TL;DR

Anthropic has signed a new agreement with Google and Broadcom for multiple gigawatts of next-generation TPU capacity, due to come online from 2027. The lab says its run-rate revenue has reached around $30 billion (£24 billion), more than triple the roughly $9 billion (£7 billion) it was doing at the end of 2025.

Compute Pre-Booking Reaches a New Scale

CFO Krishna Rao called it Anthropic’s “most significant compute commitment to date”, framing the deal as a continuation of a deliberate scaling strategy rather than a one-off. The bulk of the new capacity will sit in the United States, building on the $50 billion American infrastructure pledge Anthropic made in November 2025. It also extends an October 2025 expansion of TPU capacity with Google Cloud.

The headline customer numbers tell a similar story: Anthropic now counts more than 1,000 enterprise accounts each spending over $1 million a year on its services, double the figure it disclosed at its Series G fundraise in February. That doubling has happened in under two months.

Hedging Across AWS, Google and Nvidia

Notably, the announcement keeps Anthropic’s three-way hardware split intact. The company still trains and serves Claude across AWS Trainium, Google TPUs and Nvidia GPUs, with AWS remaining its primary cloud and training partner via the Project Rainier build-out. Claude is also the only frontier model offered on all three of the largest cloud platforms — Bedrock, Vertex AI and Azure Foundry — a positioning advantage Anthropic is increasingly leaning on commercially.

Looking Forward

For UK enterprises evaluating frontier model vendors, the practical takeaway is supply confidence: Anthropic is stacking compute reservations years ahead of demand at a moment when more than half of US datacentres planned for 2026 are facing delays. The trade-off is concentration. Anthropic’s expansion is increasingly an American story, both in where its data centres sit and where its capital is being committed — leaving Britain’s positioning as a research, talent and policy hub rather than a compute one.