Anthropic just committed $45 billion to rent compute from a company that didn’t exist before 2024. That fact alone is worth sitting with. According to TechCrunch, the deal is with Nscale, a British AI infrastructure startup founded in 2024 that has already signed agreements with Microsoft. Under the terms, Nscale will supply Anthropic with compute powered by Nvidia’s Vera Rubin chips, running out of a flagship data center in West Virginia, with capacity expected to go live in late 2027.
The Vera Rubin system is Nvidia’s most advanced chip architecture, combining six chips working together in a single system. So Anthropic isn’t just buying capacity, it’s buying the best available capacity, timed to coincide with what will likely be the most competitive period in frontier AI development.
This deal doesn’t stand alone. Over the past eight months, Anthropic has been on an aggressive infrastructure push that makes this $45 billion figure feel almost expected. Earlier this month, it signed a $10 billion, six-year agreement with Volta, a cloud startup founded in January of this year, sourcing compute from Norway. In July, there was a $5 billion deal with AMD. In May, a compute agreement with SpaceX, reportedly worth $1.25 billion per month, drawing from two SpaceX data centers. And in April, Anthropic expanded its partnership with Amazon for an additional 5 gigawatts of capacity, and separately extended relationships with Google and Broadcom.
The pattern is clear. Anthropic is not treating compute as a cost center. It’s treating it as a strategic asset, one that needs to be locked in years in advance because the supply chain for frontier AI chips is constrained and the competition for that supply is intense. OpenAI, Google, and Meta are all doing versions of the same thing. But Anthropic’s willingness to sign with very new, relatively unproven infrastructure companies like Nscale and Volta is notable. It suggests the company is prioritizing speed and capacity over vendor maturity.
For developers and founders building on top of Claude, this matters for a few reasons:
- More compute headroom means Anthropic can run larger, more capable models without the same tradeoffs on latency or availability
- Long-term supply agreements reduce the risk of capacity crunches that could affect API reliability
- The Vera Rubin chip architecture, once online in 2027, should offer meaningful performance improvements over current-generation hardware
The broader story here is that the frontier AI race has two fronts now: model quality and infrastructure scale. Anthropic is competing hard on both. Whether $45 billion commitments to year-old startups are a sign of strategic confidence or something closer to desperation to keep pace with OpenAI is a fair question. But the spending isn’t slowing down, and that tells you where Anthropic thinks the leverage actually is.
