Anthropic's S-1 Reveals $800 Billion Compute Contracts, Only 27% Serving Customers

Anthropic spent less than a third of its $8.9 billion first-half compute bill on serving customers, while committing roughly $800 billion to future TPU and GPU capacity, according to its S-1 filing. The rest of that compute bill — more than two-thirds — went to training, research and internal use of Claude.

The filing lands amid widespread doubts about an AI bubble. Anthropic's IPO is likely to beat rival OpenAI to market, after OpenAI filed confidentially before CEO Sam Altman said its own offering will not come this year.

What do the numbers say about serving Claude?

Serving customers cost Anthropic $2.4 billion in the first half of the year, against $16.3 billion in revenue. Cost of revenue rose about 2.3x year over year, but revenue rose 14x, so the company keeps an estimated 85 cents of every dollar after serving costs, up from 12 cents a year earlier.

Anthropic breaks its costs into three lines: inference for paying customers; R&D, which is mostly training and research compute plus internal use; and sales and marketing, covering channel-partner and cloud-reseller fees. R&D grew about fivefold and absorbs 66% or more of compute — the main source of the operating loss. After partner fees and marketing, the company keeps about 66 cents per dollar, up from 2%.

Revenue per compute dollar jumped from $0.40 in Q1 2025 to $2.08 in Q2 2026, which the company links to capacity purchased 12–18 months ahead. Serving has also become 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 through, in the company's words, "a combination of a lower price and improved efficiency" made "while preserving our margins."

One caveat sits inside the filing. Anthropic says it has "consistently operated in a compute-constrained environment where customer demand exceeds available compute," which suggests serving costs were muted by limited supply, not only by efficiency. Compute runs as "a single, shared resource across all workloads," and the company does not disclose how it splits the resource among training, research and serving.

Where is the $800 billion going?

Anthropic runs a "multi-chip compute architecture across AWS Trainium, Google TPUs, Nvidia GPUs, and AMD GPUs," without percentage splits or gigawatt figures. The named contracts, summed by the source, total roughly $782–807 billion:

• Broadcom: five-year equipment leases for TPU capacity, rent of approximately $161.2 billion excluding interest; some capacity online this year, more from 2027.
• Google: at least $114.5 billion between April 2026 and July 2033.
• Other TPU providers, including Google: an estimated $125–150 billion delivered 2028–2031.
• Amazon: at least $110 billion on AWS, including Trainium, from 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, expected December 2026.
• Microsoft: at least $44.9 billion on Azure, November 2026 to May 2033.
• Nscale: up to approximately $44.6 billion in Nvidia Vera Rubin NVL72 into 2033.
• Lambda: about $41.3 billion in Nvidia GPUs into 2034.
• AMD: an Instinct MI450 (Helios) supply agreement expected to exceed $20 billion, with first capacity in 2027.

TPUs are the largest line overall. 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, with about $245 billion in new minimums signed since, leases included.

Can Anthropic afford the racks on order?

The five biggest deals by stated amount would cost an estimated $103 billion a year on average once delivered, more than twice Anthropic's revenue at Q2's pace of about $46 billion a year. The company itself warns margins may compress as new hardware lands ahead of the revenue it supports.

Exit terms vary widely. The Nvidia deals, Microsoft's aside, are the easiest to escape: SpaceXAI's non-cancelable periods end in January and March 2027, after which either side can cancel on 90 days' notice. The Google and Amazon minimums are owed regardless of use, and $34.9 billion of the Lambda deal can be terminated only if Lambda fails to secure its facility lease. Each contract carries its own termination conditions.

Because expenses are recognized "as compute becomes available to us," the compute base "did not fluctuate with in-period demand," so revenue per compute dollar climbed as demand surged. When demand outruns capacity, Anthropic buys short-term compute, which "typically carries a higher cost per unit."

The filing leaves blanks an amended S-1 would normally fill before pricing: the total commitment figure, the contractual obligations table, Q3 revenue and Run-Rate Revenue. Microsoft's minimum starts in November, SpaceXAI's full delivery is expected in December, and more Broadcom TPU capacity plus AMD's first tranche follow next year — by which point most of these deals will be billing.