Anthropic Is Paying Its Direct Rival $45 Billion for GPU Capacity

Buried inside SpaceX's own IPO S-1 filing was a detail that had nothing to do with rockets: Anthropic has signed a $45 billion compute deal with xAI, Elon Musk's AI company, for access to 220,000 GPUs running on xAI's Colossus and Colossus II data centers. The agreement runs through May 2029 and works out to roughly $1.25 billion a month. It also includes 90-day walk-away clauses that let either side exit on comparatively short notice.

Rivals renting to rivals

Anthropic and xAI compete directly for the same enterprise customers, the same research talent, and the same spot at the top of frontier-model leaderboards. That Anthropic would pay xAI, rather than route around it, says something about where the actual bottleneck in AI right now sits. It isn't model quality or product design. It's physical GPU capacity, and there isn't enough of it to go around even for companies with the balance sheets to buy their own.

Why scarcity beats rivalry

Building a data center at the scale of Colossus takes years and tens of billions of dollars: land, power contracts, and chip supply that's itself rationed by Nvidia's production schedule. If xAI has 220,000 GPUs sitting in a facility and Anthropic needs more capacity than its own infrastructure and cloud partnerships can supply, the fastest way to get it is to pay whoever already built it, competitor or not. Compute scarcity is turning direct rivals into landlords and tenants, a dynamic that would look strange in almost any other software category but makes straightforward sense in one where the physical infrastructure is the actual constraint.

What the 90-day clause is really telling you

A four-year, $45 billion contract with an exit window that short is a strange combination. Long-term deals usually exist to lock in pricing and guarantee supply for both sides; a 90-day walk-away undercuts a chunk of that certainty. It suggests that neither side is fully willing to bet on the other as a stable, years-long infrastructure partner, whether because of concerns about xAI's own financial position, Anthropic's ability to bring capacity elsewhere online, or simply because both companies want the flexibility to renegotiate as GPU pricing and availability keep shifting. A deal built to be exited quickly is a deal both sides expect might need to be.

The bigger pattern

This isn't an isolated arrangement. As frontier labs run out of room to build their own data centers fast enough, compute is turning into a shared, tradable resource that gets bought and sold across competitive lines, the way electricity or bandwidth does in other industries. Watching who signs compute deals with whom, and how quickly those companies can pull out, is becoming as useful a signal of who's actually capacity-constrained as any benchmark score.