Vibeleaderboard
← All Intel
Intel / article

Anthropic Openai Compute Spend Growth

Source
epoch.ai
Author
Epoch AI publications
Date

Context

Anthropic spent $9.86 billion on compute and infrastructure in the third quarter of 2026, according to the amended IPO filing it submitted to US regulators, an S-1. Multiplied by four, that is a yearly pace of about $39 billion, against $433 million for all of 2023. Epoch AI, a research group that tracks AI inputs, fitted a growth curve to those disclosures and finds Anthropic's compute bill has grown about 4x a year, faster than the roughly 3x a year it estimates for OpenAI.

The two halves of that comparison rest on different evidence. Anthropic's numbers come from its own regulatory filing: annual totals for 2023 and 2024 and seven quarters from early 2025 through September 2026. OpenAI's come from media reports: $5.8 billion for 2024, taken from a projection The Information reported in October 2024, and $16.3 billion for 2025, also from The Information, plus a forecast rather than a result for this year: OpenAI president Greg Brockman testified in May that the company expects to spend about $50 billion on compute in 2026. Epoch says the OpenAI figures carry more uncertainty, both in accuracy and in what they include. It also warns that dollars do not convert neatly into chips or power, because the price of rented compute and the mix of hardware change over time. Both companies rent most of their capacity from cloud providers such as Amazon, Google, Microsoft, Oracle, CoreWeave and SpaceX, so the figures are mostly operating costs rather than spending on data centers they own.

Most of Anthropic's compute is not spent serving customers. The S-1 reports the cost of answering paying customers' requests as cost of revenue, and in the first nine months of 2026 that was $4.8 billion of $18.8 billion in compute spending, or 26%, down from 38% in 2025. Epoch says most of the rest goes to research and other internal uses; compute for free-tier users is booked under sales and marketing and is small. Epoch notes that media reports put OpenAI's share of compute spent on customer at about half in 2025.

The 4x pace may not continue. The S-1 lists about $447.5 billion in commitments, substantially all for compute and infrastructure, expected to be incurred between now and the end of 2030. Epoch puts that at around $100 billion a year, roughly one more doubling from the current rate, and says it is not clear Anthropic's spending will keep growing at anywhere near 4x a year, though the company notes it may spend more on compute before 2030 than it has committed so far. OpenAI, by comparison, plans to spend around $750 billion on compute over the same period.

Key takeaways · AI-distilled
  • Per Epoch AI's reading of Anthropic's S-1, Anthropic spent about $9.9B on compute and infrastructure in Q3 2026, roughly $40B annualized, up from $433M for all of 2023. OpenAI's figures come from media reports, not filings, so they carry more uncertainty.
  • Most of Anthropic's compute goes to internal use rather than customers: cost of revenue, substantially all inference compute, was 38% of 2025 compute spend and 26% in the first nine months of 2026. Reports put OpenAI's customer inference share near 50% in 2025.
  • Anthropic paid SpaceX about $5B for compute in Q3 2026, roughly half its compute spend that quarter and 34.5% of SpaceX's total revenue, according to Epoch AI.
  • Anthropic's S-1 lists $447.5B in compute commitments from September 2026 through 2030, about $100B a year. Epoch notes that is roughly one more doubling from today's rate, so it is unclear whether 4x annual growth continues, though Anthropic says it may spend more.
  • Epoch cautions that dollar spend does not convert directly into FLOPs or power, since the ratio depends on chip mix, cloud deal terms and rental pricing. OpenAI plans about $750B of compute spend through 2030.
Terms in this piece · Glossary
  • inference — Running a trained model to get answers — the phase where AI is actually used, as opposed to trained.
More from Epoch AI publications
Recommended reads
Comments

Checking sign-in…

Loading comments…