OpenAI and Anthropic Count Revenue Differently, Making Their Valuations Less Precise

OpenAI and Anthropic, the two largest independent developers of frontier models, calculate the same closely watched revenue figure in different ways. Investors rely on that figure when they decide which company is worth more, and by how much. Bloomberg reports this. Both companies remain private and do not release audited financial statements to the market. Their numbers reach the public through their own statements and leaks, and there is no common standard for how to count them.

The Bloomberg piece itself is short in the version available to us. It gives no specific amounts and no details of the methodologies, so what follows covers the mechanics of the problem rather than exact figures.

Why the same word hides different numbers

Revenue at AI labs is not as simple a quantity as it looks. A handful of decisions about what to count, and how, change it the most. The differences most often arise in these places:

• Gross or net revenue. Some model sales go through cloud partners. A company can count the full amount the customer paid, or only the share the developer keeps after settling with the cloud provider.
• Extrapolation period. An annual estimate can come from multiplying a month by 12, a week by 52, or even projecting a short peak a year ahead. When growth is fast, the difference between these methods is noticeable.
• Product mix. The figure may include API, consumer subscriptions, enterprise plans and licensing agreements, or it may leave some of them out.
• Discounts and credits. Free credits, pilots and preferential terms show up differently in the final number.
• Contracts with commitments. Prepaid volumes and multi-year agreements can be counted when the cash is paid, or according to a recognition schedule.

Each of these choices is defensible on its own. Together they produce two numbers with the same label that should not be placed side by side.

How a methodology gap becomes a valuation gap

For fast-growing companies, valuation is almost always derived from a revenue multiple. An error in the base scales along with the multiple. A hypothetical example: if one company overstates its base by 10% compared with the other, then at a 30x multiple the valuation gap is larger than it first appears, and in absolute dollars it is very large. In a market where rounds close at tens of billions of dollars, that imprecision affects real money.

There is also a second side to this. Both companies raise capital constantly, sign cloud deals and discuss timelines for going public. Any growth metric becomes a bargaining argument. When methodologies differ, each side is tempted to pick the one that looks more favorable for the comparison.

It is also worth noting that both companies depend on the cloud giants. A significant share of their distribution runs through them, and that is exactly where the gross-versus-net question is sharpest.

How to read these numbers

Until a standard exists, a simple discipline helps. Before making any comparison, find out which revenue exactly is being cited: gross or net, calendar or extrapolated, with or without consumer subscriptions. Companies' statements about "annual revenue" are better treated as a rough guide than as a fact from a financial report.

Investors and enterprise customers are better served by looking at more durable signals: the trajectory of enterprise contracts, the share of recurring revenue, and dependence on a single cloud partner. This data is harder to bend to fit a convenient metric.

When one of the companies goes public, it will have to disclose revenue under regulator rules, and the comparison will then become fair. Until that point, any "OpenAI vs. Anthropic" table should be read with a caveat: the left and right columns may hold different quantities.