Y Combinator › The Light Cone
The Case for Data Centers in Space
StarCloud CEO Philip Johnston explains why orbit is becoming the only viable place to build AI compute — and how a wax-cooled H100 dunked in an ice bath proved it works.
Tap a timestamp pill below to jump the video to that moment.
The problem
AI needs power Earth can no longer provide
The defining constraint on artificial intelligence in 2026 is not algorithms, not chips, and not talent. It is energy. Every large training run, every inference cluster, every frontier lab racing to the next frontier requires a data center — and building one means securing a large block of grid power, which now takes years and is increasingly being denied outright. New York has banned new data center construction. Several other states are moving in the same direction, citing grid strain, water use, and the general political toxicity of “big tech” projects.
Philip Johnston, co-founder and CEO of StarCloud, watched this ceiling materialize in slow motion. His insight came not from the data center industry but from a 2023 weekend trip to Starbase, Texas, where SpaceX was quietly building two gigafactories capable of producing three Starships per day. Johnston did a simple thought experiment: if launch cost fell to one-tenth of today’s price and capacity grew by a thousand times, what would suddenly make sense that doesn’t make sense now? He started sketching out space-based solar power. Then he realised the power was not the thing to beam down — it was the compute itself.
“New York now blocking data centers construction… for reasons which are not grounded in science”
New York now blocking data centers construction, you know, for reasons which are not grounded in science really. It seems like more of vibes than anything else. Building these things in space will definitely be much easier from a regulatory perspective. … It has become basically a national security issue for America and for the west to have access to intelligence. So for tech policy to sort of put us back into the stone age and possibly lose the war for us — that puts StarCloud in a very different position.
The economics crystallised when Johnston ran the break-even math two ways. Space-based solar — the classic sci-fi vision of giant orbital mirrors beaming energy earthward — needs launch costs below $50 per kilogram to pencil out, because you lose 95% of the energy in transmission. But space-based compute is different: you don’t need to re-enter a product. You just need to get the data back. Crunch the numbers and the break-even launch cost for orbital data centers is around $500 per kilogram. SpaceX’s Falcon 9 rideshare is already within reach of that figure. StarCloud pivoted overnight.
“We came to a much closer to reality number of $500 a kilo”
We reran the calculations. What would the launch cost break-even need to be to make data centers in space break even? And we came to a much closer to reality number of $500 a kilo. That’s what we think currently the break-even launch cost is. And so then we rapidly pivoted the company towards that.
How they solve it
GPUs in orbit, cooled with wax, hardened with particle beams
StarCloud’s fundamental proposition is simple: put GPU clusters in low Earth orbit where solar energy is free, continuous (in a polar sun-synchronous orbit), and unconstrained by permits. But simple propositions can hide brutal engineering. Johnston says his 20-person team in Redmond — half former SpaceX, half from hyperscalers — is split almost exactly 50/50 across two unsolved problems: how to dump heat in a vacuum, and how to keep chips working in radiation.
“Team is split basically 50/50 — get rid of heat in a vacuum, make chips work in radiation”
The two biggest ones that are outstanding — and most of our engineering team is split basically 50/50 down these two lines — is number one, as you mentioned, how do you get rid of this heat in a vacuum? And then number two, how do we make the chips work in a higher radiation environment? Some of the other problems are being solved by other people. The interconnect: we’ve just signed a contract with SpaceX for our next 20 satellites to have a Starlink laser terminal that gives us very high bandwidth, low latency connectivity. The core ones we’re solving are thermal and radiation.
The thermal problem is not about cold. Space looks cold, but a vacuum has no atmosphere to carry heat away from a chip. StarCloud’s answer is a large deployable liquid-cooled radiator — not a new physics idea, the ISS has one — but engineered to be ten times less massive per watt of dissipation and 500 times cheaper per watt than the space station’s version. The difference is manufacturing ambition: they treat it like an industrial product, not an aerospace programme. For StarCloud 1, launched in November 2025, co-founder Addy and Ezra had a more improvised solution. They submerged the entire H100 — board, memory, power delivery — in a phase-change material, a wax-like substance that absorbs heat by melting. Nobody had ever tried this for an orbital GPU cluster. To test whether the material would crack the board under thermal cycling, the team ran a 5 a.m. experiment on the day of shipment: dunking the assembly in an ice bath to freeze it, blasting it with hot-air guns to melt it, then dunking again.
“Ezra like dunking this thing in an ice bath at 5am, hot air guns to heat it back up”
We needed to know if we heat up this phase-change material and cool it down — heat it up and cool it down — is it going to crack anything. At like 5:00 a.m. the day we had to ship it, working through the night, Ezra is dunking this thing in an ice bath to cool it down. And then we pull it out and I’m like, “Are you sure it’s going to be fine with the electronics?” And they’re like, “Don’t worry about it.” Then we had these hot-air guns to heat it up, melt all the wax, and then dunking it back in the ice bath. It is a miracle that it works, to be honest.
The radiation problem is solved through brute empiricism. StarCloud ships GPU hardware to Brookhaven National Lab’s particle accelerator for heavy-ion testing, and to a cyclotron in Knoxville for high-velocity proton testing, exposing each assembly to the equivalent radiation dose of a five-year orbital mission over 24 hours. That data informs both shielding design and the software-level error correction that catches bit flips before they corrupt a model weight. Critically, Johnston says they are probably the only organisation on Earth that now knows exactly where an H100, H200, and B200 will fail under particle bombardment — which makes them indispensable to Nvidia, with whom they are co-designing a new “Rubin Space” chip for orbital use.
The roadmap has three rungs. StarCloud 1 (November 2025) was the proof of concept: one H100 in wax, five GPUs total, launched as part of a SpaceX rideshare for roughly $300,000. StarCloud 2 is the first commercial product: a 10-kilowatt spacecraft targeting government and military customers who need on-orbit inference for satellite imagery and synthetic aperture radar data. StarCloud 3 is the hyperscale play: a 200-kilowatt, three-tonne satellite that can fit 50 units per Starship, carrying 10 megawatts of compute per launch. StarCloud has filed with the FCC for 88,000 of these — yielding 20 gigawatts of orbital compute, roughly 20 times the entire US power grid, and a theoretical ceiling of 10 terawatts in dawn-dusk sun-synchronous orbit.
“20 gigawatt of new compute capacity… 20 times the entire US power grid”
That means on the order of 20 gigawatts of new compute capacity. And we can fit up to 10 terawatts in the dawn sun-synchronous orbit. That’s like 20 times the entire US power grid. … For context, the largest data center deployment on Earth today is about a gigawatt. So you’re talking about 20 times the largest data center on Earth — per Starship launch cadence.
Takeaway
The quick version
- The terrestrial ceiling is real: energy permits, land opposition, and regulatory bans are already throttling AI data center growth — and the political climate is getting worse, not better.
- $500/kg is the orbital compute break-even, already approaching viability with Falcon 9 rideshare; Starship could cut that by another order of magnitude.
- The two hard engineering problems — vacuum cooling and radiation hardening — are being solved with radiators 500× cheaper than the ISS and particle-accelerator-tested commodity GPU boards, not bespoke space-grade parts.
- StarCloud went from a $300K rideshare experiment (StarCloud 1) to a $170M raise and YC’s fastest unicorn in 17 months — by proving the physics first, then letting the narrative catch up.
“We can fit up to 10 terawatts in dawn sun-synchronous orbit. That’s 20 times the entire US power grid.”— Philip Johnston, StarCloud / Y Combinator Light Cone
Whether or not StarCloud reaches 88,000 satellites, the argument Johnston makes here is already reshaping how investors and policymakers think about where AI infrastructure can go. Once you accept that building data centers on Earth is becoming politically impossible, an H100 in a wax-filled box tumbling at 17,000 mph starts to look less like science fiction and more like the only available option.