Starcloud's Philip Johnston: Why the Cheapest Compute Will Be in Space

Summary

Philip Johnston, co-founder and CEO of Starcloud, argues that AI compute will increasingly move to space. He explains that orbital data centers avoid land and battery costs, generate eight times more solar energy, and become economical once launch costs fall to around $500/kg. Starcloud plans an 88,000-satellite constellation delivering 20 GW of inference capacity and is working with Nvidia on a space-specific chip.

• Starcloud deployed Star Cloud 1 with an Nvidia H100 and trained a model in space.
• Space data centers avoid permitted land and battery storage costs and get 8x more solar energy per square meter.
• Launch cost break-even is around $500/kg, and Starship is designed for $10-$20/kg.
• Starcloud filed for an 88,000-satellite constellation targeting 20 GW of inference capacity.
• Heat dissipation in a vacuum requires large radiators; Nvidia is developing a hotter-running space chip.
• Kessler risk is manageable with low altitudes and collision avoidance, and radiation effects are mitigated through testing and shielding.
• Inference is expected to be ~99% of the AI compute market in 5-10 years.
• Starcloud's 88,000-satellite constellation would require about $100B capex, less than terrestrial equivalent.