On May 6, Anthropic announced a compute partnership with SpaceX that gives the company all of Colossus 1: more than 300 megawatts and 220,000 Nvidia GPUs, available within the month. For users, the deal translated immediately into doubled Claude Code 5-hour limits for Pro, Max, Team, and Enterprise tiers, removed peak-hour throttling for Pro and Max subscribers, and higher API rate limits for Opus models.
The coverage has mostly run it as a compute arms race story. That frame is not wrong. It is just not the interesting part.
The catch: According to Simon Willison, flagged by Semafor, the deal reportedly includes a clause giving Musk the right to withdraw compute capacity if Anthropic's AI causes "harm," with Musk defining the harm criteria. Neither Anthropic nor SpaceX has confirmed or released the contract language. If the clause is accurately described, Musk does not need to be reasonable about that judgment for the power imbalance to hold.
What Anthropic Actually Signed
Colossus 1 was built by xAI. SpaceX acquired xAI in February, which is how a data center built for a rival AI lab became Anthropic's nearest-term compute option. Anthropic's announcement does not explain the ownership transfer; Musk's company framed it separately as a compute partnership from xAI's side.
Anthropic's compute picture, as stated in its announcement: AWS up to 5 gigawatts (with about 1 GW online by end of 2026), Google and Broadcom at 5 GW beginning 2027, a $30 billion Azure capacity deal with Microsoft and Nvidia, and a $50 billion infrastructure investment from Fluidstack. The SpaceX deal fills the gap between today and when those longer-dated commitments come online. It is the bridge.






