Deep dive: Lab economics in the open: what the filings show

Why today

Two sets of numbers changed the conversation about frontier AI economics in the past three weeks. First, Reuters reviewed a draft of Anthropic's IPO prospectus, which the company filed confidentially in June. Then the Financial Times reported that OpenAI's annualised revenue was approaching $50bn, not the $70bn that had circulated. The gap came from how sales through cloud partners get counted. AI stocks fell on that report. This is the first time we have prospectus-grade numbers on what a frontier lab spends. It also comes with a public fight over what counts as revenue. Both matter to anyone whose product runs on someone else's API.

Plainly

Start with gross versus net. Say a company buys a dollar of Claude through Amazon's cloud marketplace. By Reuters' estimate, Amazon keeps about 16 cents. Anthropic books the full dollar as revenue and records Amazon's cut as a cost. That is gross. OpenAI, selling through Microsoft, books only the share that reaches it. That is net. Both can be legitimate. Accounting rules allow gross when you are the principal, meaning you control the product and carry the risk. Anthropic says that is its position. In 2025, 47% of its revenue came through Amazon and Google, so the choice matters. Now the cost side. Think of a lab as a restaurant that pays its kitchen rent in GPUs. In 2025 Anthropic spent $7.33bn on compute and infrastructure. That is roughly 1.6x its revenue, before salaries or anything else. Total operating expenses were $12.65bn. The reports don't split that compute into training new models and serving customers. That split is the whole question. Training is a fixed bet you hope to spread across many customers. Serving cost rises with every token sold. If most of the spend is training, growth can fix the economics. If most of it is serving, growth alone won't. One more trap: the headline net loss was $42bn. About $34bn of that was an accounting charge tied to the rising estimated value of financing, not cash out the door. The operating loss is the number that tells you about the business.

What the evidence shows

Start with what is established, and how. These numbers are Reuters' reading of a draft prospectus, not a public filing. A prospectus must carry audited annual statements, so the 2025 figures are probably the firmest numbers any frontier lab has produced. We are still seeing them second-hand. Revenue grew 12-fold. Compute roughly tripled. Nearly a quarter of revenue came from two customers, and the company warns that many large clients are not on long-term contracts. Anthropic has also committed $518bn to compute partners over about a decade. About 80% of that cannot be cancelled. Then 2026 gets more interesting, and murkier. Anthropic reportedly told investors that second-quarter revenue topped $11.5bn. Quartz, citing the Wall Street Journal, reported that compute ate 71 cents of each revenue dollar in the first quarter. The projection for the second quarter was 56 cents. These are preliminary, self-reported figures shared during a fundraise. The adjusted measure reportedly excludes stock-based compensation, and none of it is audited. On OpenAI, the FT established that the $70bn came from investors restating OpenAI's numbers on Anthropic's gross basis. Yet when the FT first asked, OpenAI had described the $70bn as a "net" revenue metric. Nobody has explained that inconsistency. What is claimed but not shown is that run rates mean much at all. Anthropic's run rate annualises the prior 28 days of usage revenue, and usage is not recurring the way a subscription is.

How people read it

The AGI-forward reading says the losses are the price of scaling, and the curve is turning. Revenue grew 12-fold and compute per revenue dollar is falling. A lab that reaches operating profit while still training frontier models has shown real operating leverage. The AI-realist reading is made most forcefully by Ed Zitron, who sells a paid newsletter on the AI bubble. He argues the profitable quarter lined up with discounted compute from SpaceX, and that an adjusted, unaudited metric proves little. In his words, Costs increase linearly with revenue. The market reading sits in between. The concentration risk is real: two big customers, cloud partners who are also investors and suppliers, and mostly fixed commitments. Fixed commitments look great while demand grows and brutal when it stalls. Gary Marcus adds that Chinese open-weight models have forced price wars, which would limit how far API prices can rise. The facts don't settle this yet. A few audited quarters would.

What it means for builders and investors

Our read is that the 2025 numbers show frontier tokens were sold below full cost. Spending roughly 1.6x revenue on compute is a subsidy, whoever ends up paying for it. If the 2026 figures hold, the gap is closing through cheaper serving rather than higher prices. That is better news for startups than we expected, and it softens our view, but only provisionally. For builders, the bigger risk is probably not a sudden doubling of prices. It is that a lab carrying $518bn of mostly fixed commitments will price to fill capacity, then change terms when it needs margin. For investors, gross versus net is a live question in any startup that resells model access. One question worth exploring in diligence: what is the company's net revenue under accounting rules, and what happens to gross margin if token prices double? Clear answers to both say a lot about the business.

What to watch

Three things would settle this. First is the public prospectus, with audited interim 2026 numbers and a clear definition of adjusted operating profit. It has to appear before any listing, though the timing is not confirmed. Second is whether compute stays near 56 cents per revenue dollar once the discounted SpaceX period Zitron describes has ended. Third is any audited OpenAI filing on a net basis, which would allow a like-for-like comparison of the two labs. If audited compute stays below revenue for several full quarters, the subsidy story gets much weaker.