On August 25, Elon Musk gave a concrete date on X: the space-optimized Vera Rubin NVL72, co-designed by SpaceX and NVIDIA, will launch into orbit in the fourth quarter of 2027, with large-scale deployment beginning in 2028.
NVIDIA has already published the ground-based specs for this system: 72 Rubin GPUs and 36 Vera CPUs, paired with ConnectX-9 SuperNICs and BlueField-4 DPUs, spread across 18 compute trays and 9 NVLink switch trays. The whole rack weighs roughly 1,800 kilograms, draws 120 kilowatts per rack, and relies on a newly designed liquid-cooling scheme to manage the heat.
1,800 Kilograms Is the Easy Part, 120 Kilowatts Is Not
Start with the math. A Falcon 9 can carry 22,800 kilograms to low Earth orbit, and its fairing measures 13.1 meters tall and 5.2 meters in diameter on the inside. On weight and volume alone, a single NVL72 rack fits with plenty of margin to spare. The real problems lie elsewhere.
First, power. SpaceX has specified a solar array rated at 210 kilowatts for the satellite carrying this payload, which looks more than sufficient for a 120-kilowatt load. But solar panels in orbit pass through eclipse periods and don't deliver constant output, and SpaceX hasn't said how much margin it's building in or whether there's any energy storage backing it up.
Second, cooling. On the ground, liquid cooling works by transferring heat into cooling water and then rejecting it to the environment. In space there's no convection and no available heat sink, so the only option is radiating heat away through liquid radiators. SpaceX has mentioned it will use liquid radiators, but hasn't disclosed radiator area, operating temperature, or heat-rejection capacity per unit area. Orbital altitude, the radiation-shielding approach, and the exact launch window remain unspecified as well.
From "Any Chip Fits" to "NVIDIA Exclusive"
When SpaceX unveiled the AI1 satellite design back in June, one selling point was that its compute modules were swappable and not locked to any single chipmaker — at the time, it was also working with Tesla on an in-house Terafab chip, and keeping its options open was clearly deliberate. Two months later, the tone shifted. On the August earnings call, Musk made it explicit:
"Going forward, we've decided to build exclusively on Nvidia, because we think the Vera Rubin architecture is the best architecture. We're exclusive to Nvidia."
The newly announced Starmind first-generation satellite runs on this optimized NVL72 setup. On NVIDIA's side, the Vera CPU is positioned as the component that handles agent orchestration, code execution, and data processing; NVIDIA vice president Ian Buck described it as giving "AI agents the CPU performance to act in real time," while SpaceXAI president Mike Nicolls emphasized that Vera's memory bandwidth can sustain heavy orchestration workloads.
The Orbital Data Center Race Is Already Crowded
NVIDIA introduced its Space-1 Vera Rubin module at GTC in March, claiming space-optimized inference performance 25 times that of the H100, and named six partners at the time, including Axiom Space, Kepler Communications, Planet Labs, and Starcloud. Google's Project Suncatcher plans to launch its first TPU prototype satellite in 2027 and was already discussing launch services with SpaceX as of May. Meta, meanwhile, is pursuing a gigawatt-scale orbital solar power scheme.
All of these plans cluster around 2027, and whoever gets a full data-center-grade system running first will come down purely to execution. Musk's track record here isn't reassuring: The New York Times tallied 602 of his public commitments and found only 19% were delivered on schedule, while another 35% were delayed by years or never materialized at all.
Rough math: even if Starmind's first generation launches ten units a year, the on-orbit compute available by 2028 would amount to a fraction of a single mid-sized ground cluster. An orbital data center won't solve the compute crunch anytime soon — what it's really meant to prove is something else: whether hauling a 120-kilowatt liquid-cooled rack into vacuum is even engineering-feasible in the first place.
Sources: The Register, SpaceXAI earnings call transcript, CocoLoop, NVIDIA GTC public materials; NVL72 GPU/CPU counts, rack weight, and 120kW power draw verified against NVIDIA specifications, Falcon 9 payload capacity and fairing dimensions verified against SpaceX public parameters.