OpenAI's annualised revenue was approaching $50B at the end of September, the Financial Times reported on 8 October, citing a note sent to the company's investors. Nine days earlier the figure circulating in the press was $70B, and the $20B between them is not money OpenAI lost.
Annualised revenue takes recent sales and stretches them across a full year, so the result shifts with whichever period a company picks. OpenAI told its investors that revenue had grown more than 70% over the months the note covered, and that growth rate is the only company figure the $70B estimate rests on.
Investors built the rest themselves, and they built it on the wrong base. Their August estimate of about $40B had been restated onto Anthropic's accounting, which counts sales made through cloud partners such as Amazon Web Services and Google Cloud, and OpenAI leaves that partner revenue out of its own figure.
Applying OpenAI's reported growth to that inflated base took $40B to roughly $70B. Nobody at OpenAI published the number, and the company has not commented publicly on either figure.
The Nasdaq Moved On The Number
Technology stocks dropped within about half an hour of the report on Thursday afternoon. The Nasdaq Composite was down 378 points, or 1.37%; the Nasdaq 100 fell 1.51%; and the S&P 500 lost 0.60%.
Oil prices were already climbing before the report landed, and the interest rate the US government pays on 30-year debt sat at a more than two-decade high. Bitcoin fell 2.24% to around $81,500 in the same session.
What moved was the number attached to OpenAI rather than anything inside the business. The size of that move is a measure of how closely AI revenue figures are now watched.
OpenAI had spent the previous week showing products rather than numbers. It launched Dots, a set of AI agents aimed at office workers, together with a model called GPT-6.1 Sol at its developer conference, and neither features in the revenue figures now under discussion.
Run Rates Are Not Audited Accounts
A run rate, the measure both companies publish, is a projection rather than a result, built by taking a recent period of sales and multiplying it out to twelve months. No auditor signs it, no regulator defines it, and no rule says which revenue belongs inside it.
Private AI companies have made it the standard unit of comparison anyway, because it is the only number they publish often enough to track. Reporters, investors and rival firms all read the same figures without a common definition underneath them.
A model sold through Amazon Web Services counts for Anthropic and not for OpenAI, so two businesses of identical size would publish different numbers on the same day. That is the cost of a measure nobody has agreed on.
Reported figures for OpenAI have moved repeatedly through 2026, from roughly $25B in March to close to $50B now. Each step arrived through a note to investors or a press report rather than a filing. That is how a $20B discrepancy survived nine days in print.
What The Audited Books Show
OpenAI's last full audited year is 2025, and the accounts for it record $13B of revenue against $34B of spending. Ed Zitron, a technology writer who publishes the newsletter Where's Your Ed At, obtained the documents. The Financial Times verified them independently, and the figures appeared in June 2026.
Running at an annualised $50B would put the company nearly four times its 2025 revenue inside a single year. No audited spending figure has been published alongside that growth, so the $34B from 2025 remains the last confirmed cost base anyone can set against that revenue.
Anthropic Sits Above OpenAI
Anthropic reported a run rate above $65B at the end of July 2026, more than seven times its run rate at the end of 2025. At $50B, OpenAI sits below the rival it is usually ranked above.
Anthropic's most recent completed quarter brought in more than $11.5B, against $787M in the same quarter of 2025, and the company reported a profit on its core operations for the period. Money actually booked works out at an annual pace closer to $46B than $65B, because a run rate measures the latest month while a quarterly figure counts what came in across three months.
The dates do not line up either, with Anthropic's figure from July and OpenAI's from September. Two months is a long gap in a market growing at these rates, and neither company has published a figure calculated the other's way.
After a May funding round valued it at $965B, Anthropic filed confidentially for a stock market listing, a route that lets a company prepare without publishing its accounts until late. It is working with Morgan Stanley, Goldman Sachs and JPMorgan, and could reach public markets as soon as this autumn. OpenAI is expected to follow in 2027.
Spending Set Against Revenue
Sam Altman has put OpenAI's planned investment at $1.4T to build 30 gigawatts of computing power, roughly the output of 30 large nuclear reactors. Reuters later reported a compute spending target of around $600B through 2030, and the company has not reconciled the two figures in public.
Broadcom is seeking more than $50B of financing for a custom chip it is building with OpenAI, according to the Wall Street Journal. Borrowing on that scale was among the pressures already weighing on technology stocks this week.
Those commitments buy physical capacity, and the queue for it runs through every site where power is available. The demand reaches past the data centres and into the fibre that carries the traffic between them.
Reports in recent weeks have put OpenAI in early talks over funding at a valuation of $1.2T or more. That money would be raised against revenue near $50B a year, and against the spending commitments already made.
A Number Nobody Published
OpenAI's position is unchanged from the day before the report. Revenue grew more than 70%, the business is running at roughly four times its 2025 size, and it is still spending considerably more than it earns.
For nine days the market held a figure that put OpenAI ahead of Anthropic, and that figure came from arithmetic performed outside both companies. What the correction changed is the comparison, not the company.
