AI infrastructure is starting to look less like cloud procurement. It looks like capacity diplomacy.
Google paying SpaceX about $920M a month for access to roughly 110,000 NVIDIA GPUs changes how I read AI infrastructure: the critical object is not only compute, but the contract around compute.
In software, we talk about infrastructure as if the hard part is the API. In regulated and high-dependency systems, the harder questions are usually contractual: who has priority, who can cancel, what happens if capacity is late, and which workloads get degraded first.
That is what makes this deal interesting. A 90-day cancellation clause, a delivery dependency on promised GPUs, and bridge capacity for surging Gemini demand say more about the AI market than another benchmark. Compute is becoming something companies lease under strategic uncertainty, not consume from an infinite cloud shelf.
The architecture is only half the story. The capacity agreement is part of the system.
The model is visible. The cancellation clause is where the dependency lives.