xAI monetizes Colossus capacity

This shows xAI turning compute buildout into a two sided business, not just a cost center. When Colossus 1 was no longer the best cluster for xAI’s next training push, SpaceX leased that capacity to Anthropic instead of letting hundreds of megawatts sit partially idle. That means each Colossus generation can first train Grok, then be rented out as a giant external cloud, which helps finance the next build cycle.1 2 3 4

• The scale matters. Anthropic took access to more than 300 MW and over 220,000 Nvidia GPUs at Colossus 1 in May 2026, and SpaceX later disclosed agreements covering roughly 325,000 GPUs across Colossus and Colossus II at $1.25B per month through May 2029, with ramp pricing in May and June 2026.1 2 3 5
• This is closer to CoreWeave style infrastructure monetization than a normal AI lab model. xAI builds clusters for its own model roadmap, then sells excess or transitional capacity to another frontier lab, while Anthropic diversifies beyond AWS, Google, and Nvidia linked supply to lock in dedicated compute fast.3 4 6
• The deeper implication is architectural. Colossus 1 can be valuable even after xAI shifts frontier training toward newer Colossus generations, because an older cluster is still useful for inference, fine tuning, or overflow training. That extends asset life and reduces the economic penalty of building ahead of immediate internal demand.1 4 7

Going forward, the winners in frontier AI will look less like pure model labs and more like power merchants with software attached. If xAI can keep cycling each Colossus generation from internal use into third party leases, it creates a repeatable funding engine for larger clusters, and a lead in AI infrastructure that rivals without owned capacity cannot easily copy.2 3 4 6