# Anthropic says most compute spending goes to research

> Anthropic spent under a third of its $8.9 billion first-half compute bill serving customers. Tom’s Hardware estimates its future chip commitments at about $800 billion.

Oossa · 2026-10-29 · https://oossa.com/en/anthropic-says-most-compute-spending-goes-to-research

Anthropic spent less than one-third of its $8.9 billion compute bill in the first half of the year serving customers, according to the company’s IPO filing. Most of the rest went to training and research, as well as running Claude for Anthropic’s own use. Compute is the processing power used to train and operate AI models.

The figures show why a company’s spending on AI chips is not the same as the cost of answering customer prompts. Anthropic’s filing separates customer-serving compute from research and internal use, but the supplied figures do not break down those other categories individually or state how much it costs to serve each Claude user.

## What are the chip commitments?

Anthropic has also signed contracts for hundreds of billions of dollars in additional capacity from Google’s TPUs, its AI chips, and GPUs, the processors widely used for AI work. Tom’s Hardware estimates the commitments at roughly $800 billion. That is the publication’s calculation, not a figure presented in the supplied text as Anthropic’s own total.

These are commitments to future computing capacity, not evidence that Anthropic has already paid the full amount or is using all of it. The filing details described here do not give a schedule for when the capacity will be delivered or how much will be used for customer requests rather than research.

## Revenue per compute dollar rose

Anthropic reported that revenue per dollar of compute spending increased from $0.40 in the first quarter of 2025 to $2.08 in the second quarter of 2026. The company links the increase to capacity it bought 12 to 18 months in advance. This ratio compares revenue with compute spending; it is not a measure of profit, and the figures do not show what other costs Anthropic had.

The filing comes as Anthropic prepares for an IPO, or initial public offering, in which a company sells shares to the public. Tom’s Hardware says rival OpenAI had filed confidentially, while CEO Sam Altman said OpenAI would not go public that year. The figures leave a central question open: whether AI companies can earn enough to cover the cost of running their models for customers.

## The facts

- Anthropic’s first-half compute bill was $8.9 billion, and less than one-third went to serving customers.
- The filing says most of the remaining compute went to training, research and running Claude internally.
- Tom’s Hardware estimates Anthropic’s additional TPU and GPU capacity commitments at roughly $800 billion.
- Revenue per compute dollar rose from $0.40 in Q1 2025 to $2.08 in Q2 2026.

## Why it matters

For Claude customers, the filing shows that most of Anthropic’s compute spending supports work beyond serving current requests. It does not say whether that spending will change Claude’s prices or availability, or whether revenue will ultimately cover the cost of running the models.

## Sources & references

1. [Anthropic’s IPO filing shows most of its compute goes to R&D, not serving customers — company details massive $800 billion in TPU and GPU contracts](https://www.tomshardware.com/tech-industry/artificial-intelligence/anthropics-ipo-filing-shows-most-of-its-compute-goes-to-r-and-d-not-serving-customers-company-details-massive-usd800-billion-in-tpu-and-gpu-contracts) – Tom's Hardware, 2026-10-29

Last updated: 2026-10-29