When an AI company announces a revenue figure, the first question isn't "how big?" It's "which revenue?" The answer can change the picture more than a year of growth.
The claim #
AI revenue headlines routinely blend gross figures, net figures, and annualized run-rates without saying which is which. Until a company says whether a number is gross or net of partner shares, and whether it's recognized or extrapolated, the number tells you about ambition, not about the business.
Gross versus net, in plain terms #
Accounting standards distinguish between acting as a principal and acting as an agent in a sale. Roughly: if you control the product before it reaches the customer, you can book the full amount the customer pays as revenue (gross) and record what you pay suppliers as a cost. If you're arranging for someone else to provide it, you book only your cut (net).
AI distribution makes this murky. Models are sold directly through APIs and apps, but also through cloud marketplaces and partner platforms that take a share or pay a share. Depending on the contracts, the same customer dollar can show up as gross revenue for one company and as a revenue share for another — or, in loose public statements, for both.
That isn't fraud. It's a judgment call under the standards. But it means two companies with identical underlying economics can report very different topline numbers, and a single company can tell a bigger or smaller story depending on which line it quotes.
The case that prompted this #
On X I posted about reporting that the two largest independent labs count partner revenue differently. In March 2026, Forbes, citing an analysis by Khosla Ventures partner Ethan Choi, reported that OpenAI books revenue from its Microsoft partnership net of the roughly 20% share it pays Microsoft, while Anthropic reports revenue from its AWS and Google Cloud partnerships gross, with the cloud provider's share still in the top line.1 The flows also run both ways: Microsoft reportedly pays OpenAI a share of its own Azure OpenAI revenue.2 The specific numbers matter less than the pattern: two headline figures that look comparable are built on different bases, and most coverage compared them anyway.
Run-rate is a third thing #
The other common move is "annualized run-rate": take a recent month, multiply by twelve, and report the result. It's a legitimate internal metric for a fast-growing business. As a public headline it has two problems.
First, it picks the month. A strong month annualized looks like a trend. Second, it isn't revenue anyone has recognized yet. It's a forecast dressed as a fact. When you see "$X billion run-rate," read it as "one recent month, times twelve," and ask which month.
Why this matters beyond accounting #
These distinctions aren't just for investors. Revenue claims are used to justify valuations, which justify compute spending, which shapes what gets built and how fast. If the revenue story is inflated by mixing gross, net, and run-rate, the whole capacity plan above it is too.
They also matter for oversight. A company whose economics depend on a partner's share has a different set of incentives from one that keeps most of each dollar. Who keeps the money tells you who has leverage over decisions, including safety decisions.
A reporting standard I'd use #
When I write about AI revenue, I'll try to tag every figure with three labels:
- Gross or net of partner and distribution shares.
- Recognized or run-rate, and if run-rate, which period was annualized.
- Source: the company's filing, the company's statement, or a third-party estimate.
If a number can't be labeled, I'll say so instead of quoting it. Most companies in this space are private, so filings are rare and estimates are common. That's fine as long as the reader knows which they're getting.
What to watch #
When AI companies that report a single headline revenue figure eventually file public financial statements — through an IPO or otherwise — the recognized, net figures will come in noticeably below the headline numbers that circulated before. If the filed numbers match the headlines closely, the industry has been more careful than I'm giving it credit for, and I'll say so.
Notes
- Josipa Majic, “OpenAI And Anthropic Count Revenue Differently, And Investors Are Looking Into It”, Forbes, Mar 25, 2026. ↩
- Tom Warren, “Inside Microsoft’s complicated relationship with OpenAI”, The Verge, Jun 19, 2025. ↩