Anthropic's S-1: revenue, compute costs and $800bn of commitments

Anthropic filed its S-1 registration statement with the US Securities and Exchange Commission on 26 October 2026. It's the first full look at the company's audited numbers, and it shows a business whose revenue has grown faster than the compute bill that makes it possible. It also shows how much compute Anthropic has already promised to buy. Figures for 2026 are unaudited, and nothing here is investment advice.

Revenue overtook compute in Q1 2026

Revenue was $41m in 2023, $386m in 2024 and $4.6bn in 2025. In the first half of 2026 it was $16.3bn. Compute and infrastructure, the biggest cost, grew too, but more slowly: from $944m in Q1 2025 to $5.6bn in Q2 2026. The S-1 says compute is bought 12 to 18 months before it's available, so the cost is fixed in the short term and rises when committed capacity arrives, not with demand. As a share of revenue it fell from 251% to 48% over those six quarters.

A first operating profit

Operating expenses were $10.6bn in Q2 2026 against $11.6bn of revenue, an operating profit of $974m after stock-based pay of $1.3bn. The net loss for the quarter was still $66.3bn, because Anthropic records changes in the fair value of its convertible notes, derivatives and SAFEs through profit and loss. As its valuation rose, those instruments became more valuable, which is a non-cash loss of $59.6bn in the quarter.

How it compares with OpenAI

OpenAI is private and doesn't publish its accounts. The Wall Street Journal reported figures from documents shared with OpenAI's investors: revenue of $5.7bn in Q1 2026 and $6.7bn in Q2, and operating losses, including stock-based pay, of $9.3bn and $12.3bn. On those numbers Anthropic's quarterly revenue passed OpenAI's in Q2. The two companies may define revenue and costs differently, and OpenAI's figures are unaudited.

About $800bn of compute that isn't on the income statement yet

The S-1 lists Anthropic's largest compute agreements. Most are minimum spending commitments, which must be paid even if Anthropic uses less. Some, such as the extra TPU capacity for 2028 to 2031 and the AMD agreement, have no fixed dollar amount, so the S-1 gives estimates and leaves them out of its commitments table. Added up at the lower estimates, the deals come to about $782bn; at the upper estimate for the extra TPUs, about $807bn.

None of that is an expense yet. Under US GAAP, a commitment to buy cloud capacity is disclosed in the notes but isn't recorded as a cost until the service is delivered. The Broadcom TPU leases are finance leases, which go on the balance sheet when each lease starts, as equipment becomes ready for service, and are then depreciated. That's why the first half of 2026 shows $8.9bn of compute expense against roughly $800bn of signed deals. Some of the commitments can be cut short: the SpaceXAI agreements can be cancelled on 90 days' notice once their initial periods end in early 2027, and $34.9bn of the Lambda commitment depends on Lambda securing a facility lease.

Next: building its own compute

The S-1 says Anthropic is developing its own data centre infrastructure and that a growing share of its spending goes on building its own compute and "vertically integrating across the compute supply chain". That means capital expenditure and lease commitments made well before the capacity arrives. At 30 June 2026 it had $112.4bn of cash and cash equivalents and $19.3bn of short-term investments, mostly from the $110.8bn of financing it raised in the first half of the year.

Sources and method