Colorful shipping containers line up on railway tracks in a logistics center viewed from above
Compute & energyIndependently verified

Anthropic spent $9.9 billion on compute in one quarter, 58% of its revenue, and the share is falling

By Yash Malviya·

Photo: Tobi &Chris / Pexels

Anthropic's S-1/A shows compute and infrastructure costs of $9,863 million in the third quarter of 2026, which Epoch AI annualizes to $39 billion and calls a 4x yearly climb. Revenue is growing faster still.

Anthropic spent $9,863 million on compute and infrastructure in the three months to September 30, 2026. That figure, from the company's registration statement, is the cleanest public look at what it costs to run a frontier lab, and Epoch AI turned it into a trend line on November 18: close to $40 billion a year at the current pace, up from $433 million in 2023.

The twist inside the filing is that the bill is growing fast but revenue is growing faster. Compute is a shrinking share of what Anthropic earns, even as the absolute number keeps climbing.

The series, quarter by quarter

Anthropic's Form S-1/A, filed November 2, 2026, lists compute and infrastructure expense for seven quarters. It was $944 million in the first quarter of 2025, $1,390 million in the second, $2,047 million in the third and $2,945 million in the fourth. In 2026 it was $3,387 million, $5,558 million and then $9,863 million.

Epoch's data insight, written by Josh You, uses those numbers and says Anthropic's compute spending is increasing by about 4x per year. It reports full-year totals of $433 million for 2023 and $2,522 million for 2024. By comparison, Epoch says OpenAI appears to have roughly tripled its compute spending annually and forecasts similar growth for 2026, though data for OpenAI is sparser, drawn from press reports rather than a filing.

Revenue is outrunning the bill

The same filing lists revenue by quarter: $376 million in the first quarter of 2025, then $787 million, $1,382 million and $2,040 million. In 2026 it reports $4,731 million, $11,555 million and $16,956 million.

Set side by side, the ratio flips. In the third quarter of 2025, compute of $2,047 million was about 148% of revenue of $1,382 million. In the fourth quarter it was about 144%. By the first quarter of 2026 it was about 72%, by the second 48%, and in the third 58%. Over twelve months, revenue grew roughly twelvefold while compute grew about 4.8 times.

“Anthropic spent nearly $10 billion on compute and infrastructure in the third quarter of 2026, or close to $40 billion on an annualized basis, up from just $433 million in 2023.”

Josh You, Epoch AI, Data Insight, November 18, 2026

The third quarter also shows why the ratio is not a straight line down. Compute jumped by $4.3 billion from the second quarter, and the company reported an operating loss of $567 million after posting an operating profit of $974 million in the second. These are the company's own figures, and a single quarter is a thin base for a trend, but the direction in the filing is clear: growth is being bought with ever larger infrastructure commitments.

Why spending and capacity diverge

Epoch's companion insight explains part of what the dollars buy. Its Inference Performance Index estimates that the agentic inference capacity of AI chips sold grew 4.6x between the second quarter of 2025 and the second quarter of 2026, compared with 2.6x for peak 8-bit computing power. That is because memory and chip-to-chip links matter more than raw arithmetic for serving agents. Our explainer on agentic inference capacity walks through the method.

For a lab serving coding agents at scale, that gap means each dollar of new hardware may support more simultaneous work than the headline compute figure implies. It also means comparisons of spending across labs and years need a common yardstick.

The suppliers are now shareholders

Anthropic's November 20 closing release shows how tied this spending is to its suppliers. Seven companies, AMD, Broadcom, Google, Micron, Microsoft, NVIDIA and TSMC Partners, bought about $26.0 billion of non-voting Class N stock at $83.00 a share in a private placement alongside the IPO. Our report on the IPO close covers the structure, and our compute explainer covers what a training run costs.

Where the rest of the money goes

Compute is the largest line, not the only one. In the third quarter of 2026, total operating expenses were $17,523 million, so compute was about 56% of them. Sales, marketing and partnerships cost $2,361 million. Other operating and administrative costs were $3,588 million, up from $1,031 million in the second quarter. Personnel cost $1,711 million, down from $1,857 million in the second quarter.

Two things follow. First, the company's costs are overwhelmingly machines, not people: personnel was about 10% of operating expenses in the quarter. Second, the jump in other operating and administrative costs deserves a footnote when the company reports again, since it moved the operating result as much as compute did. The filing does not break that line down in the excerpt we reviewed, so we do not speculate on its causes.

Reading the trend with care

Epoch's 4x figure is a fit to a short, steep series, so it will not hold forever. The four quarters from the fourth quarter of 2025 to the third quarter of 2026 add up to roughly $21.8 billion of compute expense, against annualized figures that look larger because they scale a single quarter by four. Our IPO tracker keeps the running record of what the company has confirmed.

Our take

The filing turns a rumor-heavy subject into a ledger. Three lessons stand out. First, a $39 billion annual run rate is a fixed obligation, not a one-off, and the S-1/A itself warns that if revenue growth does not keep pace with compute-related payment obligations, the company may be unable to meet them. Second, the falling ratio is the real story of the year, because it shows demand has so far outrun the cost of serving it. Third, the OpenAI comparison deserves caution until OpenAI publishes comparable audited quarters. Watch the next quarterly report for whether the ratio keeps falling or rebounds as new training runs start.

Frequently asked questions

How much does Anthropic spend on compute?

Anthropic's Form S-1/A reports compute and infrastructure expense of $9,863 million for the quarter ended September 30, 2026. Epoch AI annualizes that to $39.45 billion, up from $433 million for full-year 2023.

How fast is Anthropic's compute spending growing?

Epoch AI says it is increasing by about 4x per year. The filing shows quarterly compute expense rising from $2,047 million in the third quarter of 2025 to $9,863 million in the third quarter of 2026, about 4.8 times.

Is Anthropic spending more on compute than OpenAI?

Epoch compares growth, not totals: Anthropic about 4x a year, OpenAI roughly 3x. It cautions that data for OpenAI is sparser, based on press reports, while Anthropic's comes from an SEC filing.

What share of Anthropic's revenue goes to compute?

In the third quarter of 2026, compute of $9,863 million was about 58% of $16,956 million revenue. A year earlier, compute was about 148% of revenue. The figures come from Anthropic's S-1/A and our arithmetic.

Is Anthropic profitable?

The S-1/A shows an operating profit of $974 million in the second quarter of 2026 and an operating loss of $567 million in the third. Quarterly results swing with compute spending, so one quarter is not a trend.

What is agentic inference capacity?

It is Epoch AI's estimate of how many simultaneous agent-style requests a set of chips can serve, measured with its Inference Performance Index. Epoch estimates it grew 4.6x from Q2 2025 to Q2 2026 while computing power grew 2.6x.

Sources

What each one is, and whose it is.

  1. 1

    Anthropic, PBC Form S-1/A, Amendment No. 1, U.S. Securities and Exchange Commission (November 2, 2026)

    FilingThe vendor’s own
  2. DatasetIndependent of the vendor
  3. DatasetIndependent of the vendor
  4. Press reportThe vendor’s own