Anthropic’s amended S-1/A-2 filed November 10, 2026 reports $33.2 billion of revenue for the nine months ended September 30, 2026 and says substantially all of $4.812 billion of cost of revenue is compute for customer-serving workloads. Simple arithmetic on those two lines implies a ~85.5% gross margin on that narrow cost bucket. Headlines this week shortened that to “Anthropic has 85% margins.” The filing is more precise: training, research, internal use, and free-tier compute sit outside cost of revenue, mostly in research and development and sales and marketing.
For people who pay for Claude tokens, the distinction matters. Serving margin tells you whether inference is priced above the incremental GPU bill for paying API traffic. Business margin still includes the cost of building the next Opus and subsidizing free users. Anthropic’s own text warns that operating margin will fluctuate when large new compute blocks arrive under fixed commitments.
TL;DR
| Measure (nine months to Sept 30, 2026) | Revenue | Cost bucket | Implied margin |
|---|---|---|---|
| Gross margin (filing arithmetic) | $33.242B | Cost of revenue $4.812B | ~85.5% |
| Same period 2025 | $2.545B | Cost of revenue $1.796B | ~29.4% |
| All compute in ops statement | $33.242B | Total compute and infrastructure $18.808B | ~43.4% (our arithmetic, not a labeled ratio in the filing) |
| GAAP operating result | — | Operating loss $1.3B | Includes non-serving spend |
| Adjusted operating income (non-GAAP) | — | $4.1B positive | See prospectus reconciliation |
Where the 85% number comes from
Ledger-style illustration of gross-margin arithmetic on Anthropic S-1 revenue and cost of revenue lines
Anthropic changed expense presentation in 2025 from a functional layout with gross profit on the face of the income statement to a nature-of-expense layout. Note 8 to the consolidated financial statements groups operating expenses by function and is where gross-margin arithmetic lives.
For nine months ended September 30, 2026, that table shows:
- Revenue: $33.242 billion (also stated as $33.2 billion in narrative rounding).
- Cost of revenue: $4.812 billion — described as substantially all customer-serving compute and infrastructure.
- Research and development: $19.240 billion — primarily model development, research, and internal usage compute.
- Sales and marketing: $6.619 billion — includes compute for free access to Claude.
- General and administrative: $3.831 billion.
Gross margin in investor language usually means (revenue − cost of revenue) / revenue. On those nine-month lines, that is about 85.5%.
For the prior-year nine months, revenue was $2.545 billion and cost of revenue $1.796 billion, about 29.4% by the same math. The step-change is both scale and classification, not a single quarter of “flipping profitable on GPUs.”
explainx.ai covered the revenue scale and commitment stack in the amended S-1 revenue post and the IPO roadshow terms. This post isolates margin anatomy.
Serving margin is not the whole machine bill
The statement of operations also reports total compute and infrastructure expenses of $18.808 billion for the nine months. That line includes training and other workloads not parked in cost of revenue.
If you subtract all compute from revenue as a rough “fully loaded compute burden” check, you get about 43.4% of revenue left — still positive, but nowhere near 85%. Anthropic does not label that 43.4% figure; it is explainx.ai arithmetic for comparison only.
The filing’s narrative makes the economic point in plain language: Anthropic manages compute as a fungible pool and shifts capacity between customer-serving, training, and research workloads. Gross margin answers “what do paying customers fund per incremental inference dollar?” Total compute answers “what does staying at the frontier cost regardless of who pays?”
That split shows up when you compare Anthropic to Chinese labs on the same headline. Coverage such as The Inference’s Z.ai comparison stresses revenue per dollar of compute and procurement timing, not just API list prices. Our China token-economy explainer covers how rental compute shows up differently in other filings.
Quarterly texture: Q3 compute delivery
The amendment adds quarterly color that headline 85% figures hide.
For three months ended September 30, 2026, Anthropic reports:
- Revenue $17.0 billion for the quarter (narrative in the filing).
- Operating loss $567 million versus operating income $974 million in the prior quarter — driven in part by $5.0 billion of compute under short-term commitments that cost more than multi-year deals, plus $2.0 billion of non-cash charges tied to a donation-matching program, per the filing text.
- Adjusted operating income $2.5 billion in Q3 versus $2.2 billion in Q2, while adjusted operating margin fell from 19% to 15% because of the short-term compute block.
Compute and infrastructure rose each quarter in the series, reaching $9.9 billion in Q3 2026. As a percent of revenue, that expense was 48% in Q2 2026 and 58% in Q3 2026 after the delivery spike — still below revenue, but climbing.
Builders should read that as capacity waves, not as “margin only goes up.” When committed GPUs land, near-term costs are relatively fixed while revenue catches up. The filing explicitly expects margin fluctuation when incremental compute arrives.
What outlets add beyond the SEC file
The Information on November 6 — summarized by outlets such as FourWeekMBA — reported that gross margins around 85% in the first nine months of 2026, excluding model-development compute, may leave Anthropic room to lower prices if it needs demand as compute gets costlier. That is analyst speculation, not a sentence in the prospectus.
Contrast with GAAP operating loss of $1.3 billion over nine months and the much larger net loss lines driven by fair-value accounting on convertible instruments, which we explained in the Q3 S-1 revenue post. IPO readers and API customers are looking at different loss definitions.
What this means if you build on Claude
Compute and data-center capacity behind Anthropic serving-margin versus training spend
-
Do not treat 85% as “inference is trivially cheap.” Training and free-tier spend are real and recurring while models keep moving. The filing books much of that in R&D and sales, not in cost of revenue.
-
Serving margin can coexist with aggressive list prices when demand outruns committed cheap capacity — exactly the short-term compute premium narrative in Q3.
-
Price cuts are not confirmed. If Anthropic needs share against OpenAI’s reported Q3 revenue rebound or enterprise Claude adoption stats, margin headroom is one lever. Watch Claude pricing pages and changelogs, not S-1 tables, for actual changes.
-
Efficiency still matters. High serving margin does not remove the need for prompt caching, routing smaller models, and batching. Your bill is list price times tokens, not Anthropic’s internal gross margin.
-
IPO noise ≠ product roadmap. The same week brought Australian retail IPO access and RSP threshold updates. None of those replace a model card or API release note.
Reading gross margin next to OpenAI’s investor briefing
The same week brought leaked OpenAI investor documents quoting ~$10 billion of Q3 revenue and adjusted operating margins still deep in negative territory (−109% in WSJ relays), while GAAP operating losses reportedly widened to $10.9 billion. OpenAI has not filed a public S-1 with comparable line items, so you cannot line up an apples-to-apples cost of revenue table.
The comparison that is fair is directional:
- Anthropic now shows, in a public SEC file, that paying customer inference can carry a high gross margin even while training and subsidies absorb tens of billions elsewhere in the P&L.
- OpenAI investor slides, as reported, still portray a business spending more than twice revenue on an adjusted operating basis in Q3, with revenue growth re-accelerating.
Neither disclosure changes your API invoice on November 11. Both explain why compute procurement and routing efficiency stay central in output-token inflation debates and in Menlo’s enterprise spend survey, which already had Anthropic leading API share among surveyed companies.
IPO readers versus API customers
Securities analysts use gross margin to ask whether a software-like business can fund Capex-like compute without perpetual equity raises. Builders use price per million tokens and error rate per dollar. The S-1/A-2 is written for the first audience.
Three sentences in the amendment bridge the two:
- “We manage compute as a single, fungible pool” — your traffic competes internally with training runs for the same GPUs.
- Short-term compute can cost more than multi-year blocks — Q3’s $5.0 billion spike is the filing’s example.
- Adjusted operating income is positive over nine months while GAAP operating loss remains — know which metric a headline cites.
If Anthropic lists as ANT, public-quarterly reports may make serving margin easier to track. Until then, treat every 85% headline as cost-of-revenue margin, not as “Claude is 85% profitable.”
Z.ai, MiniMax, and why cross-country margin tables mislead
Outlet comparisons pit Anthropic’s ~85% against Z.ai ~26% or MiniMax ~18% on first-half 2026 disclosures compiled by trade press. Those labs book compute differently, rent more capacity spot-style, and price aggressively in domestic markets. Our Tencent Hy4 launch post shows how Chinese vendors publish their own benchmark tables against Kimi K3.1 Pro; margin math is another dimension where vendor accounting dominates.
Use cross-border margin tables to understand investor narratives, not to pick a coding model. Run your evals on your codebase, on the host your security team approved — whether that is Bedrock with Kimi, direct Claude API, or a self-hosted open weight.
How to verify the numbers yourself
- Open Anthropic S-1/A-2 on SEC EDGAR.
- Search for “cost of revenue” and read the footnote that substantially all of it is customer-serving compute.
- Open the supplemental table grouping operating expenses by function for nine-month periods.
- Read “Non-GAAP Financial Measures” before quoting adjusted operating income.
- Compare to prior explainx.ai coverage: draft prospectus H1 revenue, compute commitment caveats, internal Claude use vs payroll.