Why the economics of orbital AI are so brutal

Elon Musk wants to move artificial intelligence off the planet entirely. SpaceX has filed regulatory requests to build solar-powered orbital data centers distributed across up to a million satellites, capable of shifting 100 gigawatts of compute power into space. "By far the cheapest place to put AI will be space in 36 months or less," Musk said on a recent podcast.

Google has announced its own Project Suncatcher launching in 2027, while other tech giants are placing similar bets on orbital computing becoming reality.

The economics remain challenging. Current orbital data centers would cost roughly three times more than Earth-based facilities, around $42.4 billion for one gigawatt of capacity. Success depends on SpaceX's unfinished Starship rocket dramatically reducing launch costs from today's $3,600 per kilogram to around $200 per kilogram.

Space presents unique hurdles beyond cost. Without atmosphere, heat dissipation requires massive radiators. Cosmic radiation degrades chips and corrupts data. Solar panels, while five to eight times more efficient in space, deteriorate faster and limit satellite lifespans to about five years.

The applications may differ too. While AI inference tasks could work well in orbit, training large models requires thousands of coordinated processors that current satellite communication links cannot support effectively.