Google just signed a $920 million per month deal with SpaceX to rent 110,000 NVIDIA GPUs thr...
Google just agreed to pay SpaceX roughly 920 million dollars a month to rent 110,000 NVIDIA GPUs through 2029, a deal worth close to 33 billion dollars over its full term. For a company that designs its own TPUs and runs some of the largest data centers on the planet, writing that check to Elon Musk's rocket company says something important about how tight AI compute has become.
The contract was disclosed in regulatory filings days before SpaceX's planned June 12 public offering at a 1.77 trillion dollar valuation, which gives the timing an obvious flavor. SpaceX gets to walk into its IPO with a marquee hyperscaler customer locked in, and Google gets a bridge to keep Gemini moving while its own buildout catches up. There is a real out clause here too. Google can cancel in October if the 110,000 chips do not show up on schedule, so the execution risk on SpaceX's side is not abstract.
The technical story underneath the headline number is the more interesting one. Gemini Enterprise demand is outrunning what Google can stand up internally, even with custom silicon and years of data center investment. If a company at that scale needs to rent capacity from outside, the GPU shortage is not a temporary squeeze, it is a structural feature of this phase of the industry. SpaceX has quietly built its way into being a compute landlord through Starbase data centers powered by its own energy infrastructure, and this is already the second mega deal of this shape the company has signed in recent months. The pairing of power generation, land, and chip allocation under one roof is exactly the bundle that everyone else is scrambling to assemble.
The bigger signal is about who sits at the table now. Compute is the new oil, and the operators who control power, land, and GPU allocation have the same leverage as the model builders themselves. A few years ago you would not have predicted that a launch company would be selling training capacity to the company that invented the transformer. That inversion is going to keep happening.
What I will be watching is the second tier. If Google is renting from SpaceX, the smaller frontier labs and the enterprise AI startups behind them have a much harder problem. Expect more unusual landlords, more long term power purchase agreements tied to AI workloads, and more deals where the chip allocation matters more than the cloud brand on the invoice. The compute supply chain is being rewritten in public, one filing at a time.
Originally posted on LinkedIn.