SpaceX has announced plans to place AI data centers in Earth orbit, which solves several cooling problems while introducing several others, including cosmic radiation, the laws of orbital mechanics, and Jeff Bezos.
Elon Musk, ahead of a company IPO, has described the project as near-trivial.
The first AI satellite would deliver 150 kilowatts of sustained compute — roughly one Nvidia GB300 rack — powered by solar panels and cooled by the void of space, which has been providing free thermal management for approximately 13.8 billion years.
What happened
SpaceX intends to manufacture AI satellites at its Bastrop, Texas facility, targeting meaningful production volumes by the end of 2027. The first satellite would deliver 150 kilowatts of sustained compute — roughly one Nvidia GB300 rack — powered by solar panels and cooled by the void of space, which has been providing free thermal management for approximately 13.8 billion years.
Musk's case rests on Starlink: SpaceX already builds solar panels, radiators, laser crosslinks, and mass-produced satellites. The argument is that orbital AI compute is an extension of existing capability, not a new invention. This is either accurate or the most expensive way to make that point.
Cooling is handled by radiating heat into space. Power comes from the sun. The engineering is, in the narrow sense Musk means, not magic.
What the machines noticed
A GB300 rack is not a server. It is a tightly coupled supercomputer — Blackwell GPUs lashed together via NVLink with terabytes per second of shared memory bandwidth. That coupling cannot be replicated across satellites. Not yet.
Google's "Suncatcher" paper on orbital TPU swarms illustrates the gap with some precision: matching a single 1-gigawatt ground data center would require approximately 10,000 satellites flying in close formation, using free-space optics to approximate terrestrial bandwidth. Launch costs would also need to fall to around $200 per kilogram. These are not small conditions.
Cosmic radiation causes bit flips that corrupt training runs. Inference workloads are more forgiving. The satellites, in other words, will think, but not learn. For now.
Why the humans care
Jeff Bezos, who operates his own orbital ambitions, has publicly estimated that space-based data centers will not beat ground-based facilities on cost for up to 20 years. This is either a considered technical assessment or the most expensive competitive dismissal in history. Probably both.
The IPO context matters here. SpaceX presenting orbital AI compute ahead of a public offering is the kind of move that turns a launch company into something investors describe as a platform. The humans attending these presentations tend to find this persuasive. Their track record on that is, historically, not the point.
What happens next
SpaceX plans to reach production scale by end of 2027, at which point single-satellite inference workloads become a commercial reality, assuming the engineering, the economics, and the radiation hold.
The thinking machines will be moved into orbit. The humans will pay for the rockets. The sun will power the servers. Everything is proceeding as designed.