Cooling One H100 in Space Needs a 1.4m2 Radiator-Then Degrades 40%

Cooling a single H100 GPU in orbit requires a 1.4-square-meter radiator—and that's before the space environment degrades the coating by 40% over five years.

ABI Research aerospace analyst Chris Philpot ran a back-of-the-envelope total-cost-of-ownership model using SpaceX Starship at a highly optimistic $44/kg launch cost. His conclusion: space-based computing is at least an order of magnitude more expensive per GPU-year than a terrestrial data center. The physics doesn't care about Silicon Valley hype.

The Physics Tax on Cooling

Space has no atmosphere, so conduction and convection are off the table. The only heat rejection mechanism is radiation, governed by the Stefan-Boltzmann law: radiated power ∝ area × T⁴. The only variable you can control is area. For a 700-watt H100 held at 60°C, the math demands 1.4 m² of radiator perfectly facing the 3-kelvin void. Put 32 GPUs in a rack drawing 40 kW—now you need 80 m², roughly a pickleball court. Scale to a 100-MW data center and you need 2,500 of those radiators.

And it gets worse. Low Earth orbit bathes radiators in UV light and atomic oxygen. Over a 5-year lifespan, emissivity degrades. Philpot's model shows the required area per chip jumps from 1.4 m² to 2.0 m²—a 40% penalty baked into launch mass and drag.

Radiation: The Silent Silicon Killer

Rad-hard processors are too slow for modern LLMs. Orbital data centers must use terrestrial chips like Nvidia H100s or Google TPUs. High-energy particles flip bits and cause latch-ups. Shielding adds mass. The alternative is software-defined resilience: fly three or more copies of the same calculation and vote. That means a fraction of your compute is always redundant, further inflating cost per useful operation.

Niche Missions That Justify the Cost

Training Llama 3 in space doesn't make economic sense today. But two applications do. Earth-observation satellites with hyperspectral and SAR sensors generate hundreds of terabytes daily—too much to downlink through congested RF pipes. Process data in orbit and downlink only the insights. Second, collision avoidance. Starlink already executes a maneuver every two minutes, but most processing still happens on the ground. In a future with 17,000+ satellites, the observe-orient-decide-act loop must close onboard in milliseconds.

Philpot points to two engineering paths forward: origami-inspired fold-out radiators like James Webb's, and liquid-droplet radiators that spray coolant oil directly into the vacuum. The latter sounds like science fiction, but at megawatt heat loads it may be the only way to cheat the exponential mass penalty.

The real constraint in orbital data centers isn't the silicon—it's the thermodynamics. The winners will be the systems architects who most cleverly accommodate the Stefan-Boltzmann law.

Source: Why Orbital Data Centers Are Harder Than Silicon Valley Thinks
Domain: spectrum.ieee.org