Anthropic’s research and development spending was 4.0 times its cost of revenue in the first nine months of 2026, up from 2.0 times a year earlier, by our arithmetic from the amended S-1 it filed on 2 November. The filing says cost of revenue is substantially all compute for customer-serving workloads.
A supplemental table in the filing’s notes puts research and development at $19,240 million for January to September 2026 and cost of revenue at $4,812 million. Split per $5 of the two lines, that is $4.00 for research and development and $1.00 for serving customers, by our arithmetic.
Business Pill · Gross Margin
A short Business Engineer explainer on gross margin and cost of revenue. It teaches the general idea only and does not discuss Anthropic’s filing.
The key insight: By our arithmetic from the S-1/A’s functional table, Anthropic’s research and development line was 4.0 times its cost of revenue in January to September 2026 ($19,240 million against $4,812 million), up from 2.0 times a year earlier and 2.3 times for 2025.
The S-1/A reports operating expenses by their nature, with one line for compute and infrastructure. Note 8 regroups the same costs by function: cost of revenue, research and development, sales and marketing, and general and administrative.
The footnotes say what sits in each. Cost of revenue: “Substantially all of cost of revenue consists of compute and infrastructure costs associated with customer-serving workloads”. Research and development “primarily consist of compute and infrastructure costs associated with model development, research, and internal usage workloads, personnel-related costs, software expenses, and allocated overhead costs”.
Sales and marketing also carries compute, the filing says, for “providing free access to the Company’s intelligence service”. So the two lines compared here are not pure compute: research and development also holds personnel-related costs, software expenses and allocated overhead costs.
Research and development vs cost of revenue, split per $5 of the two lines combined, for 2024, January-September 2025, 2025 and January-September 2026. Source: Anthropic S-1/A filed 2 November 2026, operating expenses by function (Note 8). The dollar splits and ratios are our arithmetic. Cost of revenue is substantially all customer-serving compute; research and development holds model-development, research and internal-use compute plus personnel and other costs.
What Was Posted on X
For every $5 Anthropic spends on AI:
$4 goes to building new AI models $1 goes to running Claude for customers
Wall St Engine’s post, as framed by the poster: for every $5 Anthropic spends on AI, $4 goes to building new AI models and $1 to running Claude for customers. It quotes a Tradr ETFs post and chart of Anthropic’s 2025 costs from the S-1/A. By our arithmetic, the 4-to-1 matches January to September 2026; the 2025 figures the quoted post cites give 2.3 to 1.
How the Ratio Moved
Research and development was 16.2 times cost of revenue in 2023 and 3.0 times in 2024, by our arithmetic from the table. It fell to 2.0 times in the first nine months of 2025 and was 2.3 times for the full year, with the fourth quarter alone at 2.8 times, derived by subtracting the nine-month figures from the full year.
In the first nine months of 2026 it reached 4.0 times. Research and development grew 5.4 times year on year, while cost of revenue grew 2.7 times and revenue 13.1 times, by our arithmetic. Research and development fell to 58% of revenue, from 138% in 2025.
The two lines together came to $24,052 million in the nine months, 69.7% of total operating expenses of $34,502 million, by our arithmetic. The rest is sales and marketing at $6,619 million and general and administrative at $3,831 million.
What the Post Said
A post by Wall St Engine on X on Monday framed the split as “For every $5 Anthropic spends on AI: $4 goes to building new AI models $1 goes to running Claude for customers”. It quotes a Tradr ETFs post that cites $2.8 billion for serving customers and $6.3 billion for research and development.
The 4-to-1 matches the filing’s first nine months of 2026, by our arithmetic. The $6.3 billion and $2.8 billion are the full-year 2025 figures ($6,338 million and $2,773 million), which work out to 2.3 to 1, or $3.48 and $1.52 per $5. The quoted post’s $1.60 of compute per revenue dollar in 2025 matches the filing’s 160%, and its 57% for 2026 matches the nine-month figure.
Compute Alone
The filing does not split compute dollars by workload. It says Anthropic deploys committed compute “as a single, shared resource across all workloads, and thus plan, incur, and assess it on a total basis rather than by individual workload”.
Total compute and infrastructure expenses were $18,808 million in the nine months. Treating cost of revenue as all compute, the rest is about $14.0 billion, which also covers free-tier compute booked in sales and marketing. On that basis, compute for model development, research and internal use was at most about 2.9 times serving compute, against at most about 1.6 times in 2025, by our arithmetic.
As we read the table, the line that holds model-building compute grew about twice as fast as the line that holds customer-serving compute: 5.4 times against 2.7 times year on year, by our arithmetic. Revenue grew faster than both, so each line fell as a share of revenue while the gap between them widened.
Anthropic, S-1/A filed 2 November 2026
“We operate an iterative cycle in which compute deployed to model development yields more capable and efficient models. Those efficiency gains free compute to serve additional customer workloads”
Three Implications
FOR READING THE IPO FILING The filing’s own functional table, not the single compute line, is where the serving-versus-building split shows; the filing says cost of revenue is substantially all customer-serving compute.
FOR THE 4-TO-1 FIGURE By our arithmetic it matches January to September 2026; the 2025 figures of $6,338 million and $2,773 million give 2.3 to 1.
FOR COMPUTE ALONE Treating cost of revenue as all compute, compute outside it was at most about 2.9 times serving compute in the nine months, by our arithmetic; the filing says it plans compute on a total basis.
The analysis says: “Training was a cost center: episodic, concentrated among roughly twenty frontier model labs, amortizable across trillions of tokens.”
As we read it, Anthropic’s research and development line is that cost centre on the income statement, and in the first nine months of 2026 it grew faster than the serving line. The analysis does not discuss this filing.
What Is Not Established
The filing does not say how much of research and development is compute and how much is personnel, software or overhead. It does not separate training from research or internal use of Claude, and it does not give the functional table by quarter. The free-tier compute inside sales and marketing is not quantified. The 2.9-times bound assumes cost of revenue is entirely compute; the filing says substantially all.
This Week In The AI Supercycle
The Great Convergence of Capital, Power, and Geopolitics
This week’s issue: for two years the AI build-out was priced as a supply story, and last week as a credit story. This week it was priced as a permission story, and the answer arrived as a cost, not a ban.
On Anthropic’s own functional table, the line that holds model-building compute was 4.0 times the line that holds customer-serving compute in the first nine months of 2026, up from 2.0 times a year earlier and 2.3 times for 2025, by our arithmetic.
95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. All figures come from Anthropic’s amended S-1 filed with the SEC on 2 November 2026, chiefly the operating-expense table by function in Note 8. Ratios, per-$5 splits, growth multiples and the compute bound are our arithmetic. The Wall St Engine and Tradr ETFs posts are quoted as published on X. Nothing here is investment advice.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
