Anthropic just locked in more compute than most countries use to power their cities. The company announced an expansion of its existing partnerships with Google and Broadcom, covering multiple gigawatts of next-generation compute capacity.
To put that in perspective: one gigawatt equals roughly the full output of a nuclear power plant. Modern AI data centers typically draw 100 to 500 megawatts each. "Multiple gigawatts" signals Anthropic is planning for infrastructure on a scale that would have seemed absurd two years ago.
Google has been a significant Anthropic backer since 2023, with cumulative investments that give it a stake in Anthropic's success alongside Amazon. The Broadcom angle is newer territory. Broadcom has been expanding aggressively into custom AI chips - specifically building application-specific processors (ASICs) for major AI companies who want an alternative to Nvidia's GPUs. Adding Broadcom to this compute agreement suggests Anthropic may be pursuing custom silicon rather than relying entirely on commodity GPU supply chains.
What This Means for Claude's Development
Anthropic's current Claude 3.7 models are computationally expensive to run, especially the extended "thinking" mode that reasons through problems step by step before answering. Training the next generation of models - and serving them to millions of users simultaneously - requires a steady, guaranteed supply of compute that spot purchases on the open market can't reliably provide.
This kind of infrastructure deal is how you secure the inputs for a model that doesn't exist yet. The lead time between signing a compute agreement and actually using that capacity is typically 18 to 36 months, given hardware manufacturing and data center construction timelines. Anthropic is effectively placing a bet now on where it needs to be by 2027 or 2028.
For Claude users, the practical implication is continued model development without the compute bottlenecks that have slowed some competitors. For the broader AI market, it's another signal that the handful of frontier AI labs are racing to lock up infrastructure before it becomes genuinely scarce.