Google Pays SpaceX $920M Monthly for GPUs Through 2029

Published by James Harris on June 6, 2026June 6, 2026

Google is paying SpaceX $920 million per month to rent roughly 110,000 Nvidia GPUs, locking in a contract that runs through June 2029 and is worth more than $40 billion in total value if it runs to term. The deal, disclosed in an SEC filing, lands weeks before SpaceX’s planned IPO at a $75 billion target valuation.

The sharper detail is the customer. Google operates one of the largest proprietary data center networks on the planet and has spent years developing custom TPU chips specifically to reduce dependence on Nvidia hardware. The fact that it is still writing a $920 million monthly check to a competitor’s infrastructure says more about the current GPU shortage than any analyst report has. Combined with the Anthropic deal struck in May, which gave the Claude developer full access to SpaceX’s Colossus 1 facility in Memphis, the two contracts total more than $70 billion in combined value and roughly $26 billion annually. The Colossus facility was originally built to serve Musk’s own xAI venture. It is now, functionally, a commercial cloud business.

SpaceX’s IPO timing here is almost too clean. Two anchor enterprise contracts, both disclosed in the pre-offering window, both with named counterparties that are credible at any valuation table.

The Exit Terms Matter

The contract structure carries real optionality for Google. If SpaceX fails to deliver the agreed GPU count by September 30, 2026, Google can walk after a one-month grace period or accept reduced capacity at a proportionally lower rate. After December 31, either party can exit with 90 days notice. All IP, AI model weights, and data processed on the infrastructure remain Google’s property. These are not unusual enterprise terms, but they do mean the $40 billion headline figure has meaningful conditionality attached.

For the AI infrastructure sector broadly, this deal accelerates a dynamic already visible in hyperscaler capex disclosures: even the companies best positioned to build their own compute are choosing to rent capacity rather than wait for new data centers to come online. That calculus benefits any operator sitting on large GPU clusters, and it raises the floor on what those assets are worth in a secondary or acquisition context. Coreweave’s public market debut earlier this year established one pricing reference. SpaceX’s IPO, whenever it prices, will establish another.

The ramp structure in the contract, with capacity increasing gradually through September at below-peak rates before hitting the full $920 million monthly figure, means SpaceX’s AI revenue will not fully reflect these contracts until late 2026.