Anthropic’s Serving Costs Equal 25.6% of Its Compute Bill

Anthropic’s IPO filing puts the cost of serving customers at about a quarter of its compute bill, in line with the figure Altimeter Capital’s Brad Gerstner gave on stage at the TomorrowX Summit on 17 November. “Only 26% of their compute budget this year went to generating $100 billion in revenue”, Gerstner said in a clip of the session.

Our number: in the nine months to 30 September 2026, Anthropic’s cost of revenue, which the filing says is substantially all customer-serving compute, was $4,812 million against $18,808 million of compute and infrastructure expenses, equal to 25.6%, by our arithmetic. A year earlier the same ratio was 41.0%. It is an upper bound for compute spent on paying customers, because cost of revenue also includes some personnel and overhead costs. This is not investment advice.

Business Pill · Operating Leverage

An evergreen Business Engineer explainer on operating leverage: why profit can grow faster than sales once fixed costs are in place. It teaches the general idea only and does not discuss Anthropic.

The key insight: Anthropic’s filing puts cost of revenue, which it says is substantially all customer-serving compute, at 25.6% of its compute and infrastructure bill in the first nine months of 2026, by our arithmetic, close to the 26% on Brad Gerstner’s slide; most of the rest sits in research and development, which cost four times as much.

What Gerstner Said

Gerstner said Anthropic “had two gigawatts of compute this year, and only 25% of it went to serving this all up”. He put that at 500 megawatts and said “The vast majority of it goes to training”.

He then turned to next year: “the consensus forecast is another $100 billion of revenue. They only need 500 megawatts”. In his words, “they just don’t need that much compute to hit their numbers for next year”, adding: “Of course, they want a lot more”.

The slide behind him was labelled “Altimeter’s estimated allocation” and split the budget into 26% for “Serving paying customers” and 74% for “Training, research, internal use and the free tier”. It said: “The 26% estimate uses cost of revenue as a proxy for paid inference”.

Altimeter is an Anthropic investor. Bloomberg reported on 17 November that it is among the existing backers planning to buy shares in the offering, as noted in our pricing report.

Cost of revenue as a share of compute and infrastructure expenses, nine months to 30 September: 25.6% in 2026
Cost of revenue as a share of compute and infrastructure expenses, nine months to 30 September: 25.6% in 2026 ($4,812M of $18,808M) vs 41.0% in 2025 ($1,796M of $4,381M), our arithmetic; an upper bound because cost of revenue also includes some personnel and overhead costs. Source: Anthropic S-1/A No. 3.

What the Filing Shows

Anthropic’s functional expense note puts nine-month cost of revenue at $4,812 million, research and development at $19,240 million and sales and marketing at $6,619 million. It says “Substantially all of cost of revenue consists of compute and infrastructure costs associated with customer-serving workloads”.

Research and development, the note says, consists primarily of “compute and infrastructure costs associated with model development, research, and internal usage workloads”, plus personnel and other costs. Sales and marketing includes “compute and infrastructure costs associated with providing free access to the Company’s intelligence service”.

Total compute and infrastructure expenses were $18,808 million, 57% of $33,242 million of revenue, against $4,381 million a year earlier. The filing says the increase “was driven by the delivery of committed compute throughout the first nine months of 2026 under our CSP arrangements”.

The Ratio, Two Ways

By our arithmetic, cost of revenue was 25.6% of compute and infrastructure expenses in the nine months to September, down from 41.0% ($1,796 million of $4,381 million) a year earlier. Research and development was 4.0 times cost of revenue.

Revenue was 6.9 times cost of revenue, against 1.4 times a year earlier ($2,545 million of revenue), by our arithmetic. That is the same picture as the 85.5% gross margin in our S-1 piece: serving got cheaper per dollar of sales while the training and research bill grew faster.

How Anthropic Frames It

The filing says Anthropic does not run separate budgets: “We deploy this 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”.

It describes a loop between the two uses: “compute deployed to model development yields more capable and efficient models. Those efficiency gains free compute to serve additional customer workloads”. The filing also says revenue grew 1,127% year over year in the third quarter while compute and infrastructure expenses grew 382%.

The Structural Read

As we read the filing, revenue per dollar of serving cost rose from 1.4 to 6.9 times in a year, so the paying side of the business needed a smaller share of a much larger compute bill.

As we read Gerstner’s remarks, his capacity figures turn that cost share into megawatts; the filing itself reports dollars, and dollar and capacity shares can differ.

Anthropic S-1/A No. 3, functional expense note

“Substantially all of cost of revenue consists of compute and infrastructure costs associated with customer-serving workloads”

Three Implications

FOR IPO BUYERS The filing shows serving costs at 25.6% of the compute bill by our arithmetic, against 41.0% a year earlier. Not investment advice.

FOR READERS OF GERSTNER’S SLIDE His 26% is labelled an estimate that uses cost of revenue as a proxy; the 500 megawatts and two gigawatts are his figures, not the filing’s.

FOR COMPUTE SUPPLIERS The filing says Anthropic plans compute on a total basis across workloads, not by workload.

The Business Engineer Lens

This story maps onto the Business Engineer framework The State of The Inference Economy.

The framework says: “Inference is a revenue engine: continuous, distributed across millions of applications and agents, incurring marginal cost with every query.”

As we read it, Anthropic’s filing shows that revenue engine running on about a quarter of the compute bill by cost; the framework, citing Deloitte, puts inference at about two-thirds of all AI compute industry-wide. The framework does not discuss Anthropic’s filing.

What Is Not Established

A text search of the full filing found no megawatt or gigawatt figures. The 500 megawatts and two gigawatts are Gerstner’s, and we cannot check them. Dollar shares and capacity shares can differ: the filing says short-term compute costs significantly more than committed compute.

Gerstner did not say which $100 billion figure he meant. The filing’s nine-month revenue is $33.2 billion; our run-rate piece covers the annualised figure. His slide cites a 2 November filing; we checked the ratio against the 17 November amendment. We did not contact Anthropic or Altimeter.

This Week In The AI Supercycle

AI Lost Its Central Banker, the Lab Oligopoly, and the Future of Open Source

This week’s issue: the Nvidia leadership change, read as the AI build-out losing its central banker: who now finances the GPU landlords, backstops the data centres and sponsors open models, and why two labs are becoming the bank.

Read the full issue →

Business Engineer Framework

Anthropic’s S-1 & The State of Frontier AI

The Business Engineer premium analysis of Anthropic’s S-1: how the business makes money and what it commits to spend.

Read the Map of AI →

The Bottom Line

Anthropic’s cost-based numbers are in line with the 26% on Gerstner’s slide: cost of revenue equalled 25.6% of its compute and infrastructure bill in the first nine months of 2026, by our arithmetic, down from 41.0% a year earlier. Research and development cost four times as much. The capacity figures are his. The IPO process is in our explainer. This is not investment advice.

97,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.

A note on sourcing. We read the summary financials, the compute and infrastructure discussion and the functional expense note in Anthropic’s S-1/A No. 3 (17 November 2026), and transcribed Gerstner’s remarks ourselves from the audio of a TomorrowX Summit clip posted on 18 November; the slide text is as displayed in that clip. Altimeter’s plans in the IPO are from Bloomberg’s 17 November report. The ratios are our arithmetic. This is not investment advice. We did not contact Anthropic or Altimeter.

Sources: Anthropic S-1/A No. 3, 17 Nov 2026 · Gerstner at TomorrowX, clip posted 18 Nov 2026

Scroll to Top
FourWeekMBA