Anthropic compute spending is growing about four times a year, while OpenAI’s is growing about three times, according to a new analysis. The estimate leans on figures from Anthropic’s S-1, the filing it submitted ahead of its IPO.
Epoch also finds that Anthropic spends a shrinking share of its compute on serving customers, and that its commitments through 2030 are far larger than what it spends today.
Where the numbers come from
For Anthropic, Epoch used the amended S-1 filed on November 2. It lists “compute and infrastructure” spending for seven quarters, from the start of 2025 through the third quarter of 2026, plus annual totals for 2023 and 2024.
OpenAI doesn’t publish comparable filings, so Epoch relies on media reports. That makes the OpenAI side less certain. In May 2026, OpenAI President Greg Brockman testified that the company expects to spend $50 billion on compute this year, but Epoch has no non-forecast figure for OpenAI’s 2026 spending.
To compare mixed annual and quarterly data, Epoch multiplies quarterly figures by four to get annualized run rates, then fits growth rates with a log-linear regression and reports 90% confidence intervals. It doesn’t adjust for inflation.
How much goes to serving customers
Most of Anthropic’s compute isn’t used to answer customer requests. It goes to research and other internal work. The S-1 reports customer inference as cost of revenue, which it says is “substantially all” compute.
- In 2025, cost of revenue was $2.773 billion, 38% of $7.326 billion in total compute spending.
- In the first nine months of 2026, the share fell to 26%: $4.812 billion of $18.8 billion.
Epoch notes cost of revenue slightly understates inference, because Anthropic books free-tier compute under sales and marketing. That line was $5.5 billion for the nine months. Of its $5.2 billion year-on-year increase, $4.8 billion came from revenue-sharing fees paid to cloud partners, so free-tier compute is at most in the hundreds of millions.
By comparison, media reports say about half of OpenAI’s compute went to customer inference in 2025, up from roughly one-third in 2024.
Who gets the money
Both companies rent most of their capacity from cloud partners, including Amazon, CoreWeave, Google, Microsoft, Oracle and SpaceX. Epoch says neither carries significant upfront compute spending, which makes comparing operating costs reasonable. Anthropic will use more TPUs it leases directly, but those lease payments are amortized.
One data point shows the scale: Anthropic accounted for 34.5% of SpaceX’s total revenue in the third quarter of 2026, around $5 billion. That is half of what Anthropic spent on compute that quarter.
What the $447.5 billion commitment implies
The S-1 discloses $447.5 billion in spending commitments from September 2026 through the end of 2030, “substantially all” tied to compute and infrastructure. That works out to about $100 billion a year. Epoch calls this roughly another doubling from the current rate, so it isn’t clear its compute spending will keep growing at 4x a year. The company says it may spend more than the committed total before 2030.
OpenAI plans to spend around $750 billion on compute over the same period, which is well above Anthropic’s committed figure.
Caveats Epoch lists
Dollars don’t convert cleanly into chips or power. The relationship shifts with fleet mix, deal terms and rental prices. Epoch also says it can’t split the money between accelerators, CPUs and general cloud services, though it assumes most of it supports AI research, training and inference.
Why the gap between 4x and 3x matters
A one-step difference in annual growth compounds fast. Over three years, 4x a year means spending multiplies roughly 64 times, while 3x a year means about 27 times. The two firms are running different bets on how quickly they need to scale capacity.
The split between research and customer serving also differs. Anthropic’s customer-facing share is falling as total spending rises, while OpenAI’s rose in 2025. That points to Anthropic putting more of its budget into training and internal work for now.
What to watch next
The next data will come from Anthropic’s first quarterly reports as a public company, which should show whether the 4x pace holds or flattens toward the $100 billion-a-year commitment. Epoch’s data explorer, which doesn’t yet include the S-1’s quarterly figures, is the place to track updates.
Source: Epoch AI