Anthropic S-1 Shows $800 Billion in Chip Deals, Serving Is Just 27% of Compute

Anthropic spent less than a third of its $8.9 billion first-half compute bill on serving customers, according to its S-1 filing. The rest went to training, research, and internal use of Claude. Meanwhile the company has signed roughly $800 billion in TPU and GPU contracts.

Revenue per compute dollar jumped from $0.40 in Q1 2025 to $2.08 in Q2 2026. The company attributes this to capacity purchased 12 to 18 months ahead. The compute base "did not fluctuate with in-period demand," the filing says, so the ratio rose as demand surged.

What does the filing actually show?

Anthropic spent $2.4 billion serving customers in the first half of the year, against $16.3 billion in revenue. Serving Claude covers its costs. Paying for what is on order is the harder question.

The five compute deals with the largest stated amounts would cost an estimated $103 billion a year on average once delivered. That is more than twice the Q2 revenue pace of about $46 billion a year. The company itself warns margins may compress as new hardware lands ahead of the revenue it supports.

Anthropic's three main costs:

• Cost of revenue — the inference cost of running Claude for paying customers
• R&D — mostly compute for training and research, plus internal use
• Sales and marketing — partner cuts and cloud reseller fees

Cost of revenue rose 2.3x year over year, while revenue rose 14x. Anthropic now keeps an estimated 85 cents per dollar after serving costs, up from 12 cents. R&D is up about 5x, absorbing 66% or more of compute. That is the main source of the operating loss. Sales and marketing rose 16x; including it, the company keeps 66 cents per dollar.

Where does the compute go?

Serving took at most 44% of compute in the first half of 2025 and at most 27% in the first half of 2026. The filing's own table shows revenue of $1.16 billion growing to $16.29 billion between those periods.

One caveat sits inside the filing. Anthropic says it has "consistently operated in a compute-constrained environment where customer demand exceeds available compute." Serving costs may be muted by supply limits, not efficiency alone. The company runs compute as "a single, shared resource across all workloads" and does not disclose how it splits capacity between training, research, and serving.

Serving has also gotten cheaper. In September, Opus 5.5 cut costs per task by about 40% versus Opus 5, and Fable 5.1 cut typical workload costs by about 25% versus Fable 5. Each came from "a combination of a lower price and improved efficiency," made "while preserving our margins," the company says.

How binding are the contracts?

The named deals total about $782–807 billion across 2026–2036. The S-1 leaves its own total blank.

• Broadcom — five-year TPU equipment leases, about $161.2 billion in rent excluding interest; no exit except on default
• Google — at least $114.5 billion between April 2026 and July 2033; minimum owed regardless of use
• Other TPU providers — estimated $125–150 billion, delivered 2028–2031; exit terms not stated
• Amazon — at least $110 billion for AWS, including Trainium, May 2026 to April 2036
• SpaceXAI — up to $120.5 billion in Nvidia GPUs into 2029, about $3.1 billion a month at full delivery; cancellable on 90 days' notice after January and March 2027
• Microsoft — at least $44.9 billion, November 2026 to May 2033
• Nscale — up to $44.6 billion in Vera Rubin NVL72, billed only as capacity is delivered
• Lambda — about $41.3 billion into 2034; $34.9 billion of it depends on Lambda securing the facility lease
• AMD — Instinct MI450 supply that may exceed $20 billion; no fixed commitment

Minimum purchase commitments, excluding equipment leases, grew from about $55 billion at the end of last year to about $265.6 billion by the end of Q2. About $245 billion in new minimums has been signed since. TPUs are the largest line. The Nvidia capacity, Microsoft's aside, is the easiest to exit.

What is missing?

The filing omits the total commitment figure, the contractual obligations table, Q3 revenue, and Run-Rate Revenue. An amended S-1 would normally fill these in before pricing. No date has been given.

The IPO is likely to beat rival OpenAI's, which filed confidentially before CEO Sam Altman said his company's offering will not come this year. Whether AI labs can cover serving costs remains open, and doubts about an AI bubble persist. Anthropic appears to clear the serving bar. The $800 billion question is the racks on order.