OpenAI and Anthropic's money: who really makes money from AI

In November 2026 we learned how much the two leading companies in artificial intelligence earned last quarter. OpenAI, the maker of ChatGPT, brought in more than ten billion dollars in three months for the first time. Anthropic, the maker of Claude, brought in almost seventeen billion over the same three months. The headlines said Anthropic had overtaken OpenAI.

But behind these numbers there is a more interesting question. Do these companies actually make money? Not revenue, but money. And if they don't, who is paying for it, and what happens next?

This question is not just for investors. If you build a product on ChatGPT or Claude, the prices, limits and rules you work under are shaped by exactly this economy. When a company is short of money, it changes its pricing. When it needs customers, it buys startups or releases a product that competes with yours.

In this episode we will look at where OpenAI and Anthropic get their money, how they find customers, what they spend it on, what they say about themselves and what they keep quiet. And at the end we will talk about what all of this means for people who build on their models.

One important caveat first. These two companies are open to very different degrees. Anthropic is preparing to go public, and on the tenth of November it filed a prospectus with the US Securities and Exchange Commission, an official document with its financial statements. The company is legally responsible for the numbers in it. OpenAI remains private and does not publish its accounts. Everything we know about its finances comes from leaks: investor documents seen by reporters at The Information, Bloomberg and The Wall Street Journal. No one has checked these numbers. So as we go, I will say where each figure comes from.

Part one. Why seventeen versus ten is not a fair comparison.

Let's start with the numbers themselves. Anthropic's revenue for the third quarter, July through September, was sixteen billion, nine hundred fifty-six million dollars. That is in the prospectus. OpenAI's, according to The Information and Bloomberg, was about ten billion, fifty percent more than in the second quarter.

But these two figures are calculated differently, and that matters. Both companies sell access to their models not only directly but also through cloud platforms: Amazon, Google and Microsoft. The customer pays the cloud, the cloud keeps its share and passes the rest to the model maker.

Anthropic books the full amount the customer paid as revenue and records the cloud's commission as an expense. Forty-four percent of its revenue for the first nine months of this year came through the clouds. OpenAI books only its own share, with Microsoft's part already taken out. Both methods are allowed under accounting rules. But you cannot compare them directly.

OpenAI knows this and is annoyed. In April, an internal memo from OpenAI's chief revenue officer, Denise Dresser, leaked. She told staff: "Their stated run rate is inflated," and "We report Microsoft revshare net." By her estimate, the gap was about eight billion dollars on an annualized basis. Anthropic replies that it sells to customers itself and has the right to count the full amount.

There is a second trap: quarterly revenue versus the so-called run rate. A run rate is not money earned but a projection: a company takes its revenue for the latest month and scales it up to a full year. Anthropic's stood at eighty-one billion at the end of October. For OpenAI, at the end of September, Axios first reported almost seventy billion, and nine days later the Financial Times, Bloomberg and Axios itself corrected that to about fifty. The twenty billion dollar gap came precisely from how cloud sales are counted.

Even if you strip the cloud commissions out of Anthropic's revenue, it is still larger in the third quarter. But not by as much as the headlines suggest. Remember this rule: when two companies are compared by revenue, the first question to ask is how each of them counts it.

Part two. Where OpenAI's money comes from.

OpenAI's growth story looks like this. According to leaked audited statements, the company had three point seven billion dollars in revenue in 2024 and thirteen billion in 2025. In the first quarter of this year it had about five point seven billion, in the second six point seven, in the third ten.

The most interesting thing here is that the first half of the year was hard for OpenAI. Co-founder Greg Brockman said so plainly at a Wall Street Journal conference on the fourth of November: "Q1 was pretty rough. Q2 was pretty rough." In April, reporters wrote that the company had missed its internal targets for new users and revenue. Growth in the second quarter was only eighteen percent, and Anthropic pulled ahead of OpenAI for the first time.

What changed in the third quarter? According to Brockman, business customers and coding took off. He said a line that sums up the whole year well: "We had Codex, but no one was talking about it." Codex is OpenAI's coding tool, a rival to Claude Code. According to the press, enterprise revenue doubled in the quarter, and businesses now bring OpenAI more than half of its money. At the start of the year it was the other way around: sixty percent came from ordinary users, forty from businesses.

Now the users. Officially, at the end of February ChatGPT had nine hundred million weekly users and fifty million paying subscribers. In the autumn OpenAI announced one point two billion weekly users. How many of them pay now, the company does not say. The press mentions a figure of more than one hundred million, but there is no confirmation for it.

And this is where the part OpenAI prefers not to discuss begins. Most of the growth in paying users comes from the cheapest plan, ChatGPT Go, at eight dollars a month. According to an internal forecast reported by The Information, this plan should make up ninety-two percent of all paid subscriptions by the end of the year. And the number of subscribers on the classic twenty dollar Plus plan is expected to fall by eighty percent. In other words, people are not moving up to expensive plans; they are moving down to the cheap one in large numbers. Average revenue per subscriber is falling.

How does OpenAI plan to make up for this? With advertising. In 2024 Sam Altman said: "I kind of think of ads as a last resort for us for a business model." In January 2026 the company announced that ChatGPT would carry ads after all, and on the ninth of February it launched them in the United States. Ads appear on the free plan and on the Go plan. By the end of August, according to official figures, ad revenue reached an annual run rate of one billion dollars. The target for 2026, according to the press, is two and a half billion, and one hundred billion by 2030.

Not everything OpenAI tried has worked. Built-in shopping right inside the chat, launched in the autumn of 2025, was wound down in March: people searched for products in ChatGPT but preferred to buy them on store websites. The Sora video service was shut down. By various press estimates, it cost between one and fifteen million dollars a day while bringing in very little revenue.

Part three. How OpenAI finds customers.

This is the question with the fewest public answers from OpenAI. How much the company spends on marketing, what it costs to acquire one customer, how many people work in sales, how many subscribers leave: none of this is published.

Here is what can be seen from the outside. The main growth channel is free ChatGPT and word of mouth. People try it for free, and some of them start paying. What share, the company does not disclose. If you take the official February numbers, fifty million paying out of nine hundred million, you get about five and a half percent. That is my own calculation, not a company figure.

The second channel is big, loud advertising. In February, for the second year in a row, OpenAI bought a minute during the Super Bowl, the final of American football and the most expensive advertising time in the world. The price was not disclosed.

The third channel is partners. The best-known partner is Apple: ChatGPT is built into Siri. But according to Bloomberg, Apple pays nothing for this, and the benefit for OpenAI was supposed to come through new subscribers. In the spring there were reports that far fewer subscriptions were coming through Siri than expected, and that OpenAI had even considered a claim against Apple. For business customers, OpenAI started selling through Amazon for the first time this year, after its exclusive arrangement with Microsoft ended.

So the customer acquisition picture looks like this: a huge funnel of free users, a small share of paying ones, growing advertising and enterprise sales. And the cost of this funnel is hidden. A free user does not pay, but every question they ask costs the company money. According to The Information, OpenAI's gross margin fell from forty to thirty-three percent in 2025, and free users were one of the reasons. How much a single free user costs, OpenAI does not say.

Part four. Where OpenAI's money goes.

Now the key number: an operating margin of minus one hundred nine percent. What does that mean in plain terms? For every dollar of revenue in the third quarter, OpenAI spent about two dollars and nine cents. The operating loss for the quarter, according to The Wall Street Journal, was about ten point nine billion dollars. In the second quarter the margin was minus one hundred fifty percent, that is, two and a half dollars of costs for every dollar of revenue.

The improvement looks impressive, but look at where it came from. Revenue grew by fifty percent, and costs by twenty-four. The company did not start spending less. It spent more; revenue simply grew faster. The loss in dollars keeps growing.

One more detail. This is an adjusted margin. The word "adjusted" means some costs have been left out. Judging by the second quarter numbers, what was left out is stock compensation for employees, which runs to several billion dollars a quarter. Exactly what is excluded, OpenAI does not disclose. With those payments included, the loss would be bigger.

What does the money go to? First of all, to computing: training new models and serving user requests. In December 2025 Altman put it this way: "If we weren't continuing to grow our training costs by so much, we would be profitable way, way earlier." And he added: "We have always been in a compute deficit."

There is a telling figure in Microsoft's official financial statements. For the quarter that ended in September, Microsoft booked ten billion dollars in revenue from OpenAI. That is roughly all of OpenAI's revenue for the same quarter. Most of this money is for Azure cloud capacity, which OpenAI's models run on. So a large share of what users pay OpenAI flows straight back to its supplier.

Then there are the commitments for the future. OpenAI has signed computing deals with Microsoft, Oracle, Nvidia, AMD, Broadcom, Amazon and CoreWeave. At the end of 2025 Altman put the figure at one point four trillion dollars over eight years. In February 2026, according to CNBC and Reuters, the plan was cut to about six hundred billion through 2030.

According to a forecast reported by The Information, OpenAI expects to keep burning cash: twenty-five billion dollars in 2026, fifty-seven billion in 2027, eighty-five billion in 2028. The first year with positive cash flow in this plan is 2030.

Where does the money for all this come from? From investors. In March OpenAI raised one hundred twenty-two billion dollars at a valuation of eight hundred fifty-two billion. The investors include Amazon, Nvidia and SoftBank. Now, according to Bloomberg, the company is finishing another round of up to thirty billion, at a valuation of about one point two five trillion dollars. Its stock market listing has been pushed back to 2027.

Part five. Where Anthropic's money comes from.

Anthropic's picture is completely different, and, importantly, we see it through an official document.

Look at the growth. In 2023 the company's revenue was forty-one million dollars. In 2024, three hundred eighty-six million. In 2025, four point six billion. And in the first nine months of 2026, thirty-three billion. Quarter by quarter this year: four point seven billion, eleven and a half, almost seventeen.

Dario Amodei, the head of Anthropic, described it this way at a developer conference in the spring: "We tried to plan very well for a world of 10x growth per year … And yet we saw 80x."

What does this revenue consist of? Seventy-nine percent is usage-based: businesses and developers pay for every chunk of text processed. Twenty-one percent is subscriptions: the Pro and Max plans and enterprise seats.

And here is the main difference from OpenAI. Seventy-four percent of Anthropic's revenue comes from large customers who pay more than one hundred thousand dollars a year. There are almost twelve thousand such customers. The remaining twenty-six percent comes, as the prospectus puts it, from roughly forty million smaller customers. Anthropic is first and foremost a company that sells to businesses. OpenAI has historically been a company that sells to people.

Another number Anthropic is proud of: revenue retention of nine hundred percent. This means that large customers who were with the company a year ago now pay it nine times as much. Ordinary subscribers are not included in this figure.

And one number the company is clearly not proud of, but had to disclose. By Anthropic's own estimate, five to seven percent of its revenue for the nine months came from customers who violated its usage policy or contract terms. The company writes that this revenue will not recur. Who these customers are, it does not say.

Part six. How Anthropic finds customers.

Here Anthropic has disclosed more than any company in this industry, and this is the most useful part of the story for anyone building a business.

In the prospectus the company describes its sales model in words worth remembering: "A large portion of our growth to date has been self-service and organic." "We do not rely exclusively on a traditional enterprise sales force." And here is the line that explains everything: "By the time our direct teams engage a customer, Claude is often already in use."

Here is how it works in practice. First, individual employees start using Claude, most often programmers, through Claude Code. The prospectus calls them internal champions. They bring the tool into the company. Then Anthropic's salespeople and engineers come to an already warm customer and sign an enterprise contract. Dario Amodei explained why they started with programmers: "Software engineers are the ones who are fastest to adopt new technology."

The second channel is the cloud platforms. Salespeople at Amazon, Google and Microsoft offer Claude to their own customers, and Anthropic pays them a commission. Forty-four percent of revenue goes through this channel. The size of the commission is not disclosed.

The third channel is integration partners. In March the company set aside one hundred million dollars for its partner program, and in October another hundred million for training engineers.

Now the most interesting part: what this costs. Spending on sales, marketing and partnerships in the third quarter was two billion, three hundred sixty-one million dollars. That is about fourteen percent of revenue. A year earlier, in the third quarter of 2025, it was two hundred two million. So sales spending grew more than tenfold, roughly in line with revenue.

Pure advertising, meanwhile, cost little: one hundred forty-seven million dollars for the whole of 2025. The loudest campaign ran in February, also during the Super Bowl. Its tagline was a direct hit at OpenAI: "Ads are coming to AI. But not to Claude." Altman called the spots "funny" but "clearly dishonest."

The long-term target Anthropic showed investors is to spend between seven and ten percent of revenue on sales and marketing.

It matters what is missing here. The company does not disclose the cost of acquiring a single customer, the number of subscribers, the number of active users or customer churn. The words "monthly active" do not appear in the prospectus at all. So from public data you can see how much Anthropic spends on sales overall, but not what one customer costs.

Part seven. Does Anthropic make money?

The answer depends on which line you look at.

On operating results, Anthropic has almost broken even. In the second quarter it had an operating profit of nine hundred seventy-four million dollars, its first and so far only profitable quarter. In the third it was back to a small loss, five hundred sixty-seven million, or minus three percent of revenue. On the adjusted margin the company calculates itself, it is plus fifteen percent. Compare that with minus one hundred nine for OpenAI.

At the same time, Anthropic's net loss for the third quarter was twenty-three billion dollars. Why such a gap? Almost twenty-two billion of it is a paper expense: the revaluation of convertible loans. When the company's value rises, so does the value of its debts that will turn into shares. These losses did not cost any real cash. And actual operating cash flow for the nine months is positive: plus six point six billion.

The biggest expense is computing. In the third quarter it ate up fifty-eight percent of revenue, compared with forty-eight percent in the second. Why did it get worse? The company explains it openly in the prospectus: it had to buy five billion dollars of computing capacity under short-term contracts, and those cost noticeably more than long-term ones. Demand grew faster than its own capacity could be built.

And this is where the story becomes worrying. Anthropic's total contractual commitments come to about nine hundred billion dollars, almost all of it computing and data centers. One hundred twenty-one billion to Google. One hundred ten billion to Amazon. Up to one hundred twenty billion to SpaceX, which now owns Elon Musk's former xAI. About one hundred sixty billion in leases from Broadcom. Forty-seven billion to Microsoft. More than twenty billion to AMD. Meanwhile, cash and short-term investments on the balance sheet total one hundred thirty-two billion.

Dario Amodei described this risk himself in February on Dwarkesh Patel's podcast. His point was this: if you buy computing on the assumption of a trillion dollars in revenue, then "If my revenue is not $1 trillion, if it's even $800 billion … then you go bankrupt." Anthropic's chief financial officer, Krishna Rao, says: "The compute that we procure is the lifeblood of our business," and he spends thirty to forty percent of his time on it.

Amodei also has an optimistic argument that he has been repeating since 2025: "if you consider each model to be a company, the model that was trained in 2023 was profitable." Roughly speaking, "you train a model that costs $100 million, and then you deploy it in 2024, and it makes $200 million of revenue." But that same year the company is already spending a billion on training the next one. So each model pays for itself, while the company as a whole loses money. The argument sounds convincing, but it rests on one condition: each new model has to bring in much more than the one before. The prospectus confirms this dependence: more than eighty percent of third quarter revenue came from models released in the last six months. Stop, and revenue starts to fall.

Part eight. What we are not told.

This is the most important part, because the gaps in reporting often say more than the numbers.

Here is what OpenAI does not disclose: financial statements under accounting rules, revenue by product, the number of paying subscribers after February and how they split across plans, the share of free users who start paying, churn, customer acquisition cost, marketing spend, the cost of serving one user, its largest customers and the definition of its own adjusted margin.

Here is what Anthropic does not disclose, even in its official prospectus: Claude Code revenue, the number of subscribers and active users, the names of its large customers, the size of the commission it pays the clouds, churn and a date for reaching profitability. Gross margin is not shown in the main tables either: since 2025 the company has classified its costs differently. Analysts work it out from a footnote, and it comes to about eighty-five percent for the nine months. But that is a calculation that depends on how the company splits the same computing capacity between serving customers and research. Anthropic itself writes that it does not split it by task.

Why do these gaps matter? Because that is exactly where negative economics can hide. Take heavy users on subscriptions. In January 2025 Altman himself admitted that OpenAI was losing money on its two hundred dollar Pro plan, because people were using it more than expected. At Anthropic, one user on the two hundred dollar plan calculated that in a month they had used two thousand four hundred dollars' worth of resources at developer prices. Those are prices, not costs; Anthropic's real costs are unknown. But this is exactly why the company has been introducing weekly limits since 2025 and regularly changing them, and why enterprise seats are now billed by actual usage.

Companies that resell the models went through the same problem. By analysts' estimates, the code editor Cursor, on its old plan, was paying about forty-four dollars for models for every twenty dollars of subscription. In June 2025 it switched to usage-based pricing. Since June 2026, GitHub Copilot has also turned its subscription into a wallet that tokens are deducted from.

Analysts disagree. The bank HSBC believes that, even with revenue growth, OpenAI will be two hundred seven billion dollars short by 2030. Jay Goldberg, an analyst at Seaport Research, says of Anthropic: "They will be profitable probably in the next quarter or two. But then they're going to start going back into losses." New York University professor Aswath Damodaran, one of the best-known experts on company valuation, valued Anthropic at one point four seven five trillion dollars. That is the low end of the price range at which the shares are being offered. He said that if the price comes in well above that, "I will sit out the frenzy."

Part nine. Who is really making money.

If OpenAI is losing money and Anthropic is barely breaking even, who in this chain is making money? The people selling them shovels.

For the quarter that ended in July, Nvidia had ninety-six billion dollars in revenue and about sixty billion in net profit. Nvidia's net profit for a single quarter is roughly six times OpenAI's entire quarterly revenue. Google's cloud business doubled in the third quarter, Amazon Web Services grew by thirty-nine percent, and Microsoft's Azure by forty-eight.

And this is where what analysts call circular deals come in. Nvidia has invested about thirty billion dollars in OpenAI, and OpenAI buys Nvidia chips. AMD gave OpenAI the right to buy up to ten percent of its shares at one cent apiece, and OpenAI committed to buying AMD chips. When Anthropic goes public, Google, Nvidia, AMD, Microsoft, Broadcom and other suppliers are buying twenty-six billion dollars of its shares. Mark Mahaney, an analyst at Evercore ISI, summed it up in one sentence: "For the hyperscalers, Anthropic is at once a customer, competitor, sales channel and investment." The Bank for International Settlements in Basel has calculated that more than half of the money invested in AI companies comes from other AI companies.

This is not necessarily a bubble. Analysts at Morningstar think such deals are worth watching but not yet alarming. But it does mean that a huge part of the industry's revenue is, for now, circulating among the same companies.

Part ten. What this means for those who build on these models.

Now the main point for listeners building their own products.

First. The era of ever-falling prices has ended at the top. From 2023 to 2025, the price of the strongest models fell roughly twentyfold. In 2026 the trend reversed: OpenAI's new model, GPT-5.5, costs twice as much as the previous one, and Google's new model, Gemini 4 Argon, launched at an introductory price that reportedly doubles from the first of December. Small models and Chinese open models remain cheap. If your business depends on a strong model getting cheaper tomorrow, that assumption no longer holds.

Second. Unlimited plans are ending. Cursor, GitHub Copilot, enterprise Claude: everywhere, flat subscriptions are turning into usage-based billing, and limits change along the way. If your product sells customers unlimited use while you pay for every request, you are repeating a mistake these companies have already fixed. Calculate the cost of one task for your customer, not the cost of a subscription.

Third. The platform can become your competitor or simply cut you off. In 2025, Anthropic cut off the code editor Windsurf with a few days' notice when OpenAI was set to buy it. Anthropic co-founder Jared Kaplan explained: "I think it would be odd for us to be selling Claude to OpenAI." According to an OpenAI announcement, its models were due to disappear from the Cursor editor on the twelfth of November, after SpaceX bought it. Whether that happened, there is no confirmation at the time of recording. Cursor prepared for this in advance, because OpenAI's models accounted for only about five percent of its requests. And on the tenth of November, OpenAI bought the travel agency Odessia, which ran on OpenAI's own models, and shut the service down on the day of the deal. That was the subject of our previous episode.

Fourth. Models go away. In February OpenAI removed GPT-4o from ChatGPT, and, according to the press, on the fourteenth of January 2027 it will shut it and several other models down in the API, together with all the versions fine-tuned on them. If you fine-tuned a model for your task, it will disappear along with the base model.

Fifth. Successful companies protect their margins. The average gross margin of AI applications, according to the Bessemer fund, is fifty or sixty percent; for classic software it is eighty or ninety. Cursor reached a small gross profit when it built its own models for simple tasks. The common techniques are these: use several providers, send simple tasks to cheap models, and reuse text that has already been processed, which at all three major companies costs roughly ten times less.

What you can do as soon as tomorrow. Work out how much one typical task for your customer costs in tokens, and compare it with what they pay you. Check what happens to your product if your main provider doubles its price or shuts down a model. Connect a second provider, at least as a backup. And don't build a business on the cheap unlimited use a platform gives you today: that gift is paid for by its investors, and it can end any day.

Let me sum it up in three sentences.

You cannot compare two companies' revenue until you ask how each one counts it.

Anthropic has almost broken even, OpenAI spends two dollars for every dollar it earns, and for now the ones making the most money are those selling them chips and cloud.

If you build on someone else's models, calculate the cost of a single task and keep a backup provider.

That is all for today. My name is Evgenii Arsentev. I bring artificial intelligence into companies and build AI content factories for businesses: systems that produce short videos for social media. Thanks for listening.