Anthropic signs a 45 billion dollar compute deal as Amazon triples its Nvidia chip order
Anthropic agreed to rent AI computing power from Nscale in a six-year, 45 billion dollar deal running on Nvidia's Vera Rubin chips from a data center in West Virginia. Separately, Amazon ordered 2 million additional Nvidia GPUs, roughly tripling its earlier commitment. Nvidia also unveiled NVHBM, a custom high-bandwidth memory technology for its NVLink Fusion platform that it says can lift per-chip performance by up to 30 percent at rack scale.
Artificial Intelligence··Midday
The Anthropic-Nscale Deal
Anthropic agreed to rent AI computing power from British infrastructure company Nscale in a six-year, multi-billion dollar deal. The capacity will run on Nvidia's Vera Rubin chip system from Nscale's flagship data center in West Virginia and is expected to come online in late 2027. The agreement follows Anthropic's earlier major compute commitments with Volta, AMD, SpaceX, Amazon and Google/Broadcom over the past eight months as it builds infrastructure to compete with OpenAI.[1]
Amazon's Tripled Commitment
Separately, Amazon ordered 2 million additional Nvidia GPUs — including Blackwell Ultra, Rubin and Rubin Ultra chips plus Vera CPUs — roughly tripling its order of more than one million chips placed five months earlier. The new chips are scheduled for delivery to AWS data centers in 2027 and 2028. Nvidia CEO Jensen Huang cited the order as evidence of durable demand, noting that AI is generating profitable tokens and more compute could generate more.[2]
Nvidia's NVHBM Technology
Alongside the surge in orders, Nvidia introduced NVHBM, a custom high-bandwidth memory technology for its NVLink Fusion platform, which allows cloud providers to combine their own AI chips with Nvidia's infrastructure. Nvidia says NVHBM shrinks the memory interface by up to 67 percent, freeing up silicon for compute, and can lift per-chip performance by up to 30 percent at rack scale while reducing memory power use by 15 percent. The power savings could allow roughly 15,000 additional AI chips to fit into a one-gigawatt data center.[3]
