OpenAI’s $20 Billion Gap: Why AI Revenue Figures Don’t Compare

Roughly $20 billion in revenue vanished at OpenAI in just over a week, at least on paper. In late September, an annualized revenue figure of nearly $70 billion was making the rounds; now, according to the Financial Times, the company has given investors a figure of nearly $50 billion. OpenAI did not lose any business in the process. The difference is a matter of arithmetic, and that is exactly why it deserves a close look: it shows how poorly the AI industry’s most quoted metrics hold up when you put them side by side.

Key takeaways

• According to the Financial Times, OpenAI reported annualized revenue of nearly $50 billion for the end of September to investors; earlier reports had put it at nearly $70 billion.
• The higher figure came from adding partner sales on a gross basis to make OpenAI comparable with Anthropic, which reports exactly that way.
• Growth remains strong: according to CNBC, the revenue run rate rose 77 percent in the third quarter, and 107 percent in the enterprise business.
• On the stock market, Oracle shares fell more than 5 percent on October 8, and the Nasdaq was down about 1.4 percent in the afternoon.
• At the same time, according to Bloomberg, OpenAI is negotiating at least $30 billion in new capital at a valuation of $1.4 trillion.

Where the $20 billion came from

Annualized revenue, known in the industry as the run rate, is not a balance-sheet figure. It extrapolates revenue from a short period, usually one month, to a full year. For fast-growing companies it is a common yardstick, but it depends heavily on what gets counted in the base month. That is where the difference lies. On September 29, the news site Axios reported that OpenAI’s rate was approaching $70 billion. According to a person familiar with the numbers quoted by CNBC, however, that figure included gross revenue from deals made through partners. The investor materials showed the leaner figure of about $50 billion.

Gross means that when an AI model is sold through a cloud provider’s platform, the entire amount the customer pays counts as the model maker’s revenue, and the platform operator’s share shows up as a cost. Net counts only what reaches the model maker after that share is deducted. Anthropic reports gross, while OpenAI reports some of its partner deals net. Both methods are permitted in principle; what matters is who counts as the actual provider in a given deal. They simply produce numbers of different sizes.

Why the comparison with Anthropic is so tricky

How large the effect can be is shown by Anthropic’s confidential IPO documents, which Reuters was able to review. According to them, about $2.16 billion ran through the cloud marketplaces of Amazon and Google in 2025, 47 percent of annual revenue. The platforms received about $351 million in fees, which Anthropic books as sales and partnership costs. The company justifies gross reporting by saying it sets the prices and delivers the service itself. OpenAI had already told investors in June that this approach inflates Anthropic’s revenue by billions. What we know about Anthropic’s numbers and risks from the filing is summarized in our analysis of its IPO prospectus.

The twist in the current correction: investors had scaled up OpenAI’s figure so they could set it next to Anthropic’s run rate, which according to Bloomberg stood at more than $65 billion at the end of July. That comparison is shaky to begin with. Neither company is publicly traded, both figures come from investor communications that only become known secondhand, and neither is audited revenue. Anyone who reads a race for the top into this is comparing two extrapolations measured with different rulers.

What the market read into it

The market nonetheless took the news as a setback. On October 8, Oracle shares lost more than 5 percent according to Benzinga; Nvidia, AMD and Microsoft also slipped, and the tech-heavy Nasdaq was down about 1.4 percent in the afternoon. That said, rising oil prices and bond yields also weighed on stocks that day. The nervousness is understandable all the same: Oracle, Nvidia and other infrastructure providers have multibillion-dollar contracts with OpenAI whose value depends on whether the customer’s income keeps pace with its data center plans. We put the size of those commitments in context in our look at OpenAI’s spending plans.

For OpenAI’s fundraising, the correction changes the math noticeably. According to Bloomberg, the company is negotiating at least $30 billion at a pre-money valuation of $1.4 trillion. Measured against a $70 billion run rate, that would be about 20 times revenue; measured against $50 billion, closer to 28 times. As recently as March, OpenAI raised up to $122 billion at a valuation of $852 billion. Sam Altman has ruled out an IPO this year; 2027 is considered possible.

What the number actually stands for

Amid all the noise about the gap, it is easy to overlook that the business itself is growing. According to CNBC, OpenAI showed investors 77 percent growth in its revenue run rate in the third quarter, and 107 percent in its enterprise business. By year end, the company expects a rate of at least $70 billion according to Bloomberg, driven mainly by business customers. So the $70 billion may well arrive, just a quarter later and under the company’s own, narrower method of counting.

Bottom line: the metric is the problem, not the revenue

The episode is less a sign of weakening AI demand than of an information problem. As long as the big AI labs stay private and their numbers only leak out of investor rounds, valuations, supplier stock prices and public horse races all rest on extrapolations built by each company’s own rules. With their IPOs at the latest, both companies will have to present audited, comparable financial statements. Until then, the first question about any new run rate is not how high it is, but how it was counted.