The Claude maker is assembling a power-hungry empire of chips and data centers that would rival the electricity consumption of a small country.
Anthropic is racing to assemble roughly 10 gigawatts of compute capacity by the end of 2027, a figure that would represent a sevenfold increase from the approximately 1.4 gigawatts the company had online at the close of 2025.
The expansion is being built on a web of infrastructure deals spanning Google, Broadcom, AWS, AMD, and a handful of massive data center contracts across multiple continents.
The deal architecture
The latest piece of the puzzle landed in April 2026, when Anthropic announced a partnership with Google and Broadcom to secure approximately 3.5 GW of next-generation TPU capacity. That capacity is expected to begin rolling out starting in 2027.
Anthropic’s agreement with AWS includes provisions for up to 5 GW of new Trainium capacity, Amazon’s custom AI training chip. A separate collaboration with AMD, centered on MI450 GPUs, accounts for up to 2 GW more, with the initial 1 GW tranche expected to come online by the first half of 2027.
Anthropic signed a lease for a 2.16 GW data center campus in Australia, purpose-built for inference workloads and scheduled to launch in 2027. The company also struck a deal with Nscale for 460 megawatts of capacity under a contract reportedly worth $45 billion over six years, with operations beginning in late 2027.
The company’s compute footprint is projected to hit roughly 5 GW by the end of 2026 before reaching the 10 GW target twelve months later.
Revenue is keeping pace with the ambition
By April 2026, the company’s annualized revenue run-rate had surpassed $30 billion, a staggering jump from the roughly $9 billion run-rate reported at the end of 2025.
Anthropic now counts more than 1,000 customers each spending upwards of $1 million per year on its Claude AI models.
Projections suggest Anthropic’s annualized revenue could exceed $65 billion.
A multi-supplier bet in a single-supplier world
Anthropic is spreading its bets across Google TPUs, AWS Trainium, and AMD’s MI450. By planting a 2.16 GW campus in Australia alongside its North American infrastructure, Anthropic is positioning itself for data sovereignty requirements and latency optimization across different markets, as inference workloads are inherently more distributed than training.
