his views to the point where it erodes his credibility. There is a legitimate critique of the value of frontier firms in that model capabilities quickly commoditize. But that's different from Zitron, who just believes this whole AI thing is a worthless racket. That's where he loses me because I already get significant value from the technology
“The entire industry is looking past the fact that the nearly $3 trillion in near-term spending commitments needed completely overwhelm the cash flows of the biggest hyperscalers, so much so that they rely on a cash-hemorrhaging startup like OpenAI as the largest, most critical link in the circular funding chain.
One wonders where investors get that confidence.” — Michael Burry
WTH? "His data shows definitively that OpenAI makes far more from ChatGPT than it costs to serve." It literally doesn't. You have a serious math problem if you think $20B on losses is a growth indicator. You yourself wrote a book criticizing the AI industry oligarchs, so why this hit piece on a fellow critic who's on the same team? I don't get it.
The piece is trying to make a basic point that the unit economics of serving AI models are positive, which has profound implications for the fundamental economics of the whole industry.
The upshot of Ed's body of work is that the bubble will burst, which deflates energy to organize and resist the industry.
Well it comes across as petty friendly fire. Marcus also fights with Ed. I think in-fighting with other critics harms the cause much more than predictions about economic bubble bursts. I still don't understand your math. OAI are making more money, but they're spending more money in order to do so. Tokenomics is doomed.
We really need to understand what's true to organize effectively, and Zitron consistently misrepresents reality to fit his his narrative. This example is just one of many.
The tokenomics are positive, i.e. it costs OpenAI less to serve an API call than the price it charges. Given this, if OAI can grow revenue faster than operating costs (which is true of 2024-2025), then enough growth would get to profitability.
But even if it never gets there and there's a bursting of a bubble, this technology will not go away.
Full disclosure, I love Ed Zitron. When he really gets his rage on it is hilarious (+ also CATHARTIC). (Seriously, him going off with David Gerrard about OpenClaw/Moltbook is probably one of my favourite things on the internet.)
But I've been listening to Ed say the bubble is about to pop for a long time, and I wonder at what point his thesis veers into the unfalsifiable. Like... this bubble can’t pop fast enough, and we're still waiting.
(As an aside, tbh I worry that the literal trillions of dollars Silicon Valley, Wall Street, + other connected elites have poured into AI means the bubble won't be allowed to pop. We're already seeing rule changes & market manipulations to juice this stuff. (Note Ed doesn't think this is anything to worry about either.)
Worth noting tho, I saw Ed on "Breaking Points" the other day. He pointed out that these massive training costs & data-centre buildout costs are not going away. They're not one-and-done things. So, even if he's wrong on the inference stuff, a strong bear case can still be made.
The main issue of your argument is that you give OpenAI a HUGE benefit of the doubt by assuming that the Cogs are truly reflective of all revenue related costs and conclude that overall the gross margin is positive and improving.
If you are familiar of accounting you should be aware that from a cost nature perspective, it is highly likely that the cogs are probably only inference costs of running models and the main bulk of costs are reassigned to R&D (training/maintenance etc). This is not just Open AI, Mag 7 companies also do that.
For the arguments sake, let's just say that Cogs are truly all cogs and they have a positive margin. It appears to me in order to have a thriving genAi business relying on cutting edge models like frontier labs do at the moment, they pretty much have to keep up these R&D costs to innovate and maintain models in order to compete. Is management truly believing "our AGI will be achieved and turned into a monopoly product with eternal moat, so no R&D expense anymore after this point"? Unless they actually achieve AGI or Jensen starts selling GPUs for peanuts, none of this makes sense. Do you really want to debate the odds of these scenarios happening?
I want to clarify that I never claimed that OpenAI is a profitable business, or even has a good or sustainable business model, but simply that this data falsifies a big claim that Zitron has repeatedly made. He seems to have moved the goal posts to claiming that training costs should count toward the gross margins, but that is a totally different argument.
If you make money from your customers on average, you can grow your way to recouping even very high fixed costs. If you instead lose money from the typical customer, as Zitron repeatedly contended, then growing more will make you run out of money faster. This is an incredibly important distinction, and getting it wrong for so long (in the face of reporting from the Information, which Zitron favorably cites elsewhere, as well as independent estimates) is frankly disqualifying for a commentator on AI industry economics.
You haven't actually proven that inference is profitable, you are making very big assumptions about what the line items reported mean.
I’m really not. I’m interpreting the line items in the most straightforward way possible. If you have evidence that the cost of revenue isn’t covering the inference costs, please let me know.
Moreover, this lines up with reporting in the information and independent estimates from epoch ai, both of which I cite.
Investors are able to see OpenAIs books with more detail than this, and they would not be investing if the unit economics were negative.
Finally, if they were losing money per token, why are the models so damn verbose?
Both of the Chinese labs which filed to go public obfuscate their inference costs by putting some of it in the marketing category (since they heavily subsidize their compute). It is likely OpenAi is doing the same types of financial gymnastics.
Seems like mostly no:
The Information reported this in April:
"To be fair to Cursor, Anthropic also doesn’t count inference costs for non-paying users as COGS, though it breaks out the costs separately so investors can include them in their COGS calculations if they want. The inference costs are a separate line item in the operating expenses part of the income statement.
OpenAI, however, includes inference costs for non-paying users in its COGS by default. Both companies had around 40%gross margins in 2025."
https://www.theinformation.com/newsletters/ai-agenda/nuances-cursors-gross-margin-comparing-gpt-5-5-claude-mythos
So your argument is that OpenAI spent more on standard marketing in 2025 than the CocaCola Corporation? I find that difficult to believe. It seems much more likely they’re including subsidies and free tokens in this cost
As someone who's been in startups for more than a decade this tale is hilarious. LTV/CAC is everything... And you are just completely disregarding sales and marketing. I know a lot of companies that were amazing businesses if you just looked at LTV compared to AWS payments. Another question is where GPU depreciation is hiding.
It’s also been discussed rather thoroughly that OpenAI is likely putting free tokens given to companies during their roll out under marketing and the training costs - which functionally are a core business expense - under R&D. Which completely skews the numbers versus what would be traditionally reported.
With how many SPV’s, PIK’s, and other financial chicanery going on right now I would hope articles like this would be more critical and honest about what they are writing but alas here we are.
GPUs have appreciated in recent years no? As in you can charge more today to rent an A100 or H100 than two years ago. So might not be that relevant.
Newer nvidia GPUs generations keep coming out, and the efficiency gains of them makes the previous generations not worthwhile to run. Another problem is they are run quite hot, and they don't live forever under those thermal loads.
Previous commenter was right. GPUs are renting for more now than they were a year or 2 ago:
https://newsletter.semianalysis.com/p/the-great-gpu-shortage-rental-capacity
Come on, that's a blip.
As Zitron has pointed out many times, R&D costs are part of the ongoing costs of delivering frontier LLMs. They're broken out as R&D but they're really COGS, unless the frontier LLM play is just going to suddenly go away, in which case OpenAI and Anthropic do not have moats (not to mention the vast data center buildout that this entire bubble is built on would be mostly unnecessary).
Revenue to (COGS + R&D) ratio is also improving — if you can trust revenue numbers.
What about marketing, which is clearly obfuscating some of the COGS?
Even if you throw in *all* cost of operations you still have improving EBITDA ratio.
Ah yes well I'm sure they definitely have enough runway and liquid capital to reach equilibrium
The inference profitability is certainly only the case for API usage paying by the token. We know that subscriptions have been incredibly subsidized, and that is likely being written down under sales and marketing which is why that number is astronomical. The question isn’t whether or not inference is profitable at token rates, the question is whether or not the industry is willing to fully shift to per token spending. Signs have pointed to that being too high, or needing to cut back dramatically on token use.
There is also the question of whether the entities who are buying that API usage are, themselves, making a profit on what they are doing with it. A portion of that is coming from AI startups that are still on an investor runway, and selling to THEIR customers at a loss. So even if OpenAI is selling inference at a profit to them, that doesn't mean the whole system is sustainable all the way down. Unsustainable subsidy can enter a market at more than one level.
Now that's two comments which demolish the article :)
I remember reading (from Zitron actually) that Anthropic makes 85% of its revenue through its API. For OpenAI the % must be lower but still substantial.
Whenever someone uses Cursor, Perplexity, etc that generates revenue at Anthropic and OpenAI. But the companies making these API calls are themselves making massive losses, so this doesn't look very sustainable.
And the neoclouds who provide the compute run at a loss so actual compute costs are much higher the OpenAI currently experiences.
Surprised that you don't mention the caveats associated with Anthropic's profitability report when you cite it. There is probably more hallucination in Substack posts about AI than AI output itself.
This is what I wrote: Indeed, these companies are famously unprofitable (well, before Anthropic’s unprecedented revenue tear in recent months)...
I think it's fair to say that Anthropic being on track to being profitable for any quarter this early, caveats included, is counter to the "famously unprofitable" description (which is otherwise a fair description of the genAI industry to date).
Depends on how you judge the caveat. I very much appreciate your overarching point about not relying on a bust to solve the societal problems the tech companies poses, but I still think this passage is not objectively informative.
This is similar to debates I heard around 1998 and 1999 about whether Amazon was about to collapse from its losses.
It is similar but also opposite; one thing which really strikes me is that 25 years ago the debate with respect to Amazon was "GAAP losses don't matter because it is generating so much cash" and here we are with Anthropic saying "cash burn doesn't matter because it has achieved GAAP profitability". Doesn't necessarily mean anything definitively one way or the other but it is a funny way for history to rhyme
Yeah and Amazon and a bunch of other tech stocks got destroyed by the 2000 bubble. Doesn't mean a lot of them didn't become hugely successful companies but market valuations and company profitability are two separate things.
Great point, Amazon stock only lost 90% of its value shortly after that…and actually had a legitimate business.
The margin numbers answer the wrong question. R&D here doesn’t behave like infrastructure—it’s ripening fruit, a few months from worthless once a competitor (or OpenAI itself) ships the next one. Fold that into cost of goods rather than R&D and the picture becomes deeply negative. The trend is real: gross margin and R&D-to-revenue are both moving the right way. But improving isn’t closing—not soon, especially with token prices still falling underneath it. The interesting metric is economic half-life of model advantage.
This comment is 100% ai generated per Pangram:
https://www.pangram.com/history/8773bb25-c217-4783-b0c9-bd105b8208ee?ucc=9L9QmJmITQJ
Your reliance on Pangram is beside the question. If R&D is a decaying asset rather than a stock, the margin numbers fold into cost of goods and the picture changes. Does the reframing hold up, or not?
If unit economics are positive (as is overwhelmingly clear by now), then revenue has to grow faster than fixed costs (incl R&D), which is what OpenAI's figures show (and Anthropic's to a much greater degree), and you'll eventually reach profitability. These companies will likely have high R&D costs for a long time, and maybe in perpetuity, but if revenue is high enough, they'll be profitable.
That's not guaranteed to happen ofc, but it's what the cos are projecting and raising money off of.
You’ve restated ‘more revenue than costs = profit’ without showing why revenue should outrun a cost you’ve just admitted never goes away.
Start with the half-life: how fast does a model’s competitive advantage decay? You can’t answer the revenue question without it. Pick a number, then tell me what growth rate beats it.
A lot depends on that "sales and marketing expense". It seems to have made the difference between positive and negative in 2024, made the difference between positive and negative in 2025 and grown at roughly the same rate as revenue. There are plenty of businesses which get into bad trouble because they look at their big gross margin and say that they can grow into profitability, ignoring the cost of getting the customers through the door.
I'd like to know a lot more about what actually goes into that line - are there sales commissions, how do they account for compute swaps and discounts etc? Basically what is the story whereby that line stops growing but revenue continues to rise.
One, don't be that hard on Ed. He's awesome. It isn't that he's wrong, it's just impossible to predict the future. He just points out that the path to success is not rosy and certain. I call it a parlay bet that needs many things to happen correctly for the money numbers to make sense. But to say he's wrong in thinking that they don't make sense is unfair. They do not make sense in any traditional framework. I think like a simple mind....anything that requires this much money to create better not subsidize the product forever. When $200 buys you $4,000 worth of tokens, 60% margin ain't enough. If it doesn't, you will see the customers go elsewhere, since switching costs are low. You can't pick winners and losers, and you can't forecast when the reckoning will happen. But not everyone will win. In that sense, I defend Ed.
He's not awesome, he's dogmatic in his views to the point where it erodes his credibility. There is a legitimate critique of the value of frontier firms in that model capabilities quickly commoditize. But that's different from Zitron, who just believes this whole AI thing is a worthless racket. That's where he loses me because I already get significant value from the technology
any company can be considered profitable if you decide that some of their operating costs don't count lmao
“The entire industry is looking past the fact that the nearly $3 trillion in near-term spending commitments needed completely overwhelm the cash flows of the biggest hyperscalers, so much so that they rely on a cash-hemorrhaging startup like OpenAI as the largest, most critical link in the circular funding chain.
One wonders where investors get that confidence.” — Michael Burry
WTH? "His data shows definitively that OpenAI makes far more from ChatGPT than it costs to serve." It literally doesn't. You have a serious math problem if you think $20B on losses is a growth indicator. You yourself wrote a book criticizing the AI industry oligarchs, so why this hit piece on a fellow critic who's on the same team? I don't get it.
The piece is trying to make a basic point that the unit economics of serving AI models are positive, which has profound implications for the fundamental economics of the whole industry.
The upshot of Ed's body of work is that the bubble will burst, which deflates energy to organize and resist the industry.
Well it comes across as petty friendly fire. Marcus also fights with Ed. I think in-fighting with other critics harms the cause much more than predictions about economic bubble bursts. I still don't understand your math. OAI are making more money, but they're spending more money in order to do so. Tokenomics is doomed.
We really need to understand what's true to organize effectively, and Zitron consistently misrepresents reality to fit his his narrative. This example is just one of many.
The tokenomics are positive, i.e. it costs OpenAI less to serve an API call than the price it charges. Given this, if OAI can grow revenue faster than operating costs (which is true of 2024-2025), then enough growth would get to profitability.
But even if it never gets there and there's a bursting of a bubble, this technology will not go away.
Full disclosure, I love Ed Zitron. When he really gets his rage on it is hilarious (+ also CATHARTIC). (Seriously, him going off with David Gerrard about OpenClaw/Moltbook is probably one of my favourite things on the internet.)
But I've been listening to Ed say the bubble is about to pop for a long time, and I wonder at what point his thesis veers into the unfalsifiable. Like... this bubble can’t pop fast enough, and we're still waiting.
(As an aside, tbh I worry that the literal trillions of dollars Silicon Valley, Wall Street, + other connected elites have poured into AI means the bubble won't be allowed to pop. We're already seeing rule changes & market manipulations to juice this stuff. (Note Ed doesn't think this is anything to worry about either.)
Worth noting tho, I saw Ed on "Breaking Points" the other day. He pointed out that these massive training costs & data-centre buildout costs are not going away. They're not one-and-done things. So, even if he's wrong on the inference stuff, a strong bear case can still be made.
~Peace & love.
Good article. I appreciate Zitron’s work, but don’t take anything he says as gospel. Your analysis here is useful!
Re: making people obsolete, did you see my lengthy article on this? I’m sure you’re aware of most of what I discuss, but still might be of interest, if you haven’t seen it: https://www.realtimetechpocalypse.com/p/a-pro-human-manifesto-part-1
The main issue of your argument is that you give OpenAI a HUGE benefit of the doubt by assuming that the Cogs are truly reflective of all revenue related costs and conclude that overall the gross margin is positive and improving.
If you are familiar of accounting you should be aware that from a cost nature perspective, it is highly likely that the cogs are probably only inference costs of running models and the main bulk of costs are reassigned to R&D (training/maintenance etc). This is not just Open AI, Mag 7 companies also do that.
For the arguments sake, let's just say that Cogs are truly all cogs and they have a positive margin. It appears to me in order to have a thriving genAi business relying on cutting edge models like frontier labs do at the moment, they pretty much have to keep up these R&D costs to innovate and maintain models in order to compete. Is management truly believing "our AGI will be achieved and turned into a monopoly product with eternal moat, so no R&D expense anymore after this point"? Unless they actually achieve AGI or Jensen starts selling GPUs for peanuts, none of this makes sense. Do you really want to debate the odds of these scenarios happening?
This was one hell of a spin job. OpenAI might have found their new PR guy!
Read literally anything I've written about OpenAI lol
e.g. https://www.nytimes.com/2024/09/29/opinion/ai-risks-safety-whistleblower.html?unlocked_article_code=1.OU4.-Lcq.-p2uHNAe66sn&smid=url-share