Starcloud: building AI data centers in orbit

Executive overview

Earth's data centers are hitting hard limits on energy, land, and water. Starcloud's answer is to move the data center itself into space — powered by uninterrupted sunlight, cooled by radiating heat into deep space, using zero fresh water.

In November 2025 they launched the first satellite carrying an Nvidia H100 GPU, beating any prior startup's timeline from founding to orbit by years.

The core bet: launch costs are falling fast enough that orbital compute will undercut terrestrial data centers on energy, before the decade is out.

Why space solves the data center problem

• Terrestrial data centers evaporate vast quantities of fresh water to stay cool — already draining rivers and lakes in parts of the US
• Space offers uninterrupted solar power (sun-synchronous orbit), no land constraints, and infinite thermal sink via infrared radiation
• Zero fresh water required; large deployable radiators dissipate heat passively into vacuum
• Frees compute from grid capacity limits, enabling near-unlimited scale
• Target: 40 MW data centers (~100 tons, fits one Starship payload bay) competing on cost with the largest terrestrial facilities

From idea to orbit in 15 months

• Founded after pivoting from space-based solar: transmitting power wastes 95% of energy; sending the data center up sidesteps that entirely
• Break-even launch cost for space solar: ~$50/kg; for orbital compute: ~$500/kg — already within reach
• Applied to YC as LumenOrbit on third attempt (Summer 2024 batch); YC pushed them to own the full vision
• Satellite 1: ~60 kg, H100 GPU, built and tested in 15 months — record for a startup going from founding to orbit
• 100x more powerful than any previous computer operated in space
• Demo workloads: first model training in space, first fine-tuning, first Gemini run

Team and technical moat

• Founder Philip Johnston: applied math and theoretical physics, then software engineering — not a traditional space background
• Co-founder Adi: 20 years building data centers at Microsoft, then principal software engineer at SpaceX
• Co-founder Ezra (CTO): decade designing satellites, NASA Lunar Pathfinder, PhD in engineering
• Core IP: large-scale, low-mass deployable radiators — the thermal dissipation system critics said was impractical
• Half the engineering team focused solely on the radiator

Next steps and competitive landscape

• Satellite 2 launches October 2026: at least 10x more powerful, Nvidia Blackwell architecture, 24/7 optical high-bandwidth connectivity
• Long-term vision: multi-gigawatt orbital data center constellations handling anything not requiring ultra-low latency
• Google, SpaceX, and Amazon now exploring the same space
• Hard-company principle: one hard problem (cheap orbital compute) — if solved, hiring, press, and fundraising all follow