Anthropic's S-1 shows serving paying customers is at most a quarter of its compute bill

Anthropic's S-1, filed with the SEC on October 26, is mostly the document you'd expect: revenue up from $1.16 billion to $16.29 billion year over year for the first half, an operating loss that shrank to $693 million, and pages of risk factors about power grids and export controls. Tucked into the notes to the financial statements is something more useful for anyone trying to understand what frontier AI actually costs to run: a split of the company's expenses by function, with a footnote on which compute workloads land where.

The short version is that serving paying customers is the smaller part of the bill.

Two ways of slicing the same money

The headline income statement groups expenses by type. On that view, "compute and infrastructure" was $8.945 billion in the first half of 2026, Anthropic's largest cost by far, against $2.334 billion a year earlier.

A supplemental table (on page F-40) regroups the same total operating expenses by function, and its footnotes say what each line contains:

• Cost of revenue: "substantially all" compute and infrastructure for customer-serving workloads, plus some personnel and overhead.
• Research and development: "primarily" compute and infrastructure for model development, research and internal usage, plus personnel, software and overhead.
• Sales and marketing: among other things, the compute used to give people free access to Claude.

Put the two tables side by side and the shape of the compute spend becomes visible, at least in outline.

The percentages are our arithmetic, and they overstate the serving share, because cost of revenue also carries some salaries and overhead. Even so, in the first half of 2026 at most about a quarter of Anthropic's compute and infrastructure bill went to serving paying customers. The rest sat in the other functions: training and research, Anthropic's own use of Claude, and free usage.

What the numbers say, and what they don't

Two trends run in opposite directions. Serving became dramatically cheaper relative to what customers pay: cost of revenue was 87.6% of revenue in the first half of 2025 and 14.6% in the first half of 2026. Meanwhile the functional R&D line grew more than fivefold, from $1.85 billion to $9.74 billion, faster than the compute bill itself.

The filing's own explanation is that compute is bought 12 to 18 months ahead under multi-year contracts, recognized when it becomes available rather than when it's used, and "generally fixed in the near term". Anthropic says it treats that capacity as "a single, shared resource across all workloads" and plans it "on a total basis rather than by individual workload." So revenue that grows 1,368% year over year in a quarter, against compute expenses that grew 300% (the company's figures for the three months to June 30), mechanically drives the serving ratio down. The company says it continuously directs that pool "to its highest-value use case", so capacity that customers don't absorb is available for research and internal work.

That framing matters for reading the table. The functional split is an accounting allocation of a shared pool, not a meter on each GPU. The S-1 doesn't break R&D compute into pre-training, reinforcement learning, evaluations and internal agent use, and it doesn't give a separate figure for free-tier compute, so nobody outside the company can say how much of that $9.7 billion trained the next model and how much ran Claude Code for Anthropic's own engineers.

Internal use is clearly not a rounding error, though. The filing says Claude wrote more than 90% of contributions to Anthropic's codebase in September 2026, and that daily code volume per engineer in the third quarter was over 45 times the 2024 average. It also describes Claude "running experiments, generating and processing training data, automating evaluations and safety audits", which the company calls "closing the loop". All of that is compute that, by the footnote's definition, is booked as R&D rather than cost of revenue.

The efficiency claim

Anthropic leans on compute efficiency as a selling point. It says Claude Opus 5.5 cut customer cost per task by about 40% compared with Opus 5, and it uses "compute multipliers", which it defines as "algorithmic and methodological improvements in training and inference that increase the effective output of a given amount of compute", as a running theme. The filing also says the company trains and serves Claude on four chip families (AWS Trainium, Google TPUs, NVIDIA GPUs and AMD GPUs) through a proprietary orchestration layer, and is moving from cloud contracts alone toward leasing accelerators directly in dedicated facilities, an arrangement it calls Anthropic Direct.

These are company claims in an offering document, not independent measurements. The revenue-versus-compute growth comparison is real arithmetic, but it says as much about prices, mix and the timing of capacity deliveries as about the efficiency of the models.

One more detail for anyone hoping to size the future bill: the paragraph on compute commitments signed through September 30 still has blanks where the numbers go. It reads "aggregate committed spend of more than $ , substantially all of which is related to compute and infrastructure," with part of it expected before the end of 2030 and "substantially all of the remainder" between 2031 and 2045. The preliminary filing tells us the commitments run for two decades. It doesn't yet say how large they are.