Starcloud's Philip Johnston: Why the Cheapest Compute Will Be in Space
- Source
- youtube.com
- Author
- Sequoia Capital
- Date
Context
Starcloud co-founder and CEO Philip Johnston argues that space will become the cheapest place to power AI computing, and his case rests on cost comparisons with a terrestrial solar project. In a short talk hosted by Sequoia Capital, he says a space project avoids buying permitted land, avoids batteries because the orbit is in constant sunlight, and needs far fewer solar cells because a panel in space yields about eight times the energy of one on Earth.
The remaining major cost is launch. He puts the break-even at roughly $500 per kilogram, about a tenfold drop from today, and says SpaceX's Starship is designed for $10 to $20. On heat, he says radiators need about a quarter more area than the solar panels, and chips built to run hotter, which he says Starcloud is working on, would shrink them.
These are the CEO's own estimates in a company pitch, with no supporting data shown. He says Starcloud's first satellite trained a small and ran in orbit, which shows operation there, not the cost case.
- Philip Johnston says space solar removes the two largest costs of an Earth solar build, permitted land and battery backup, and needs about 8x fewer cells. He puts the break-even launch cost near $500/kg, roughly 10x below today, versus Starship's $10-20/kg design target.
- Johnston's radiator math: panels yield about 200 W per square meter while a radiator at 50 C sheds about 800, so radiator area is roughly a quarter of panel area. Because dissipation scales with to the fourth power, running at 80 C nearly halves it.
- Starcloud handles radiation bit flips mainly through ground testing, according to Johnston: four rounds of proton testing at a Knoxville cyclotron plus heavy-ion runs at Brookhaven, compressing about 5 years of dose into 24 hours to guide shielding and software choices.
- Johnston says the planned 88,000-satellite constellation (about 200 kW each, roughly 20 GW total) targets inference only. Large-scale would need a gigawatt-scale structure about 4 km by 4 km that he estimates is at least 15 years away.
- On debris risk, Johnston argues low orbits limit Kessler-style cascades: Starcloud's first satellite flies near 400 km and deorbits naturally within a few months, and he cites SpaceX running about 10,000 satellites with sophisticated collision avoidance and no collisions.
- Model — A computational system that uses learned patterns to produce an output from information supplied to it.
- Inference — Using a trained model to produce an output from a supplied input.
- Training — The process that adjusts a model using examples and a measure of how well it performs.
- Sampling controls — Parameters that shape how deterministic or varied model generation becomes.
videoBringing Continual Learning to Every Company: Trajectory’s Ronak Malde
videoCrusoe CEO Chase Lochmiller: Turning Stranded Energy Into Intelligence
videoDavid Silver and Lasse Espeholt of Ineffable Intelligence: A Superlearner for the Age of Experience
videoCracking the Code on Offensive Security With AI ft XBOW CEO and GitHub Copilot Creator Oege de Moor
Checking sign-in…
Loading comments…


