Anthropic’s IPO filing shows most of its compute goes to R&D, not serving customers

Anthropic spent less than a third of its $8.9 billion first-half compute bill on serving customers, with most of the rest going into training, research, and running Claude internally, its S-1 filing shows. It has meanwhile signed up for hundreds of billions of dollars more in TPU and GPU capacity. By our math, it has meanwhile signed up for roughly $800 billion more in TPU and GPU capacity, and revenue per compute dollar jumped from $0.40 in Q1 2025 to $2.08 in Q2 2026, which the company links to capacity purchased 12–18 months ahead.

Anthropic’s IPO is likely to beat that of rival OpenAI, which filed confidentially before CEO Sam Altman said its own won’t come this year, amid widespread doubts about an AI bubble. Whether AI labs can cover what it costs to serve their models is an open question.

Anthropic appears to do so easily: it spent $2.4 billion serving customers in the first half of the year, versus $16.3 billion in revenue. Paying for what it has on order is more difficult. Its five compute deals with the biggest stated amounts would cost an estimated $103 billion a year on average once delivered, more than twice its revenue at Q2’s pace of about $46 billion a year. The company also warns that margins may compress as new hardware lands ahead of the revenue it will support.

The company has three main costs: the cost of revenue, which is the inference cost of running Claude for paying customers; R&D, which is mostly compute for training and research into new models plus any internal use; and sales and marketing, which covers the cut taken by channel partners and cloud platform resellers.

Comparing the first half of last year to this year’s, cost of revenue is up about 2.3x, but with revenue up 14x, that means Anthropic keeps an estimated 85 cents for every dollar after serving costs, up from just 12 cents. R&D is up about 5 times, with 66% or more of the compute going to it, and that’s the main source of its operating loss. Sales and marketing is up 16x, and doesn’t show up in cost of revenue, but its partner fees and marketing would lower what it keeps to 66 cents a dollar. The trend here is that serving Claude now more than covers its costs.

Serving has kept getting cheaper for customers since June: in September, Anthropic’s Opus 5.5 model cut costs per task by about 40% versus Opus 5, says the company, and Fable 5.1 cut the cost of typical workloads by about 25% against Fable 5; each came from “a combination of a lower price and improved efficiency,” made “while preserving our margins.”

As for compute, which runs through all three costs, Anthropic runs it as “a single, shared resource across all workloads” and doesn’t disclose how it decides the split between training, research, and serving. Serving was limited, as the company has “consistently operated in a compute-constrained environment where customer demand exceeds available compute,” which suggests serving costs were muted by a lack of supply, not just efficiency.

Anthropic runs a “multi-chip compute architecture across AWS Trainium, Google TPUs, Nvidia GPUs, and AMD GPUs,” without giving any percentage splits or GW measurements. For TPUs, it has five-year equipment leases with Broadcom for “TPU-based compute capacity,” with rent of “approximately $161.2 billion excluding interest”; some of that capacity came online this year, with more from 2027. Another $114.5 billion or more is with Google “between April 2026 and July 2033,” with more from Google and other TPU providers from 2028 to 2031, estimated to be in the $125–$150 billion range.

Trainium comes through Amazon, with a deal for “at least $110.0 billion between May 2026 and April 2036,” though the filing doesn’t split AWS spend by chip. Nvidia GPUs come through a series of deals, including “up to $120.5 billion … extending into 2029” with SpaceXAI, “at least $44.9 billion” extending into 2033 with Microsoft, “up to approximately $44.6 billion” also into 2033 with Nscale, and about $41.3 billion into 2034 with Lambda. There is also an AMD hardware supply agreement which, in aggregate, may “exceed $20.0 billion” if Anthropic takes the full initial capacity. These deals each have their own conditions for termination.

Anthropic’s minimum purchase commitments, which exclude equipment leases such as Broadcom’s, have gone from about $55 billion at the end of last year to about $265.6 billion by the end of Q2 this year. Since then, it has signed about $245 billion in new minimums, leases included, and TPUs are the biggest line overall. The Nvidia capacity, Microsoft’s aside, is also the easiest for Anthropic to walk away from.

With all these multi-year commitments, bought well ahead of delivery, expenses are paid and recognized “as compute becomes available to us,” the company says. That compute base “did not fluctuate with in-period demand,” so revenue per compute dollar increased as demand surged. When demand outruns that capacity, Anthropic buys short-term compute, which “typically carries a higher cost per unit.”

Our $103 billion estimate is straight-line, spreading each deal evenly over its term except SpaceXAI, which is at its full monthly rate, while payments ramp up as capacity is delivered. Also not included in the estimate is interest on the Broadcom leases. Compute also does not necessarily have to earn revenue, especially in R&D.

The filing has many blanks, such as the total commitment figure, the contractual obligations table, Q3 revenue, and Run-Rate Revenue, which an amended S-1 would normally fill in before pricing. No date for that has been given. Microsoft’s minimum starts in November, and SpaceXAI’s full delivery is expected in December, with non-cancelable periods ending in January and March 2027, after which either side can cancel on 90 days’ notice. More Broadcom TPU capacity and AMD’s first tranche follow next year, by which point most of these deals will be billing.