SpaceX Wants to Own the Orbital Economy. Nvidia Is Selling the Shovels.
SpaceX says Nvidia's Rubin GPUs and Vera CPUs will run the first Starmind satellites. Nvidia hasn't confirmed it, and SpaceX is already building the chips meant to replace them.
The men who built America's first railroads were not selling freight. Whoever owned the track decided which towns became cities, which crops reached market, and who paid to move. The Vanderbilts and the Huntingtons sold transport; what they held was the right to set the terms of everyone else's commerce. The second Gilded Age is being wired the same way, one layer at a time, and last week it added a layer in orbit.
SpaceX said on August 4 that it is designing the compute payload for its Starmind AI satellites with Nvidia, using the chipmaker's Rubin GPUs and Vera CPUs for what it called "datacenter class space compute." The claim came from SpaceX's own account. Nvidia, whose silicon anchors the entire pitch, has said nothing: no release, no confirmation, and still no place for SpaceX on the space-computing roster it unveiled in March, which named Axiom Space, Kepler Communications, Planet Labs and Starcloud but not the largest space company on Earth. The silence can be read more than one way. What SpaceX is describing is a machine to move part of the intelligence economy off the planet, and to own most of it on the way up.
SpaceX is partnering with @Nvidia to design the Starmind AI1 satellite compute payload. Each of the Starmind satellites will include NVIDIA Rubin GPUs and Vera CPUs for datacenter class space compute → spacex.com/spacexai/starm…
Start with what one company already controls. SpaceX builds the rocket that lifts the hardware, the satellite that holds it, the laser links that connect it, and the Starlink network that carries the answers back down. It means to generate the power from solar arrays above the atmosphere and to manufacture the satellites at a new Bastrop, Texas plant called Gigasat. Transport, hardware, energy, network, compute: five layers of an industry, inside one balance sheet. Add the sixth. SpaceX runs Terafab, a chip venture with Tesla, and is developing its own radiation-tolerant orbital processor, known internally as D3. The Nvidia deal is the fastest route to world-class silicon in the first satellites. It is unlikely to be the last word. Musk replaced bought launch vehicles with Falcon, bought satellites with Starlink, bought battery cells with Tesla's own; depending on Nvidia for the one part that makes Starmind worth building is the kind of reliance he tends to engineer away.
Watch @ElonMusk provide a technical update on SpaceX’s capability to manufacture, launch, and operate AI satellites at scale → spacexipo.com
For Nvidia, that question can wait. Its business is to sit underneath whatever wins. Carnegie sold steel to whoever laid track and Rockefeller sold oil to whoever burned it; Jensen Huang sells the engines of machine intelligence to everyone racing to build it, including the ones planning to leave. At its March launch Huang said, "Space computing, the final frontier, has arrived," and rated the Rubin part at up to twenty-five times the space inference of an H100. Whether Starmind's first satellites carry the full seventy-two-GPU Rubin rack Nvidia sells on the ground is not known; the payload is built to swap chips, and what flies may not match the brochure.
The scale is real whatever one makes of it. SpaceX filed with the FCC on January 30 for up to a million satellites, at altitudes from 500 to 2,000 kilometres, and the agency has not yet ruled. It unveiled AI1, the first craft, in a June 8 video three days before an IPO that raised about seventy-five billion dollars at a valuation above 1.75 trillion, the largest listing on record. Each AI1 spans roughly seventy metres, wider than a 747's fuselage, and draws around 120 kilowatts on average and 150 at peak, close to a single high-end Nvidia rack on the ground, figures SpaceX disclosed at the reveal and Tom's Hardware detailed. Musk has since pushed the peak target toward 250 kilowatts. Two prototypes are meant to fly in early 2027, with volume production to follow at Gigasat and each Starship carrying dozens of satellites at a time. One satellite is a demonstration. A constellation is infrastructure, and infrastructure is the toll booth.
The Gilded Age of intelligence
The optimistic interpretation of Starmind is irresistible. Humanity’s demand for computation is colliding with limited electrical grids, slow permitting, water consumption and local opposition to data-centre construction. Space offers enormous solar resources and an industrial territory measured not in acres but in orbital planes.
Starmind could process scientific observations where they are produced, coordinate fleets of autonomous spacecraft, support lunar and Martian industry and eventually provide additional computing capacity for Earth.
The darker interpretation is that one company could end up controlling the rockets, satellites, energy supply, data network and computing infrastructure of an emerging orbital economy. The original railroad tycoons did not merely make transportation cheaper. They acquired the ability to decide who moved, who waited and who paid.
SpaceX is assembling a similar position above the atmosphere.
Why put any of it in orbit? Musk's theory of space runs through three claims, and Starmind sits at the leading edge of it. The first is insurance: a species on one planet has a single point of failure, an asteroid or a war or an engineered pathogen or a mistake no one has pictured yet. Mars, in his telling, is a backup for consciousness itself, a second copy of the human record kept far enough away to survive the loss of the first. The second is morale. Musk argues that a civilisation needs to believe its future is worth building, and that a frontier is what supplies the belief; people, he says in as many words, should wake up wanting the future to arrive.
Call it billionaire romanticism if you like, it is still the half of the argument that explains why he spends the money. The third is the one most accounts skip: progress is not automatic. The United States put people on the Moon in 1969 and then could not do it again for half a century. Musk treats that as a warning, not a footnote: capability rots when the institutions holding it stop trying, and technical possibility does not keep the window open by itself.
Starmind is where the philosophy stops being about rockets and starts being about wiring. A Mars settlement will have to think locally, its construction robots and power grids and medical systems unable to wait on a signal from Earth. Moving compute, energy and communication off the planet before people follow is the Mars project seen from the other end: the nervous system laid down ahead of the body.
The case against all this comes down to thermodynamics and money. Space cools nothing for free, whatever the popular image of orbit as a giant refrigerator suggests. In a vacuum there is no air to carry heat off, so every watt a GPU burns has to be radiated from a physical surface, which is why AI1 is mostly not computer. Its seventy-metre span is solar array and radiator: SpaceX says up to 110 square metres of deployable liquid radiators, plus pumps and shielding, to dump waste heat into the dark. Outside engineers doubt the thermal and mass numbers close, and no AI1 has flown, so no one can settle it with data. Radiation is a second problem, since data-centre chips are built for buildings rather than years in a particle bath, and the sceptics of orbital compute, as Data Center Dynamics has reported, include OpenAI's Sam Altman, the short-seller Jim Chanos and AWS's Matt Garman.
Musk's answer is the one he gives to every hard number: Starship changes the arithmetic, and nothing in the design needs technology that does not already exist, since most of AI1 is adapted from Starlink's latest satellites. On the economics he is blunter, and this is where the bet lives: in space, he says, power "becomes actually cheaper and easier over time," while on Earth it only grows harder and dearer. That may hold. It is also what every railroad prospectus promised, that once the network exists its costs fall and its rivals cannot follow. There was no magic in the transcontinental line either. There was steel, borrowed money, political muscle, and years of losses absorbed until the track became the only way through.
The first Gilded Age barons did not grow rich moving goods. They grew rich owning the road everything else had to travel. SpaceX is building that road again, overhead this time, and asking Nvidia to pave the first mile.
Dr. Evelyn Reed writes on AI governance, policy and the funding implications of regulation. A former policy advisor, she reads the rules so builders don't have to.
