The twenty-cent dollar
The cleanest way to understand the split is one transaction. A customer buys $1 worth of tokens through a cloud marketplace. OpenAI's arrangement with Microsoft means it records about 20 cents of that dollar as its own revenue and Microsoft keeps the rest. Anthropic takes the opposite position on the same question: it is the principal, so it records the whole dollar, then lists the cloud partner's share as a cost of doing business.
Both treatments are legitimate under US accounting rules. Principal-versus-agent judgments are routine for SaaS and marketplace businesses and turn on how much control the company retains over pricing and the customer relationship. The point is not that one company is right. It is that the two headline numbers do not measure the same thing, and every ranking, valuation multiple, and “who is winning” narrative built on comparing them directly has been wrong by construction.
How an $8 billion accounting gap became a $20 billion error
The accounting divergence was only the starting point. In late summer, OpenAI told investors its annualized revenue had grown more than 70 percent off a roughly $40 billion baseline, a figure Bloomberg had confirmed in August. Some investors then tried to make OpenAI's net-basis number comparable to Anthropic's gross-basis number by extrapolating OpenAI's growth rate using Anthropic's methodology. The arithmetic landed at approximately $70 billion.
That number, first reported by Axios, circulated through investor decks and financial media. A source familiar with the investor documents later confirmed to multiple outlets that the $70 billion figure never came from OpenAI. The Financial Times, citing documents shared directly with OpenAI's financial backers, put the September figure at about $50 billion. The gap between what the market had priced in and what the documents showed was $20 billion, and the selloff that followed was the market repricing the misunderstanding.
The actuals are the interesting part
Lost in the run-rate theater are the two actual revenue numbers, which tell a plainer story than any extrapolation. Anthropic collected roughly $4.6 billion in 2025, half of what its run rate implied. OpenAI is on track for about $35 billion in 2026, half of its target. Annualized run rates are extrapolations from short-term performance and contract value; they were never the same as money actually recognized over a full year. Analysts have said so all year. The market listened anyway.
The draft prospectus adds the other half of the balance sheet: $518 billion in future cloud and computing obligations. That is the cost of staying in the race, committed before the revenue that would pay for it has been collected. It is also, incidentally, the kind of line that ends up in an S-1 whether or not the company wants it there. Both labs have filed confidential IPO paperwork. A public listing forces audited, standardized financials, which is the one development that will finally end the apples-to-oranges era. Until then, every comparison arrives with an asterisk, and the asterisk is doing most of the work.
A reading rule for the next run-rate headline
The practical takeaway is narrow and reusable. When you see a run-rate number from either lab, ask one question first: gross or net. Anthropic's number includes money that flowed through Amazon and Google to their owners. OpenAI's counts only what stayed with OpenAI after Microsoft took its share. Adjust for that before you compare, and the race looks far closer than the headlines suggest — not because anyone hid anything, but because the rulers were never the same.
Sources
- [1] TechTimes — “Investors Built $70B OpenAI Revenue Estimate Using Wrong Method; AI Stocks Fell When FT Corrected” (Oct 9, 2026)Read source
- [2] KuCoin Flash — “Differences in OpenAI and Anthropic revenue reporting spark market concerns” (Oct 10, 2026)Read source
- [3] Cointime — “Concerns Arise Over Revenue Reporting Differences Between OpenAI and Anthropic” (Oct 10, 2026)Read source
- [4] Exploding Topics — “Anthropic Revenue and Valuation Statistics (October 2026)” (cites WSJ, Reuters, Axios)Read source
- [5] ValueAddVC — “OpenAI-Anthropic Revenue Dispute: Gross vs Net ARR Explained” (cites Semafor)Read source
- [6] AInvest — “The $65 Billion Question” (cites Bloomberg, OpenAI investor memo)Read source