Anthropic Buys Into SpaceX's Colossus to Break Free from Hyperscaler Compute Lock
Anthropic's full-capacity deal at SpaceX's Colossus 1 Memphis site signals a strategic push to diversify compute away from Microsoft- and Google-aligned infrastructure.
4. Anthropic Buys Into SpaceX's Colossus to Break Free from Hyperscaler Compute Lock
Anthropic announced on May 4, 2026 that it has signed a compute agreement with SpaceX granting full capacity access to the Colossus 1 data center in Memphis, Tennessee. The deal pairs with a separate announcement of higher usage limits for Claude across API tiers. Colossus 1, originally built to train xAI's Grok models, is one of the largest GPU clusters in the world, reportedly housing over 100,000 H100-class GPUs. Financial terms were not disclosed, but the arrangement gives Anthropic dedicated infrastructure outside its existing agreements with Google Cloud and AWS.
This is the more consequential detail. Anthropic's primary compute relationships run through Google, which has committed roughly $3 billion to the company, and Amazon, which has pledged up to $4 billion through AWS. Both hyperscalers are also direct competitors in the foundation model market via Gemini and Bedrock respectively. Securing independent capacity at Colossus 1 reduces Anthropic's exposure to infrastructure partners who have clear incentives to prioritize their own model families. For xAI, leasing unused Colossus capacity generates revenue from an asset that would otherwise sit idle between Grok training runs. The arrangement is transactional, not strategic alignment, but it hands Anthropic real negotiating leverage the next time Google or Amazon revisits pricing or priority access.
The broader pattern worth tracking: frontier AI labs are moving aggressively to own or contract dedicated compute rather than rely on shared cloud allocations. Meta builds its own clusters. OpenAI is co-investing in the Stargate infrastructure project. Anthropic, without a hyperscaler parent, has fewer options but is clearly assembling a patchwork of independent capacity. Whether Colossus access translates into faster Claude iteration cycles or just rate-limit headroom will show up in the release cadence over the next two quarters.
Source: Higher usage limits for Claude and a compute deal with SpaceX