At half past eleven on a Monday night, a developer at a ten-person legal-software startup opens the configuration file that decides which model reads her customers’ contracts. For a year, that model has been Claude. She has spent the evening running the same two hundred clauses through Claude and through a rival model, which now costs less for work that is close enough in quality. She changes one line, runs the tests and goes to bed. Nobody signs anything, and nobody at either lab is told. This developer is hypothetical, a sketch of the customer that Anthropic describes in its own risk factors. The ease of her decision is not hypothetical. It is written into the contracts that Anthropic disclosed on 26 October.

That day, Anthropic filed its registration statement for a public listing with the US Securities and Exchange Commission. In the S-1, a frontier lab sets out its economics in a document for which it carries legal liability. It describes a business with two very different time horizons. On one side are compute contracts that run into the 2030s and must be paid “regardless of actual usage”. On the other side is revenue that is mostly metered by the token, from customers whom Anthropic cannot hold.

What Anthropic has promised to pay

The commitments appear in the section of the filing about liquidity. In November 2025, Anthropic agreed to spend at least $44.9 billion with Microsoft between November 2026 and May 2033, “regardless of actual usage”. In March 2026, it agreed to spend at least about $114.5 billion with Google between April 2026 and July 2033. If its actual spending falls short, it must pay Google the difference. It has also committed to buy additional TPU capacity for delivery between 2028 and 2031, and it estimates those payments at $125 billion to $150 billion. In April 2026, it agreed to spend at least $110 billion with Amazon between May 2026 and April 2036, again with a true-up if usage falls short.

Those three cloud contracts add up to more than $269 billion in minimum spending, before the leases and the GPU deals are counted. Rent on Anthropic’s five-year equipment leases with Broadcom, which give it access to TPU hardware in data centres that others operate on its behalf, comes to about $161.2 billion excluding interest. The filing describes those leases as non-cancellable by either party except in the case of a default. At their maximum, further agreements with SpaceXAI, Nscale, Lambda and AMD are worth more than $200 billion, although some accrue fees only as capacity arrives and the SpaceXAI deals become cancellable on 90 days’ notice in 2027.

Much of this spending was agreed this year. According to the financial statements, agreements that Anthropic signed or amended after 31 December 2025 carry about $417 billion in future minimum payments. Another note puts the minimum payments under agreements signed or amended after 30 June 2026 at about $245 billion. In its risk factors, Anthropic says that its contracts “obligate us to pay for resources regardless of actual utilization.” It adds: “These commitments represent substantial fixed obligations.” Some of them run until 2045. The preliminary prospectus still leaves the aggregate total blank.

Revenue that can leave with little notice

The revenue side is growing fast, but it is not locked in. Anthropic’s revenue was $41 million in 2023, $386 million in 2024 and about $4.6 billion in 2025. In the first half of 2026 it reached $16.3 billion, against $1.2 billion a year earlier, an increase of 1,300%. Consumption-based fees, which customers pay per token for the model they choose, made up $13.1 billion of that, or 80%. Subscriptions, which generally run for 12 months or less, made up the rest.

Anthropic spells out what that means. “Many of our customer contracts allow customers to reduce usage or terminate with limited or no notice,” it writes. “Customers are not obligated to renew their subscriptions or contracts with us.” It adds that customers may reduce or stop their usage at any point within a period because of budgets or procurement rules, “even where demand for our models, platform, and applications remains strong.”

A growing number of enterprise customers do sign non-cancellable annual or multi-year contracts, which set minimum spending in exchange for a committed price list. The contracted part is still small next to the revenue. As of 30 June 2026, remaining performance obligations, the amounts that customers have already committed to under non-cancellable terms, amounted to $10.8 billion. Anthropic expects to recognise about three-quarters of that within a year.

In the risk factors, the company puts the two sides of its business in one sentence. Its revenue “is primarily consumption-based, and much of it is not subject to long-term contractual commitments, while our compute and infrastructure commitments generally extend over multiple years. As a result, we may enter into these obligations based on revenue expectations that may not be realized.” If revenue falls behind the payments, a default under some commercial agreements could trigger cross-default provisions in its credit agreement.

Why the margins look good for now

For now, Anthropic benefits from the fact that its compute costs are fixed. Its spending on compute and infrastructure rose from $2.33 billion in the first half of 2025 to $8.95 billion in the first half of 2026. As a share of revenue, however, that spending fell from 201% to 55%. The filing explains that revenue grew “against a compute base that was primarily committed in advance and did not fluctuate with in-period demand.” Its operating loss narrowed from $2.1 billion to $693 million.

The data scientist Max Woolf wrote on 25 October that he would “be very curious to see the Anthropic IPO prospectus and the inference profit margins once it becomes public soon.” The main statements no longer show a gross profit line, because Anthropic changed its presentation in 2025. A note to the accounts still groups costs by function. Cost of revenue, which Anthropic says consists substantially of compute for customer-serving workloads, was $2.38 billion in the first half of 2026. By this article’s arithmetic, measured against $16.3 billion of revenue, that implies a gross margin of roughly 85% on serving customers. On the same basis, the margin was about 12% in the first half of 2025. Most of the company’s spending goes to research and development. That spending reached $9.74 billion in the same half-year, and Anthropic says that it consists primarily of compute for model development, research and internal use, along with staff costs.

The revenue also depends on a steady supply of new models. Model versions released in the six months to 30 September accounted for more than 80% of revenue in the third quarter. Net revenue retention, which compares what the same customers spend a year apart, stood at about 900%. Those customers expand quickly while the models improve. The figures do not show what they would do if a rival’s model became better or cheaper.

A market that keeps cutting prices

That question became sharper during the past week. On 20 October, Apollo’s chief economist Torsten Slok argued that the slower growth in what companies spend on tokens suggests that compute faces “a flat, highly price-sensitive demand curve, where even modest changes in token prices can drive buyers to throttle usage or substitute cheap open-source alternatives.” Five days later, Woolf assembled the price changes of OpenAI and Anthropic since June. His list included cuts of up to 80% and Anthropic’s pricing of Opus 5.5 at $4 per million input tokens and $20 per million output tokens. His conclusion was that “the idea that token prices will suddenly rise is more a fantasy.” On 26 October, the same day as the filing, The Register reported that OpenAI’s new GPT-6.1 Terra was priced at $1 per million input tokens and $5 per million output tokens, which OpenAI says is 50% below the earlier Terra model.

Anthropic says that with Opus 5.5 it cut customers’ cost per task by about 40% compared with Opus 5 while preserving its margins. In the risk factors, it warns that it “may be required to reduce our prices, offer shorter contract durations or offer alternative pricing models”, and that agentic products are “structurally more compute-intensive” than its other products.

In an NBER working paper published this month, Mark Egan, Dominic Russel, Claire Shi and Vincent Yao studied card transactions from 174 million US consumers. They found that annual retention for paid AI services ranges from 20% to 45% across major providers, and that switching rises around major model releases. Their structural model finds consumers “highly price sensitive”. The study covers consumer subscriptions rather than the enterprise and API customers who provide most of Anthropic’s revenue. It does show how little loyalty people show when the alternative is a click away.

What the filing leaves open

These are company figures in a preliminary prospectus, and net revenue retention is Anthropic’s own operating metric. Anthropic also estimates that 5% to 7% of its revenue in the first nine months of 2026 came from customers that it later found violating its usage policy or commercial terms. Stronger enforcement since August may reduce that revenue and the company’s growth rate.

The filing also records past forecasting errors. “In the past, we have significantly underestimated our compute requirements,” Anthropic writes, which left it short of capacity. It warns that compute may arrive before demand and produce “periods of excess or underutilized compute and committed cash outflows that do not produce immediate revenue.” Both errors are possible under contracts that Anthropic can adjust only with difficulty, while many of its customers can change course with limited or no notice.

On Tuesday morning, the hypothetical developer checks her dashboard. Her bill for the night is a few dollars lower, and her customers have noticed nothing. Her customers’ contracts are now read by the cheaper model. The payments that Anthropic owes for its data centres will still come due, quarter after quarter, through the next decade.

References

  • Anthropic, PBC (2026). Form S-1 Registration Statement. U.S. Securities and Exchange Commission. sec.gov
  • Claburn, T. (2026). OpenAI rolls out budget-priced mid-tier GPT-6.1 Terra. The Register. theregister.com
  • Egan, M. L., Russel, D., Shi, C., & Yao, V. (2026). Competition in Generative AI: Evidence from Consumer Demand. NBER Working Paper 35953. nber.org
  • Slok, T. (2026). Slowing Token Spend Shows Compute Demand Isn’t Unlimited. Apollo, The Daily Spark. apollo.com
  • Woolf, M. (2026). OpenAI and Anthropic cut their LLM prices by up to 80% within a few months. Why? Max Woolf’s Blog. minimaxir.com