Anthropic’s S-1: Who Pays for the Next Compute Cycle After 90% Incremental Margins?

Anthropic filed its S-1 on October 26, 2026, proposing a Nasdaq listing under ANT. The number of shares and offering price range remain undetermined. First-half revenue was $16.286 billion; Q2 revenue reached $11.555 billion and GAAP operating income was $974 million. These figures allow a more useful analysis than an annualized run-rate: customer payments, service costs, research and distribution can now be examined together.^[1][2]

SemiAnalysis Weekly Ep.041 discussed roughly 90% incremental gross margin and 40% incremental EBIT. Saito’s subsequent post framed this as selling a dollar of service for ten cents of cost. This report reconstructs the economics independently from the filing. A third-party automatic transcript was used to locate the podcast discussion; the original audio was not checked word for word. This is financial analysis, not a translation of the episode.^[3][4][5]

Research cutoff: October 28, 2026, 10:20 a.m. New York / 10:20 p.m. Shanghai. Financial comparisons use the quarters and half-years ended June 30; September operating metrics are identified separately. The 2026 interim statements are unaudited. Blank fields in the preliminary prospectus are not filled with outside estimates.^[1]

Reconstructing 91.0%: the numerator is incremental profit

The face of the income statement groups expenses by their nature, including compute and personnel, rather than showing a conventional gross-profit line. Note 8 supplies a functional cost-of-revenue presentation. H1 2025 revenue of $1.163 billion less $1.019 billion of cost of revenue yields $144 million of gross profit. H1 2026 revenue of $16.286 billion less $2.380 billion of cost of revenue yields $13.906 billion. These gross-profit amounts and ratios are Atlas calculations, not a separately presented headline metric. (MD&A and p.F-39)^[1]

Incremental revenue was $15.123 billion, incremental cost of revenue $1.361 billion and incremental gross profit $13.762 billion. Dividing incremental gross profit by incremental revenue gives 91.0%. Dividing current-period gross profit by current-period revenue gives 85.4%. One measures the change between periods; the other measures the whole current business before operating expenses.

Ten cents of cost is therefore a shorthand for a particular historical revenue increment. It is not a marginal-cost guarantee for every model, customer or future dollar. It also excludes research, free-user access, distribution and corporate overhead. Price, mix, utilization and expense allocation can all change the ratio. Slower revenue growth against rising fixed capacity costs could sharply reduce incremental margins even while existing services remain gross-profitable.

What the 85.4% gross margin leaves out

Note 8 puts mainly customer-serving compute and infrastructure, plus some personnel and overhead, in cost of revenue. Model development, research and internal compute use sit in R&D. Channel-partner expenses, marketing staff and compute used to provide free access sit in sales and marketing. Training remains essential to product competitiveness even though it is excluded from the derived service gross margin. (p.F-39)^[1]

Subtract $9.741 billion of R&D, $3.808 billion of sales and marketing and $1.050 billion of general and administrative expenses from H1 gross profit of $13.906 billion. The result is a $693 million operating loss. R&D alone consumed 59.8% of revenue, with sales and marketing at 23.4% and G&A at 6.4%. Strong service profitability had not yet made the entire half-year GAAP-profitable.

Distribution matters. In the expense-by-nature presentation, sales, marketing and partnerships totaled $3.168 billion, up $3.048 billion, with $2.7 billion of the increase attributable to channel-partner fees. That $3.168 billion is not the same line as functional sales and marketing of $3.808 billion. The two are alternative groupings of total expenses and must not be added together. (p.149 and p.F-39)^[1]

Expansion within existing customers can reduce selling effort without eliminating the cost of distribution. Cloud marketplaces place Claude inside established procurement systems and charge for that reach. Sustainable profitability must be measured after distribution, research and overhead. Comparisons with other model companies first need consistent treatment of free access, training and channel fees; otherwise apparently comparable gross margins describe different economic boundaries.

A 38.7% incremental operating margin is not a 40% current margin

Q1 2026 revenue was $4.731 billion with an operating loss of $1.667 billion. Q2 revenue was $11.555 billion with operating income of $974 million. The $2.641 billion improvement in operating income divided by $6.824 billion of incremental revenue gives 38.7%. This is a reproducible calculation close to the podcast’s 40% reference, but it uses sequential quarters. The 91.0% calculation above uses year-on-year half-years. They cannot be assembled into one period’s income statement. (pp.157–161)^[1]

Q2’s actual GAAP operating margin was 8.4%; its adjusted operating margin was 19.3%, not 40%. Anthropic’s adjustment excludes stock-based compensation, non-cash charitable contributions and the Bartz settlement. In Q2, adding back $1.261 billion of stock-based compensation to $974 million of operating income produces $2.235 billion of adjusted operating income. Non-cash compensation still has an economic cost through equity and ongoing remuneration.

The disclosed quarters do not support an unqualified claim of two profitable quarters. Q1 had a $1.667 billion GAAP operating loss, a $617 million adjusted operating loss and a $310 million adjusted net loss. Q2 turned positive on the operating and adjusted-net measures. Undisclosed Q3 earnings cannot be inserted to complete the claim. (pp.160–161)^[1]

Conversely, H1’s $80.748 billion GAAP net loss is not equivalent to burning $80.748 billion of cash. Changes in the fair value of convertible notes, derivative liabilities and SAFEs generated $72.388 billion of expense, alongside $8.720 billion of other mark-to-market changes excluded in the company’s adjusted reconciliation. A higher company valuation can increase the accounting value of financing liabilities. Those movements should be separated from operating losses, without assuming that every adjustment is economically costless. (pp.156–157)^[1]

Revenue outrunning compute reflects timing as well as efficiency

Q2 revenue rose from $787 million to $11.555 billion year on year, an increase of 1,368.2%. Compute and infrastructure expense rose from $1.390 billion to $5.558 billion, up 299.9%. Revenue per dollar of that expense increased from $0.57 to $2.08. Starting instead in Q1 2025, the ratio rose from $0.40 to $2.08, or about 5.22 times its original level. This is an accounting revenue-to-expense ratio, not a measurement of GPU speed or token-level physical efficiency. (p.157)^[1]

The filing explains that much of the compute is contracted 12–18 months ahead under multiyear arrangements. Expense and payments arise as capacity becomes available, leaving the near-term cost base relatively fixed. Fast demand growth can therefore monetize previously purchased resources much more effectively. This is operating leverage, but the disclosed data do not isolate how much comes from hardware, engineering, utilization, pricing or product mix. (p.146)^[1]

The mechanism can reverse. When the next capacity tranche arrives, its expense may precede the customer revenue needed to absorb it. Anthropic explicitly warns that quarterly operating results can fluctuate with incremental compute delivery. Multiplying Q2 profits by four, or treating 38.7% as a steady-state margin, would ignore that delivery cycle.

A deliberately static sensitivity illustrates the exposure. Hold Q2 revenue and other operating expenses fixed and raise compute expense by 20%. The additional $1.1116 billion exceeds the quarter’s $974 million operating income, implying an operating loss of roughly $137.6 million. If incremental revenue could still convert at the historical 38.7% incremental operating margin, approximately $2.872 billion of additional revenue—24.9% of the baseline—would offset that cost shock. This combines explicit cost and contribution assumptions; it is not a forecast or a claim that new capacity produces no revenue.

900% retention measures expansion, not contracted future revenue

Anthropic reports approximately 900% net revenue retention as of September 30, 2026. Its definition compares the same qualified customer cohort’s current-quarter revenue with the corresponding quarter a year earlier, excluding Claude Pro and Max fixed-price subscriptions. It measures net expansion and contraction in that cohort, not a customer renewal rate. It does not cover every user and cannot simply be inserted into next year’s revenue forecast. (definition p.3 and MD&A p.135)^[1]

At the same date, more than 11,900 customers generated at least $100,000 of trailing-twelve-month revenue, more than 1,900 generated at least $1 million, and approximately 250 generated at least $10 million. These are nested thresholds, not additive populations. They show a growing high-spend base. Expansion beyond coding into general knowledge work still requires evidence on task frequency, willingness to pay and service cost; spending by the heaviest users is not a universal seat-price assumption.

Concentration requires both period and denominator discipline. The top five customers accounted for about 25% of revenue in the first nine months of 2026. Cloud-marketplace revenue represented 48% of H1 revenue. These describe customers versus channels over different periods; adding them does not establish 73% concentration. In Q3, customers outside the top five generated 79% of revenue, implying roughly 21% for the top five, versus approximately 36% in Q1. Anthropic also says revenue from each of its Q3 top five reached a record. A falling share can coexist with rising absolute spend when other customers grow faster. (pp.58 and 143)^[1]

Consumption pricing transmits expansion quickly, but it can also transmit budget reductions. Total customer spend in the two weeks ended January 2, 2026 fell approximately 20% against the preceding two weeks. This was not a full-month or quarterly revenue decline. It shows why working patterns and holidays matter to short-term run-rates. Durable demand should be evaluated through comparable cohorts, recognized revenue and collections rather than a single day annualized. (p.143)^[1]

Read compute commitments contract by contract

The aggregate obligations table on page 170 still contains blanks, while the following descriptions provide amounts, terms and exit conditions for individual agreements. The table below selects major arrangements to illustrate those differences; it is not a comprehensive total. Minimum purchases, maximum contract values, cancellable service agreements and finance-lease principal are not interchangeable. A dollar-per-gigawatt ratio from one project cannot reliably convert all these contracts into capacity. (pp.170–172)^[1]

Certain Microsoft, Google and Amazon agreements require minimum spending regardless of actual use, leaving utilization risk with Anthropic. Nscale and Lambda fees begin with delivery and acceptance, with specific financing, facility and delivery conditions. SpaceXAI offers greater exit flexibility, but only after initial non-cancellable periods and with 90 days’ written notice. That is materially different from an immediate, costless cancellation right.

Broadcom equipment leases change the asset and cash-flow structure. Aggregate five-year rent is approximately $161.2 billion excluding interest, generally starting when equipment is ready for service; some circumstances, including facility delays, can bring payments forward. Facilities, operations and financing create additional obligations. Equipment rent is not the entire data-center cost, and related financing availability must not be added again as incremental compute demand.

For infrastructure suppliers, the disclosures improve visibility into customer demand and payment conditions. Converting a contract into revenue still requires financing, construction, power connection, equipment delivery and acceptance. Exit rights and payment timing allocate risk among the model company, cloud provider, equipment lessor and equity holders. Better Anthropic margins can support a credit assessment without establishing attractive capital returns for every supplier.

A large cash balance is not the same as future self-funding

At June 30, 2026, cash and equivalents were $112.4 billion and short-term investments $19.3 billion, a combined $131.7 billion. Approximately $3.2 billion of available credit was separate, not cash. H1 operating cash inflow was $3.929 billion, while financing cash inflow was $110.774 billion. The liquidity cushion is substantial, but its main source remains capital raised rather than accumulated operating earnings. (pp.161 and 169)^[1]

H1 investing cash outflow of $17.181 billion primarily reflected purchases of short-term securities, alongside higher server and networking capital expenditure. Calling the entire amount CAPEX or operating burn would misclassify treasury investment. Equally, positive operating cash flow does not show that future equipment leases and advance payments are covered. Customer collections, vendor payments and working-capital timing also affect cash conversion. (p.169)^[1]

Anthropic expects greater investment in owned or leased infrastructure to require payments before capacity is operational. It warns that operating and investing activities may use significant cash even if financial performance improves. High service margins and continuing financing needs can therefore coexist. The relevant comparison is available liquidity versus payments over a matching horizon, including lease principal, interest and capital expenditure—not conditional financing promises versus a multiyear headline commitment. (p.162)^[1]

With offering price and share count undetermined, the filing does not yet support a reliable IPO valuation conclusion. High margins improve potential earnings power, but financing terms, stock-based compensation, convertible instruments and ongoing capital requirements also shape shareholder returns. This report does not assign a target price or rating to an undetermined offering valuation.^[2]

The next two quarters should test profit absorption and cash conversion

First, watch operating profit as new capacity arrives. If revenue absorbs the stepped-up compute base while still covering distribution and R&D, Q2 becomes more persuasive as a durable commercial turning point. If service gross margins stay high while operating profit turns negative, decompose capacity additions, research and distribution rather than using the 85% figure alone to declare the business a success or failure.

Second, examine the quality of customer expansion. A 900% retention rate benefits from an early low base; a decline need not mean the business is shrinking. Absolute spend by existing customers, new-customer contribution, concentration and collections should be examined together. As adoption moves beyond developers, the economics may not transfer if paid-task growth lags acquisition support and distribution costs.

Third, align cash obligations with delivery. Compare the next year’s minimum payments, acceptance milestones, lease principal and capital spending with operating cash generation. Financing or delivery delays affect suppliers’ revenue recognition; capacity arriving ahead of paid demand first pressures Anthropic’s earnings and cash. A complete subsequent quarterly report and amended S-1 will be more informative than another run-rate headline.

Atlas’s assessment is that the filing supports a high-gross-margin customer-serving business and a Q2 operating break-even milestone. That is substantial counterevidence to a blanket claim that model providers can only sell subsidized services. It does not show that research can stop, that every future revenue dollar costs ten cents, or that future cash generation already covers all compute commitments. Earnings growth and capital consumption must now be evaluated together.