SpaceX is putting NVIDIA inside its orbital data centers. The company announced August 4 that it’s partnering with NVIDIA to design the compute payload for the Starmind AI1 satellite.
Each unit will carry Rubin GPUs and Vera CPUs for what SpaceX calls datacenter-class space compute.
That’s the headline. The story underneath is bigger: SpaceX is betting the AI buildout eventually runs out of room on Earth, and the fix is a constellation of solar-powered supercomputers in orbit.
What Starmind AI1 actually is
Starmind is SpaceX’s plan to move data centers off the ground. AI1 is the first satellite design, and the numbers are not small: 30 meters tall deployed, a 75-meter wingspan of solar panels, and a compute payload rated at up to 250 kW peak and 175 kW average.
SpaceX claims 75 kW per ton of vehicle efficiency.
The chips are NVIDIA’s newest generation. The Space-1 Vera Rubin Module delivers up to 25 times the AI compute of an H100 GPU, according to NVIDIA, with volume shipments starting this fall.
Why space beats Earth for AI compute
The pitch is power and cooling. In sun-synchronous orbit the sun is always up, so the panels run continuously with no weather and no atmospheric loss.
Heat radiates straight into vacuum, which SpaceX says cuts cooling overhead by an order of magnitude compared to terrestrial data centers with their chillers, cooling towers, and fans.
The satellites talk to each other and to the ground through the Starlink laser mesh. Results get beamed back at low latency instead of shipping raw data down.
The NVIDIA pick closes a gap
Here’s the part I find most interesting. NVIDIA launched its space computing platform in March with six named partners: Aetherflux, Axiom Space, Kepler Communications, Planet Labs, Sophia Space, and Starcloud. SpaceX wasn’t on the list.
Now it’s the biggest customer on the board. Musk confirmed the direction, saying SpaceX has committed to using NVIDIA GPUs exclusively “because they are the best.”
That’s a notable line for a company that also runs a chip fab project with Tesla called TeraFab.
Starmind’s architecture is chip-vendor agnostic by design, so future satellites can swap compute modules from any provider without a spacecraft redesign. AI1 just happens to be all NVIDIA.
The scale is the hard part
SpaceX filed with the FCC in January for permission to deploy up to one million orbital data center satellites between roughly 500 and 2,000 kilometers up, connected by a petabit laser mesh.
For scale, about 15,000 active satellites orbit Earth today. A million would be a 66x jump.
The FCC accepted the filing in February but hasn’t ruled. Until it does, as one market analyst put it, Starmind is a chip order without a license.
Timeline: prototype testing for Starmind AI1 starts early 2027, with mass production at SpaceX’s Gigasat Factory in Bastrop later that year. The company says the factory is designed to produce thousands of AI satellites.
Neither company said how many Starmind satellites are planned, when the first one launches, or what the deal is worth. Those gaps matter, because the plan depends on more than chips.
The path from today to a working orbital data center is long. SpaceX needs the FCC to approve the constellation, Starship to fly at the cadence and cost the plan assumes, and the Gigasat factory to actually produce these things at scale. Any one of those slips and the timeline slides. None of them are safe bets.
The timing was not an accident
The announcement landed hours before SpaceX reported its first quarterly results as a public company. Revenue jumped 92% year over year across Space, Connectivity, and AI. NVIDIA shares gained about 3% on the day; SpaceX stock climbed nearly 9%.
The numbers that matter are harder. SpaceX’s AI unit lost $6.4 billion on $3.2 billion of revenue last year.
SpaceX stock still sits about 35% below its early July peak. Orbital compute is a story about the future, and the future is expensive before it pays.
What to watch
Three things, really. The FCC ruling on the million-satellite constellation. Whether NVIDIA’s Rubin production stays on schedule for this fall. And whether SpaceX can make the economics of space compute work, because a satellite with a 250 kW payload needs Starship-class launch economics to justify itself.
I covered the SpaceXAI rebrand and Grok Build going open source earlier this summer, and this fits the same pattern: Musk keeps consolidating the AI stack under the SpaceX umbrella.
Starmind is that play aimed at the data center itself.
Starship’s next flight matters here too, because none of this works without cheap mass to orbit.
The NVIDIA partnership is real, and it’s the clearest signal yet that SpaceX intends to be an AI infrastructure company, not just a launch provider. Whether the FCC agrees is a question for another quarter.
