Anthropic Commits to Gigawatts of TPU Capacity from Google and Broadcom

Anthropic just dropped a big one. They’ve signed a new agreement with Google and Broadcom for multiple gigawatts of next-generation TPU capacity, expected to come online starting in 2027. This isn’t just a bump in compute—it’s a whole new level of infrastructure to feed Claude‘s growing appetite.

Krishna Rao, Anthropic’s CFO, called it “our most significant compute commitment to date.” He’s not wrong. The company is clearly betting that demand for Claude will keep exploding, and they need the hardware to back it up.

The numbers here are staggering. Anthropic’s run-rate revenue has hit $30 billion—up from about $9 billion at the end of 2025. That’s more than tripling in roughly 15 months. And the enterprise side is accelerating too: back in February, they had over 500 business customers each spending over $1 million annually. Today that number has doubled to over 1,000. That kind of growth doesn’t happen without serious compute behind it.

Most of this new compute will be built in the US, expanding on their November 2025 pledge to invest $50 billion in American computing infrastructure. That’s a smart political and logistical move, especially with all the chatter about AI sovereignty and supply chain resilience.

What I find interesting is Anthropic’s hardware strategy. They train and run Claude on AWS Trainium, Google TPUs, and NVIDIA GPUs. That’s not just vendor diversity for the sake of it—it lets them match workloads to the best chip for the job. Better performance, better resilience. And it keeps them from being locked into any single provider. Amazon remains their primary cloud and training partner (Project Rainier is still on), but Claude is the only frontier model available on all three major clouds: AWS Bedrock, Google Vertex AI, and Microsoft Azure Foundry. That’s a flex.

This partnership deepens their existing work with Google Cloud (which got more TPU capacity last October) and Broadcom. It’s a clear signal that Anthropic sees the compute race as a multi-player game, and they’re not afraid to write big checks to stay ahead.

I’ll be watching to see how this plays out. The 2027 timeline means we’re looking at a couple of years before this capacity is live. In AI, that’s an eternity. But if demand keeps growing at this pace, they’ll need every watt.