Anthropic sources 220K GPUs from Colossus

Anthropic said on May 6 that it had signed an agreement with SpaceX to use all of the compute capacity at Colossus 1, a data center operated with xAI that the company said provides more than 300 megawatts of power and over 220,000 Nvidia GPUs. Anthropic said the capacity would come online within a month and would be used to improve service for Claude Pro and Claude Max subscribers first. xAI separately confirmed the partnership and described Colossus 1 as one of the world’s largest AI supercomputers. ### Did Anthropic actually move off Amazon Web Services? Amazon remains Anthropic’s primary cloud and training partner, according to Anthropic’s Nov. 22, 2024 announcement expanding its AWS relationship. In a separate April 20, 2026 post, Anthropic said it had signed a new agreement with Amazon for up to 5 gigawatts of capacity for training and deploying Claude, including nearly 1 gigawatt of Trainium2 and Trainium3 capacity expected online by the end of 2026. (anthropic.com) Anthropic also said on April 10, 2026 that it was expanding work with Google Cloud and Broadcom for multiple gigawatts of U.S.-based compute. Anthropic said it trains and runs Claude on AWS Trainium, Google TPUs and Nvidia GPUs, and called platform diversity a source of resilience for customers. ### So what is new in the Colossus deal? SpaceX is the new named counterparty in the Colossus agreement, and Anthropic said it would use all of the compute capacity at Colossus 1 rather than rent a smaller slice. xAI’s May 6 statement said SpaceXAI had signed the agreement to provide Anthropic access to the facility for training, fine-tuning, inference and high-performance computing workloads. (anthropic.com 1) (anthropic.com 2) More than 300 megawatts is the number that makes the arrangement stand out. Data Center Dynamics, citing the companies’ statements, reported that Anthropic had effectively taken the full Colossus 1 footprint and said the company had also expressed interest in multiple gigawatts of orbital AI compute capacity with SpaceX. Anthropic’s own post included that same statement of interest. (anthropic.com) ### Why does the 300-megawatt figure matter? Three hundred megawatts is not a normal cloud reservation. In data center terms, that figure points to a dedicated power envelope, racks, networking and cooling capacity that must be physically provisioned at a site rather than spun up as ordinary public-cloud instances. AWS says generative AI infrastructure requires accelerated computing, high-performance storage and networking systems designed to keep accelerators highly utilized. (datacenterdynamics.com) More than 220,000 GPUs also suggests Anthropic is buying access to a facility-scale cluster, not just negotiating for overflow capacity. xAI said Colossus 1 includes dense deployments of H100 and H200 GPUs, while Anthropic said the added compute would directly increase available capacity for paying Claude users. ### Is Anthropic the only lab doing this? Anthropic is not relying on a single supplier. (aws.amazon.com) Google Cloud said in October 2025 that Anthropic planned to expand its use of Google technologies, including up to 1 million TPUs, with well over a gigawatt of capacity expected online in 2026. Amazon and Anthropic then announced the separate 5-gigawatt expansion in April 2026. (x.ai) Data Center Dynamics reported on May 11 that Anthropic had also signed a $1.8 billion cloud contract with Akamai, adding to agreements with Google, AWS, CoreWeave and now SpaceX-xAI. That mix shows Anthropic spreading workloads across dedicated and cloud-style supply rather than naming one exclusive provider. (anthropic.com) ### What should readers watch next? Ross Nordeen, an xAI co-founder, said on May 7 that he was joining Anthropic “to focus on compute,” one day after the Colossus deal was disclosed. Data Center Dynamics reported the move as Anthropic accelerated its data center build-out after rate-limit issues and outages. (datacenterdynamics.com) By the end of 2026, Amazon and Anthropic expect nearly 1 gigawatt of Trainium2 and Trainium3 capacity to be online under their latest agreement, while Anthropic said the Colossus 1 capacity would arrive within a month of May 6. Those are the next concrete milestones to watch across Anthropic’s compute build-out. (anthropic.com) (datacenterdynamics.com)