Jensen Huang stood on stage at Nvidia’s GTC and announced that “space computing, the final frontier, has arrived.” SpaceX is talking up a constellation of orbital data centers. Google has a project to put its TPUs in space. A startup called Starcloud asked the FCC for permission to launch 88,000 satellites. The pitch - endless solar power, free cooling in the cold of space, and no earthquakes, floods, or neighbors filing zoning complaints.
It’s a great story. It’s also a bad idea.
Math is Hard#
The fantasy hinges on the idea that space is cold, so cooling is free. It isn’t. Space is cold and hot and empty, and emptiness is a problem. On Earth we shed heat by blowing air or running water across hot chips. In a vacuum there’s nothing to carry the heat away: your only option is to slowly radiate it as infrared light, and that takes enormous, heavy radiator panels that you have to build, launch, and unfold in orbit. The more compute you pack in, the more radiator you haul up. Cooling isn’t free in space; it’s incredibly complex and expensive.
Orbit is a radiation bath that is constantly corrupting data and fries the chips these designs depend on, so you’re stuck either adding heavy shielding or running redundant hardware that double-checks itself. On the International Space Station, all of the main computers are running on triple redundant systems because the radiation in space corrupts data so much, triple redundancy was the only way to effectively handle error correction.
Your solar panels degrade every year. Keeping everything pointed the right way (panels at the sun, radiators at the void, antennas at Earth) means finicky, failure-prone hardware in a place where nobody can come fix it. Turns out, this is incredibly expensive to do!
So why does anyone want to?#
There genuinely are some compute jobs that may make sense in space. AI datacenters are not one of them.
Modern imaging satellites generate more data than we can beam back down, so processing it in orbit and only sending the useful bits might make real sense. With orbit getting dangerously crowded, satellites increasingly need to dodge collisions in milliseconds, too fast for a round trip to a ground station. These could be excellent examples of good additional compute in space.
The point was never that it was a good idea. Most of the people hyping this are billionairs (or trillionaire…) who have figured out that they can distract you with shiny sci-fi fantasy to get people to invest huge amounts of money and ignore all of the problems of everything else they are doing. Everything has become a huge grift, trying to build the next hype cycle to siphon all of the hype from whatever the current hype cycle is, and somehow they always end up with more money.
For a few narrow, latency-sensitive missions, space compute is clever and probably inevitable. For the trillion-dollar “data centers in orbit will power AI” narrative, it’s a bad idea, and the math doesn’t make sense: you’d spend an order of magnitude more to fight a cooling problem that’s trivially solved on the ground.
Don’t take my word for it. The IEEE spectrum article goes in to much more detail, and shows their work.