Anthropic, Google, and Broadcom are spending big on next-gen TPUs — here’s what that means

Anthropic just announced a massive expansion of its compute partnership with Google and Broadcom. We’re talking multiple gigawatts of next-generation TPU capacity, expected to start coming online in 2027. That’s not a typo — gigawatts. For context, that’s enough to power a small city, and it’s all going toward training and running Claude.

The numbers here are honestly 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 a year. And the number of business customers spending over $1 million annually? It doubled in less than two months, from 500 to over 1,000. I’ve seen a lot of growth curves in tech, but this one is steep.

Krishna Rao, Anthropic’s CFO, framed it as a disciplined approach to scaling — which is corporate speak for “we need a lot more compute, fast.” But I appreciate that they’re not just throwing money at any hardware. They’re diversifying across AWS Trainium, Google TPUs, and NVIDIA GPUs, matching workloads to the best chip for the job. That’s smart engineering, not just blind spending.

Most of this new compute will be built in the US, which builds on their earlier $50 billion commitment to American infrastructure from November 2025. It’s easy to be cynical about these “American jobs” announcements, but given the geopolitical stakes around AI chips, it’s hard to argue against domestic capacity.

One thing that stands out: Amazon remains their primary cloud provider and training partner, with Project Rainier still in the works. But Claude is the only frontier AI model available on all three major clouds — AWS Bedrock, Google Vertex AI, and Microsoft Azure Foundry. That’s a strategic advantage that most competitors can’t match. OpenAI is basically Azure-only. Google’s models are mostly on their own cloud. Anthropic is playing the field, and it’s working.

The big question, of course, is whether all this compute will actually translate into better models. More TPUs don’t automatically mean smarter AI. But given the demand they’re seeing — and the revenue to back it up — it’s a bet worth making. I’ll be watching to see if Claude’s next major version actually justifies this kind of infrastructure spend.