Musk is betting that a stripped-down rack and cheap launches make compute in orbit cheaper than compute on the ground. Analysts, for now, say the opposite.
SpaceX's data centers in orbit now have named silicon and a date. The company and Nvidia have formalized the partnership behind the plan, and SpaceX says the Starmind systems it intends to operate in space will be built around the NVL72 rack from Nvidia's Vera Rubin line. The first of those racks is meant to be working above the planet by the fourth quarter of 2027.
Vera Rubin is Nvidia's newest rack, and the successor to the Grace Blackwell systems now filling data centers on the ground. It holds 72 Rubin graphics processors and 36 Vera CPUs — the two chips the platform is named for — wired to run as a single computer, and it goes into production on Earth only this autumn.
The version bound for orbit is a pared-back edition of that system. By Elon Musk's account it is far less complex, cheaper, denser and lighter than what normally gets deployed on the ground, and reworked to survive the radiation and the heat of space.
Analysts estimate that compute in orbit currently runs at more than four times the price of compute on the ground. Musk claims that gap will flip in orbit's favor within the next few years.
Some of the pull toward orbit is a push from the ground. Opposition to data center buildouts in the United States is mounting, and a site with no nearby neighbors is an appealing alternative for anyone hoping to keep scaling AI infrastructure.
Sam Altman dismissed the idea of running data centers in orbit as ridiculous earlier this year. SpaceX now intends to fly Nvidia's newest hardware about a year after it reaches the ground.