OpenAI's Revenue Fell $20 Billion in Nine Days, on Paper

On 29 September 2026 Axios reported OpenAI’s annual recurring revenue was nearing $70 billion, as Yahoo Finance relayed the same day. On 8 October the Financial Times reported that OpenAI had told investors its annualised revenue was “approaching $50 billion,” as TechCrunch quotes it. Nine days, $20 billion apart. OpenAI’s business did not shrink by that much in nine days. The yardstick changed.

Rogue AI’s read: the two leading AI labs do not count revenue the same way, and the gap between their methods is now big enough to move chip stocks. In our view, any OpenAI-versus-Anthropic revenue comparison is marketing until both companies publish audited numbers in a public S-1. Our earlier piece on Anthropic’s two-customer concentration asked who pays the lab. This one asks a prior question: what the headline number is actually counting.

Why did OpenAI’s revenue figure drop by $20 billion?

The $70 billion was never OpenAI’s own number. According to the FT, as quoted by TechCrunch, it came from “attempts by OpenAI’s own investors to produce a direct comparison with Anthropic’s annualised revenues.” CNBC confirmed roughly $50 billion at the end of September and was told by a person familiar with the matter that the widely reported $68 billion “included gross revenue from OpenAI’s partners, which helps investors make a more direct comparison with its chief rival, Anthropic.”

A second explanation circulates alongside it. Yahoo Finance, citing the FT, reports that investors first tried to work out whether OpenAI’s sales reached $40 billion in August, then added the 70% growth OpenAI had told them about to that estimate and landed on $70 billion.

The two accounts are compatible, and the net figures hang together. On 19 August OpenAI’s CFO showed staff a slide with the run rate up 35% quarter to date, while CNBC put the run rate above $40 billion. By our arithmetic, that implies a start-of-quarter base near $30 billion. The 77% third-quarter growth OpenAI gave investors, per CNBC, takes that base to roughly $52 billion, close to the roughly $50 billion OpenAI reported. Apply 70% to the mid-quarter $40 billion instead and you get $68 billion. One growth rate, applied to the wrong base, and $18 billion of revenue appears on paper.

How do OpenAI and Anthropic count revenue differently?

OpenAI and Anthropic differ on one large item: sales of their models through cloud partners such as AWS, Google Cloud and Microsoft Azure. Semafor, building on reporting by The Information in March and citing people familiar with the matter, gave the cleanest example: when a customer buys $1 of tokens through a cloud partner, “OpenAI counts its 20-cent cut as revenue” while “Anthropic counts the whole $1 as revenue.” A person familiar with the numbers told Semafor the difference could reach $8 billion in annualised revenue.

The definition of the run rate itself differs too. Reuters Breakingviews, citing a person familiar with the matter, reported that Anthropic annualises the last 28 days of consumption-billed sales by multiplying by 13, multiplies the monthly subscription take by 12, and adds the two, as Simon Willison quotes from the Breakingviews column. We found no public statement of OpenAI’s method. Here is how the two sets of numbers line up as of 10 October 2026:

Note the fourth row. Anthropic itself put its run rate at approximately $9 billion at the end of 2025, in an announcement on 6 April 2026, while the leaked prospectus shows 2025 revenue of nearly $4.6 billion. For a company growing that fast, a year-end run rate double the full year is normal. It also shows how far a run rate sits from money already booked.

Is Anthropic’s gross accounting wrong?

Not on the evidence available. Anthropic’s position, as CNBC summarised it in April, is that it books partner sales gross because it is the principal in the transaction and the cloud partners are the distribution channel, that the treatment depends on each deal, and that it is consistent with GAAP. OpenAI’s then revenue chief Denise Dresser saw it differently in an internal memo CNBC reported on 13 April: “They use accounting treatment that makes revenue look bigger than it is, including grossing up rev share with Amazon and Google.” Her contrast: “We report Microsoft rev share net, which is more inline with standards we would be held to as a public company.”

In our view, both labs picked the treatment that suits their distribution. The partner revenue Dresser objects to runs through Amazon and Google; the revenue share she says OpenAI books net goes to Microsoft. Auditors will rule on the principal-versus-agent question when the S-1s go public. The telling detail is what OpenAI’s own investors did: to compare, they grossed OpenAI up to Anthropic’s method, not the other way round. That tells you which number the market treats as the scoreboard. It is also why Dresser’s memo, which on its face criticises gross accounting, reads as a pitch too. The lab-versus-lab fight over money and leverage has moved into the accounting policies.

Why a definition moved the stock market

The correction had a price. On 8 October, the day the FT figure spread, Nvidia fell 3%, Oracle nearly 6% and CoreWeave nearly 8%, according to CNBC. Futurum Group CEO Daniel Newman posted on X that “the whole OpenAI drama today was people not knowing the difference between gross and net revenue.”

Newman has the accounting right, and the reaction still made sense. Chip and cloud valuations are priced partly on what the labs will spend, and the labs’ spending is justified by numbers like these. When the most-quoted AI revenue figures are unaudited, self-defined and leaked through investors, a reclassification looks exactly like a collapse until someone explains it. In our view, the same figures also underwrite the physical build-out that towns are now pushing back on.

What should CIOs and investors do with AI lab revenue numbers?

Treat every run rate as a press release until an audited filing backs it. That does not make the numbers useless. It means asking four questions before one goes into a vendor assessment or a board slide:

• Gross or net? If cloud partner sales are in, the figure is gross. Compare like with like or say which is which. A comparison that mixes them, as the $70 billion did, can overstate the gap by tens of billions.
• Over what window? Twenty-eight days multiplied by 13 rewards a strong month. For a vendor viability check, booked revenue and cash matter more than any annualised snapshot.
• Where does your own money land? On Semafor’s reporting, a dollar you spend on Claude through AWS or Google Cloud can count in full in Anthropic’s headline, and a dollar spent on OpenAI models through Azure counts only as OpenAI’s share. That matters when you weigh buying through a cloud marketplace or directly, and when you budget what AI integration costs.
• Who is telling you? The $70 billion came through investors with a reason to want a flattering comparison. An internal memo from a rival is not a neutral audit either.

For buyers in the EU, the same discipline applies to the sovereignty questions already in AI procurement. Vendor size is one input to vendor risk, and a size figure nobody can reproduce belongs in the risk register with a question mark next to it.

When will comparable numbers exist?

Comparable numbers arrive with the public S-1s, and neither is out. As of 10 October 2026, a search of the SEC’s full-text system shows no S-1 filed by either company. OpenAI’s CFO Sarah Friar told staff in August that OpenAI “will be a public company in 2027,” per CNBC, and TechCrunch reports the listing has been pushed to early 2027. Reuters reported that Anthropic’s debut is likely to come after the US midterm elections in November 2026.

An S-1 still will not settle the comparison by itself. Both companies can follow GAAP and keep different principal-versus-agent conclusions. What the filings add is audited revenue, a disclosed policy and, most likely, the split between direct and partner sales. With those, an analyst can put both labs on one basis. Without them, every chart that ranks the two is someone’s choice of method.

Why it matters

In our view, the $20 billion is less a story about OpenAI than about the metric. Annualised revenue was built for subscription software with steady monthly bills. AI labs sell usage that swings with model launches and promotional credits, through channels where the same dollar can be counted in full or by a fifth. Until the filings land, the safe reading of any lab-versus-lab revenue ranking is that it tells you who chose the method. If you are signing a multi-year AI contract or sizing a position, decide on numbers you can check, and keep the option to switch until you can.

Related reading

• Two customers, nearly a quarter of Anthropic’s revenue on who pays the lab, from the same leaked draft.
• OpenAI’s government stake on OpenAI’s money politics.
• Self-hosted AI vs cloud APIs on what switching actually costs.
• Local opposition to AI data centres on the build-out these numbers pay for.