OpenAI's Revenue Is $20 Billion Smaller Than the Number Everyone Repeated, Nasdaq Notices Within Half an Hour

On September 29, multiple outlets reported that OpenAI’s annualized revenue was approaching $70 billion. On October 8, the Financial Times reported that OpenAI had told investors it was approaching $50 billion. The difference is $20 billion, which is roughly the amount by which a rounding error is no longer a rounding error.

First, the precise facts, which are the funny part. According to the FT as relayed by TechCrunch and Investing.com, the $70 billion figure came from information shared with OpenAI’s investors, and was produced by investors’ attempts to compare OpenAI directly with Anthropic. The gap is an accounting difference: Anthropic counts sales made through cloud partners such as AWS and Google Cloud in its annualized revenue; OpenAI does not. Put OpenAI on Anthropic’s ruler and you get the larger number. Use OpenAI’s own ruler and you get the smaller one. Nobody, per the reporting, fabricated anything. Two companies simply measured the same elephant from different ends.

OpenAI also told investors its revenue grew more than 70% over the period, per the reports. TechCrunch said it reached out to OpenAI for comment; its article includes no response.

The markets did not treat this as a footnote. Yahoo Finance reports that the Nasdaq 100 fell more than 300 points in about half an hour after the FT report on Thursday, with futures already down about 0.5% earlier that morning, a 30-year Treasury yield at its highest in over two decades, and Brent crude above $100. The article also mentions a separate report that Broadcom is seeking more than $50 billion in financing for a custom chip it is building with OpenAI. So the Nasdaq’s difficult morning had many authors, and this was one of them. We’d like to be fair to the $20 billion: it was not working alone.

Still, imagine being the number. For nine days, $70 billion was a headline. It had a run rate. It was going to be compared with Anthropic’s. Then it was revealed to be a figure that had been, in the FT’s words, devised by investors trying to line two companies up. It is like being a Rorschach test that turned out to be an Excel sheet.

A word on “annualized revenue,” because the phrase is doing a great deal of work. As Reuters-sourced coverage noted when the $70 billion figure was first reported, annualized run rate is often misleading since it typically multiplies one month’s revenue by 12. It is a projection, not money already booked. Yahoo Finance points to audited documents, cited by Quartz, showing OpenAI booked $13.07 billion in revenue for all of 2025. So the real range of numbers in this story runs from thirteen to seventy, depending on whether you are asking what happened, what is happening, or what might happen if every month were September.

TechCrunch also recaps the rest of the backdrop: the $70 billion figure would have put OpenAI near Anthropic’s reported run rate, OpenAI raised $122 billion in a March funding round alone, and its IPO has reportedly been pushed to early 2027. Both companies are preparing to go public, which, as Reuters-sourced coverage observed, could eventually give investors a clearer view of the finances. This is the nicest thing anyone has said about an S-1: that it will contain the same number every time you read it.

The deeper joke is that the entire AI economy is currently priced against a metric that two leading companies compute differently on purpose, or at least on policy. One counts the money that flows through the cloud partner. The other doesn’t. Both can say, with a straight face, “this is our annualized revenue,” and both would be correct, in the sense that a ruler in inches and a ruler in centimeters are each correct about your height.

And, to be clear, $50 billion in annualized revenue is not a small business. It is, by any normal standard, a mind-boggling amount of money. It’s only small next to the other number, the way a very tall man looks short standing next to a rumor.

Investors, for their part, now have a new item on the due-diligence checklist, directly under “what does the model do” and above “what does the model cost to run”: which ruler?

Sources: