WhoBuilt

Inside The Startup Launching AI Data Centers Into Space

Nov 13, 2025· 13 min· 333K views·Watch on YouTube

Speaker

Philip Johnston, co-founder of StarCloud, describes the company’s effort to build orbital data centers during a visit to its Redmond, Washington headquarters with Y Combinator.

Why put data centers in orbit

StarCloud aims to start by providing GPU compute to other satellites, then build larger orbital facilities. Johnston says constant sunlight could supply power, while radiators would release heat into space without using fresh water. He argues this could reduce pressure on Earth’s electricity grids and water supplies.

From solar power to orbital computing

Johnston studied applied math and theoretical physics and began his career in software. His team first explored space-based solar power, but estimated launch costs would need to fall to about $50 per kilogram for the idea to break even. After accounting for energy lost when beaming power to Earth, they shifted to sending data centers into orbit, with an estimated break-even launch cost of $500 per kilogram.

The H100 satellite and its engineering

StarCloud says it designed, built, and tested its first satellite in 15 months. The roughly 60-kilogram spacecraft carries an NVIDIA H100, and the team built its compute module, antennas, power system, and thermal hardware in-house while contracting Astro Digital to make the satellite bus. Johnston identifies radiation shielding and a deployable radiator as key engineering challenges.

A team built for the challenge

StarCloud joined Y Combinator in summer 2024, after three application attempts, under the name Lumen Orbit. Johnston says YC encouraged the founders to discuss the larger ambition of serving data centers beyond satellite customers. Co-founder Addi brings data center and SpaceX software experience, while CTO Ezra has a decade of satellite design experience and a PhD in engineering.

What comes after the first launch

Starcloud-1 launched on November 2 with the first NVIDIA H100 deployed in space. The company planned tests including model training, inference, and running Gemini, though the transcript describes these as future demonstrations. Johnston says a planned second satellite would have at least ten times the first satellite’s capability, use NVIDIA Blackwell GPUs, and add high-bandwidth optical links.

Takeaways

  • Estimate the economics of a new infrastructure model early; StarCloud compared launch-cost break-even points for two concepts.
  • Choose an initial customer use case that can validate technology before pursuing a much larger market.
  • Identify the engineering constraint that could decide feasibility; StarCloud emphasizes heat dissipation and deployable radiators.
  • Pair expertise across the system: compute software, data center operations, satellite structures, and thermal design.
  • Separate completed milestones from planned demonstrations when assessing a technically ambitious product.

Mentioned in this talk

Search inside this talk

What was claimed 18

Related talks