OpenAI's Run Rate Lost $20 Billion to an Accounting Choice
OpenAI told investors its September annualized revenue was nearing $50 billion, not the roughly $70 billion that had circulated. The gap is about how cloud sales get counted.
The number that moved AI stocks this week was not a new number. It was an old one, restated.
According to documents reviewed by the Financial Times, and confirmed to Reuters by a source, OpenAI told investors its annualized revenue was nearing $50 billion at the end of September. Weeks earlier, figures approaching $70 billion had circulated, including at a separate OpenAI event; CNBC had cited $68 billion. The difference is roughly $20 billion of run rate, and almost all of it comes from a single accounting decision.
Gross versus net, cloud edition
Both leading labs sell a large share of their models through cloud platforms. When an enterprise buys Claude through AWS or Google Cloud, Anthropic counts that sale in its revenue, then pays the cloud partner a cut, about 16% of the revenue earned through them, per Reuters. OpenAI does not include revenue from sales through cloud partners in its own figures.
The higher OpenAI number came from an attempt to put the two companies on the same basis. In the source's words to Reuters, "the discrepancy mainly arose from an attempt to produce a direct comparison." Adjust OpenAI upward to count partner sales the way Anthropic does and you get something near $70 billion. Report it OpenAI's own way and you get something near $50 billion.
Neither figure is fake. They answer different questions. But they cannot both be the headline, and the market had been reading the bigger one.
Why a restated run rate stung
Run rate is a fragile metric to begin with. Analysts treat annualized revenue as sometimes misleading, because it often means one month's sales multiplied by twelve. It flatters a company on a steep growth curve and hides seasonality. It is also the figure private AI companies prefer to talk about, because they do not file quarterly reports.
Context made the restatement land harder. Reuters reported that in the second quarter, OpenAI booked $6.7 billion in quarterly revenue to Anthropic's $11.5 billion, the first quarter in which Anthropic came out ahead. Anthropic's annualized revenue passed $65 billion in July, according to sources Reuters cited earlier, with a target of $100 billion by year-end. Those Anthropic figures include the cloud partner channel, which a Reuters analysis found made up half of Anthropic's revenue last year.
So the comparison investors care about most, which lab is bigger, depends on whose method you pick. On OpenAI's method, the gap is wide. On Anthropic's, it narrows. Neither company has published audited financials that settle it.
OpenAI still has a growth story to tell. It began 2026 at about $20 billion in annualized revenue, up from about $6 billion in 2024, and reports say it told investors of 77% run-rate growth in its third quarter and 107% in its enterprise business. Those are large numbers. They are also numbers OpenAI has not publicly broken down.
The market's reaction
CNBC reported that Nvidia, Oracle, CoreWeave and other AI stocks fell after the report. SoftBank, which has invested almost $65 billion in OpenAI, saw its shares drop as much as 7.3% in Tokyo on Friday on concerns about OpenAI's revenue growth.
The sell-off says more about how much of the AI trade is priced off a handful of private companies than about OpenAI's actual demand. Oracle and CoreWeave have built enormous capacity commitments around OpenAI. When the best-known number about OpenAI's revenue shrinks by a quarter overnight, every contract denominated in its future growth gets re-examined, even if nothing about this month's usage changed.
What the 16% cut implies
The cloud channel is not a rounding error. If roughly half of a lab's sales flow through AWS or Google Cloud, and the cloud partner keeps about 16% of what passes through, the choice between counting the full sale and counting only the lab's share moves the headline by a meaningful amount, before anyone argues about growth. It also changes how margins read. A lab reporting gross sales looks larger and carries a cost line for the partner's cut; a lab reporting net sales looks smaller with that cost already removed.
Neither presentation is wrong. Investors simply need to know which one they are looking at before they line two companies up side by side.
The real lesson: IPO season needs one ruler
Both companies are preparing to go public. That is the cleanest fix for this kind of confusion, because a listing forces a single, audited revenue definition, a gross-versus-net decision under accounting standards rather than investor-deck preference, and quarterly disclosure that makes run-rate theater unnecessary.
Until then, every comparison between the two labs will come with an asterisk. The useful question for anyone reading the next headline figure is simple: does it include revenue that flows through a cloud partner, and if so, is the partner's cut included or netted out? A $20 billion swing just showed how much that answer is worth.
None of this means OpenAI's business is shrinking. It means the market had been comparing two companies with two different rulers, and only noticed when one of them put its own ruler back on the table.