xAI shifts to datacenter REIT with $1.25B/month Anthropic GPU deal
Anthropic's capacity crisis solved by xAI's Colossus datacenter leasing.
xAI is pivoting from frontier AI lab to datacenter REIT, leasing huge GPU capacity to competitors. In May, Anthropic signed a deal for 300MW (220k GPUs) from xAI's Colossus datacenter at $1.25B/month, immediately lifting peak-hour restrictions on Claude subscriptions. Google then locked in 110k GPUs for $920M/month. Both deals have cancellation clauses (90 days notice after lock-in), but if sustained 18 months, xAI recoups all capex and still has spare capacity. The move comes as Anthropic faced severe compute crunches, forcing rationing during overlapping European/US peak hours.
Critics note Elon Musk's OpenAI lawsuit and Google's SpaceX stake as potential motives. But the real driver is a global GPU shortage: even OpenAI's Stargate UAE datacenter is threatened by geopolitical risks. xAI's 122-day Colossus build dwarfs hyperscalers' multi-year timelines, giving it a massive execution advantage. This leaves Grok's own training and inference capacity reduced, but xAI's financial engineering ahead of the SpaceX IPO may outweigh model ambitions. For professionals, the takeaway is clear: compute leasing is becoming a core AI business model, and infrastructure speed matters more than model performance.
- Anthropic pays xAI $1.25B/month for 300MW of H100s (220k GPUs), ending peak-hour Claude restrictions.
- Google signed $920M/month for 110k GPUs, with both deals cancellable after lock-in with 90 days notice.
- xAI's Colossus datacenter was built in 122 days, far faster than hyperscaler projects, giving it a critical supply advantage.
Why It Matters
Compute leasing is becoming a core AI business model; infrastructure speed now rivals model innovation.