Nvidia CEO Jensen Huang declared space computing the final frontier at GTC in March, sparking a rush to orbit‑based AI farms. SpaceX, fresh from acquiring xAI, plans a constellation of data‑center satellites; Google unveiled Project Suncatcher with Planet to launch two TPU‑armed satellites by early 2027. Startup Starcloud has filed an FCC request for an 88,000‑satellite orbital data‑center network.

Proponents cite abundant solar power, free cooling and immunity to earthquakes, floods or protests. Space provides no convection, so heat must radiate through large panels. A single Nvidia H100 GPU drawing 700 W needs about 1.4 m² of radiator to stay at 60 °C; a 32‑GPU rack would need an 80 m² panel, roughly a pickleball court. Radiation degrades panels, raising required area by 40 % over five years.

ABI Research modeled total‑cost‑of‑ownership for an H100 rack launched on a Starship at an optimistic $44 per kilogram. The yearly expense exceeded terrestrial operation by at least tenfold, even assuming $0.20/kWh electricity. Consequently, general‑purpose orbital data centers appear uneconomical, though niche tasks—pre‑processing satellite imagery, real‑time hypersonic missile tracking, or LEO collision avoidance—might justify the steep price despite thermal and radiation hurdles.