SpaceX and Nvidia Designed a Vera Rubin Rack for Orbit. The Heat Still Has to Get Out.

Elon Musk confirmed it in a single sentence on X. "SpaceX, in partnership with Nvidia, has designed a space-optimized Vera Rubin NVL72 system for launch to orbit in Q4 next year, with significant scale in 2028," he posted on August 24, hours after Nvidia's own press release had pointedly avoided naming any date. That line put a delivery window on the most aggressive orbital-compute claim any company has made, and shifted the argument from whether AI data centers belong in space to whether a rack engineered for a liquid-cooled machine room can be re-engineered for vacuum in roughly fifteen months.

What Nvidia actually announced is narrower than the headlines suggest. The August 24 release is primarily a CPU deal: SpaceXAI will deploy Nvidia's standalone Vera processor, an 88-core chip with up to 1.2 TB/s of memory bandwidth, to orchestrate the tool calls, code execution and data processing that surround model inference. "Vera gives us the CPU performance and memory bandwidth to run enormous amounts of orchestration, code and data processing while keeping GPUs doing what they do best," said Mike Nicolls, president of SpaceXAI, in the release. "That means higher-performance AI agents and more useful work from every watt of compute."

The orbital piece occupies three sentences. Nvidia says SpaceXAI's "planned first-generation Starmind AI satellite will be based on the optimized NVIDIA Vera Rubin NVL72 rack-scale system," and that the two companies "are working to adapt that foundation to the requirements of orbital computing." No launch date, no unit count, no deal value. Ian Buck, Nvidia's vice president of hyperscale and high-performance computing, framed it as continuity: "SpaceXAI is taking this architecture from massive AI factories to the next frontier of computing in orbit." Musk supplied the date the press release would not.

The specifications that do exist are modest by data-center standards and enormous by satellite standards. The terrestrial NVL72 packs 72 Rubin GPUs and 36 Vera CPUs into one fully liquid-cooled rack. SpaceXAI's first-generation AI1 satellite carries a 120 kW compute payload peaking at 150 kW on a spacecraft wider than a Boeing 747: roughly one rack strapped to a solar array the size of a wide-body airliner.

The heat problem

That ratio is the whole story, and it is not a marketing problem. Vacuum is an excellent insulator. Terrestrial racks dump heat into air and water; an orbital rack can only radiate it away as infrared, governed by the Stefan-Boltzmann law, where radiated power scales with the fourth power of absolute temperature. Industry estimates put the requirement at roughly 1,200 square meters of radiator per megawatt rejected at a stable 20C, about four tennis courts. An IEEE Spectrum analysis cited by Brookings calculated that a full-scale orbital data center would need on the order of 2.15 million square feet of radiators.

"It's counterintuitive, but it's hard to actually cool things in space because there's no medium to transmit hot to cold," Voyager Technologies CEO Dylan Taylor said earlier this year. Engineers can shrink the radiator by running it hotter, since 60C instead of 20C roughly halves the required area, but that pushes silicon toward its junction limits and trades longevity for mass.

Radiation is the second tax. Orbital hardware faces trapped-belt particles, solar energetic events, single-event upsets, latchup and cumulative total ionizing dose. Radiation-hardened parts typically lag commercial silicon by a generation or more, defeating the purpose of flying frontier accelerators. The alternative is tolerating faults with ECC memory, watchdog resets and redundancy, and accepting a higher failure rate on hardware nobody can service.

Then there is cost. Andrew McCalip, an aerospace engineer who published an open first-principles model comparing orbital and terrestrial builds, priced 1 GW of orbital solar infrastructure at $31.20 per watt and $891/MWh against $14.80 per watt and $398/MWh for a terrestrial gas-fired build, requiring roughly 22.2 million kilograms delivered to low Earth orbit across about 222 Starship launches. His verdict is blunt. "If you run the numbers honestly, the physics doesn't immediately kill it, but the economics are savage," he wrote. "This is not a 25% mismatch. It's 400%. Closing that is the whole job."

Why It Matters

Orbital compute has become a financial instrument as much as an engineering program. SpaceX's June 12 IPO priced more than 555 million shares at $135 and featured orbital data centers prominently in its filings; the stock carried the company toward a $3 trillion valuation within days, on a technology with no proof of concept at scale. Brookings fellow Lauren Tokos warned that space communications firms may be presenting "the illusion of a solution" to justify valuations, in a market where rivals already depend on SpaceX for launch.

The comparison set is instructive about pace. Google's Project Suncatcher concluded that space-based machine learning compute is "not precluded by fundamental physics or insurmountable economic barriers," yet is flying only two prototype TPU satellites with Planet in early 2027 and projects economic feasibility around 2035. Starcloud, which raised $250 million at a $2.3 billion valuation in August, already has an H100 operating in orbit and trained a small model there; its next launch, in October, adds Blackwell. Nvidia's own Space-1 Vera Rubin Module, announced at GTC in March alongside Jensen Huang's line that "space computing, the final frontier, has arrived," is still listed as available at a later date. Against that backdrop, a full NVL72-derived rack in orbit by late 2027 is an outlier claim.

What to Watch

Three things separate program from press release. First, whether SpaceXAI publishes a radiator architecture and thermal budget for AI1: 150 kW peak implies deployable radiator area never flown at that scale. Second, whether the Q4 2027 window survives contact with Nvidia's Space-1 module schedule, which still carries no announced availability date. Third, supply: SpaceXAI has acknowledged that orbital compute at its target scale requires substantially more chips than it currently has access to, a constraint the Vera Rubin commitment does not resolve. Starcloud's October Blackwell flight will return real thermal and radiation telemetry long before Starmind flies. Watch that data, not the valuation.