Anthropic paid close to $10 billion for compute and infrastructure in the third quarter of 2026, and the bill has quadrupled each year since 2023, according to an analysis by Epoch AI, the research group that tracks AI scaling. Epoch’s Josh You published the estimate on November 18. The raw figures come from Anthropic’s amended S-1, the IPO filing it submitted on November 2, so the Anthropic side rests on the company’s own disclosures.

The exact quarterly number is $9.863 billion. Multiply by four and the run rate lands near $40 billion a year. In 2023, Anthropic spent $433 million. Epoch fits a growth curve through the filing’s annual and quarterly points and arrives at the 4x annual pace.

OpenAI is the comparison, and the data there is weaker. Epoch estimates that OpenAI has grown its compute spending about 3x a year. Its figures come from press reports rather than financial statements: $5.8 billion in 2024 and $16.3 billion in 2025, both reported by The Information. For 2026, Epoch has no reported result, only a forecast. OpenAI president Greg Brockman testified in May that the company expects to spend about $50 billion this year. That would put OpenAI’s annual total ahead of Anthropic’s, even as Anthropic’s curve is steeper.

One landlord explains a lot of the recent jump. In the third quarter, roughly $5 billion of Anthropic’s compute payments went to SpaceX, about half of the period total. Epoch backs this with a second data point: Anthropic made up 34.5 percent of SpaceX’s revenue that quarter. Amazon, Google, Microsoft, CoreWeave, and Oracle also supply capacity to the two labs.

The filing also shows where the compute goes. Epoch reads the S-1’s “cost of revenue” line as customer inference. In 2025 that was $2.773 billion, or 38 percent of $7.326 billion in compute spend. Across the first nine months of 2026 it fell to 26 percent, or $4.812 billion out of $18.8 billion. Most of Anthropic’s compute budget therefore goes to research and other internal work, not to serving paying users. Epoch notes that media reports put OpenAI’s customer inference share near 50 percent in 2025, up from about a third in 2024. The two labs are spending on different mixes.

The growth rate may not hold. The S-1 discloses $447.5 billion in spending commitments from September 2026 through the end of 2030, substantially all of it compute and infrastructure. That is around $100 billion a year, roughly another doubling from today’s pace. Epoch says it is unclear that Anthropic’s spending will keep climbing at anything like 4x, although the company’s own filing warns it may spend more than the committed total. OpenAI, by Epoch’s account, plans around $750 billion over the same period, nearly two thirds more than Anthropic has locked in so far.

Epoch lists its own limits. The OpenAI numbers are reported, not audited. Dollars do not convert cleanly into chips, FLOPs, or megawatts, because fleet mix and rental pricing vary by deal. Epoch also does not adjust for inflation, and it cannot say how much of the spend is accelerator capacity versus ordinary cloud services.

The practical reading is about dependency. A customer of Anthropic is buying from a company that sends half of its compute budget to a single supplier. For any team building on Claude, supply terms between those two firms matter as much as the model roadmap, and the S-1’s $447.5 billion commitment is the document that will show how firmly they are tied.

Epoch AI data insight by Josh You, published November 18, 2026, using figures from Anthropic’s amended S-1 and press reports on OpenAI.