Google rents the chips it didn't build

Google designs its own AI chips. For a decade its TPUs have been the in-house answer to depending on NVIDIA, the whole point being not to pay a rival for the scarcest hardware in computing. Then SpaceX filed to go public, and the disclosure included a second compute lease: starting October 2026 and running through June 2029, Google will pay SpaceX about $920 million a month — roughly $11 billion a year — for around 110,000 NVIDIA GPUs sitting inside xAI's data centers.

The buyer trying hardest to escape NVIDIA is leasing NVIDIA chips from a rival to do it.

So the buyer trying hardest to escape NVIDIA is renting NVIDIA capacity, from Elon Musk, to keep up with demand. Google calls it short-term bridge capacity for Gemini Enterprise, its agent platform — an admission that even a company with its own chip line and its own data centers can't pour concrete fast enough for the present moment, and will lease its rival's overflow to cover the gap.

It is the second such lease in the same filing. SpaceX already rents Anthropic far more compute, and the two deals together put its projected income from renting out chips above what it made in 2025 from rockets, Starlink, and AI combined. The rocket company that built GPU clusters to train Grok is on track to earn more as a landlord for other people's models than from launching anything. And whoever rents, NVIDIA still gets paid — the quiet constant under every route the industry takes to get out from under it.