Where the compute money goes
According to Anthropic's IPO filing (S-1), the company spent $8.95 billion on compute and infrastructure in the first half of 2026. Only $2.38 billion of that went to serving Claude to paying customers. Training, research and internal use took at least $5.93 billion, meaning at most 27% of compute went to customer service.
Revenue more than covers compute
Revenue reached $16.29 billion in the same period. The share kept after serving costs rose from 12% in the first half of last year to 85%. Once partner fees and marketing are included, that figure drops to 66%. Anthropic says Opus 5.5 cut cost per task by about 40% versus Opus 5, while Fable 5.1 cut the cost of typical workloads by about 25% versus Fable 5.
A massive capacity commitment
By Tom's Hardware's math, the company has ordered roughly $800 billion more in TPU and GPU capacity. A five-year TPU equipment lease with Broadcom is worth about $161.2 billion, and a deal with Google between April 2026 and July 2033 is worth at least $114.5 billion. Anthropic acknowledges that its five largest compute deals would cost an estimated $103 billion a year on average once delivered, more than twice its current revenue run rate.
A compute-constrained environment
Anthropic says it operates in an environment where customer demand exceeds available compute, so serving costs were muted by supply constraints rather than efficiency alone. The company runs a multi-chip architecture across AWS Trainium, Google TPUs, NVIDIA GPUs and AMD GPUs but does not disclose the split. The filing also warns that margins may compress as new hardware lands ahead of the revenue it will support.
