Starmind: Don't Want AI in Your Backyard? Good Old Musk Will Put It in Orbit!
A familiar Musk pattern: two of his companies, SpaceX and Tesla, have formed special committees of independent directors to study a merger. A combined entity would be worth something on the order of $3.5 trillion, the largest corporate combination in history, tying rockets, satellites, electric vehicles, batteries, robots, and AI computers under one roof. Shareholders would mostly just be exchanging one set of Musk-company paper for another. Fairness opinions will be written. The real question is whether the industrial logic is as clean as the shareholder logic?
One piece of that logic is already being stress-tested in public: the idea that the fastest way to resolve the terrestrial fight over AI data centers is to stop building them on Earth. Musk has been explicit. SpaceX (now with xAI under its roof) is pursuing Starmind, a constellation of up to one million solar-powered satellites that would run AI inference and, eventually, training workloads in low Earth orbit. Prototypes are aimed at 2027; volume production is planned at a Texas factory. Each early unit is roughly the size of a large aircraft's wingspan, generates on the order of 120–175 kW of continuous compute, and is meant to fly in tight clusters of about ten satellites linked by multi-terabit laser connections. The pitch is simple and, on paper, attractive: solar intensity is several times higher outside the atmosphere, sunlight can be nearly continuous in the right orbit, there is no land, no water, no grid interconnection queue, and no local zoning board.
That last point is the political one. Data-centre proposals on Earth are running into power constraints, water use for cooling, and classic NIMBY opposition. Moving the load off-planet sidesteps all three. The vacuum is also a giant heat sink. Radiators can dump waste heat directly into deep space without chillers or evaporative towers. Musk's claim is that within a couple of years the lowest-cost place to put new AI computers will be orbit, and that a merged SpaceX–Tesla entity would be uniquely positioned to industrialise it, Starship for mass, Tesla for batteries and solar, robotics for manufacturing and eventual servicing.
The engineering and operational problems are not small, and they are well documented by people who are not trying to sell the concept.
Cooling is the hardest physics constraint. There is no air and therefore no convection. Heat leaves only by radiation. For the power densities AI chips produce, that requires large radiator area, tens to hundreds of square meters per satellite for early designs, and far more at true gigawatt scale. Radiators add mass, which multiplies launch cost, and they degrade under atomic oxygen and radiation, further increasing the area you must carry. Some analyses put the radiator requirement for a single high-end rack on the order of a basketball court or larger once realistic operating temperatures and degradation are included.
Radiation is the second fundamental issue. Galactic cosmic rays and solar particles flip bits and can latch up or permanently damage commercial silicon. Radiation-hardened parts exist but are slower, more expensive, and lower-density than the GPUs used in terrestrial clusters. Software error correction, redundancy, and selective shielding can mitigate the problem, but they impose a performance and mass tax. Over multi-year lifetimes the cumulative dose still forces design compromises.
Maintenance and replacement are the third. You cannot send a technician with a screwdriver. Failed GPUs, batteries, or solar arrays currently mean a lost or degraded satellite. In-orbit servicing is still developmental. The practical answer for early systems is high redundancy, short design life, and cheap enough launches that you simply replace the entire vehicle. That only works if Starship (or equivalent) drives launch costs down by another order of magnitude and flies at very high cadence.
Scale and orbital traffic are the fourth. A million satellites is roughly two orders of magnitude beyond today's Starlink fleet. Even if they operate in dense clusters rather than as fully independent nodes, the collision and debris risk rises. SpaceX has published safety concepts that rely on frequent ephemeris sharing, onboard sensors, and autonomous manoeuvring, the same techniques it uses for Starlink, but the absolute numbers remain extreme. Astronomers already complain about Starlink; a denser compute layer would intensify the interference.
Economics follow the engineering. Today the combination of launch cost, radiator mass, radiation mitigation, and replacement cycles makes orbital compute more expensive than a well-sited terrestrial data centre. Multiple independent estimates put the premium at 1.5–3× until launch costs fall into the low hundreds of dollars per kilogram and radiator and chip designs improve. Power is cheap in space; everything else is not, yet.
None of these problems is purely theoretical, and none is automatically fatal. They are the same class of constraint Musk's companies have treated as engineering problems rather than showstoppers: reusability, mass production of satellites, high-bandwidth optical links, and rapid iteration. A merged SpaceX–Tesla entity would have, in one balance sheet, the launch vehicle, the satellite manufacturing, the solar and battery expertise, the robotics for both ground production and eventual on-orbit work, and the AI models that need the compute. That vertical integration is the actual industrial case for the combination, beyond the overlapping shareholder base.
Whether it works on Musk's preferred timeline is a different question. Two-to-three years to cost parity is aggressive; five-to-ten years with continuous design improvement is more plausible if Starship cadence materialises. Early deployments will almost certainly be inference or specialised workloads rather than frontier training runs. The terrestrial data-centre build-out will not pause while the orbital one is proven. But the directional logic is coherent: the constraints that are binding on Earth (power interconnection, water, land, local politics) are largely absent in orbit, and the constraints that are binding in orbit (heat rejection, radiation, launch mass, servicing) are exactly the ones SpaceX and Tesla have spent the last decade attacking with capital and iteration.
The committees will produce fairness opinions. The harder test is whether the combined company can turn the orbital data-centre thesis from a filing and a few prototypes into a production system whose cost curve actually undercuts the next generation of Earth-based clusters. If it can, the NIMBY fights over megawatt-scale server farms become a transitional problem rather than a permanent one. If it cannot, the merger still leaves you with the world's most capable rocket company and the world's most valuable automaker under one roof, which is not nothing.
