Elon Musk told the World Economic Forum that the cheapest place to run AI will be in space within two to three years. Days later SpaceX submitted an FCC application for an orbital data‑center constellation of up to 1 million satellites in low‑Earth orbit, 500‑2,000 km up, as the company eyed a public listing.

Current orbital capacity makes the math unrealistic. About 14,500 satellites orbit today, two‑thirds of them already belong to Starlink. Launching 1 million data‑center units would require roughly 16,700 Starship flights, each carrying 60 payloads. Even at SpaceX’s record 165 launches in 2025, a tenfold increase would still span a decade, and it also requires a tenfold jump from Starlink’s 4,000‑unit annual output.

Cooling proves another blocker. A single Nvidia H100 GPU draws 700 W and needs 1.4 m² of radiator at 60 °C; a 40‑kW rack requires 80 m², and a 100‑MW data centre would need 2,500 such panels. Starcloud’s test satellite barely ran the chip, and astronomers warn a million radiators could obscure the night sky. Costs would far exceed ground‑based options. The hype outpaces engineering reality.