Massive Hype Around Orbital AI Data Centers Already in Space

In January 2025, Elon Musk stated at the World Economic Forum in Davos that within two to three years, space-based data centers would become the cheapest way to host artificial intelligence. Shortly afterward, SpaceX filed an application with the Federal Communications Commission to create a constellation of 1 million satellites in low Earth orbit (500–2000 km). Musk also mentioned the initial specifications of the satellite data center AI-1. However, estimates suggest that launching this many satellites would require 16,666 Starship launches (60 satellites per launch) — even with a tenfold increase in SpaceX's current launch rate, this would take ten years. Producing 1 million satellites at Starlink's current capacity (~4000 per year) and a tenfold growth would take 25 years. Cooling a single Nvidia H100 (700 W) in space requires a radiator area of 1.4 m² at 60°C; a 40 kW server rack requires 80 m², and a 100 MW data center would need 2,500 such radiators. Startup Starcloud, which applied for 88,000 satellites, sent one H100 into orbit, but the radiator proved too weak for full power. Analysts believe the project could become economically viable only in 5–10 years, and the only near-term application is inference, not model training, due to synchronization delays. According to an IEEE Spectrum editor, Musk is building a vertically integrated model: xAI builds data centers, SpaceX delivers them to orbit, and Tesla produces solar panels.