An IEEE Spectrum analysis by an ABI Research aerospace analyst shows the binding constraint on space based compute is radiator area, not launch cost or chip speed, and the next eighteen months of announcements should be read as radiator problems wearing GPU launch costumes.
The Stefan-Boltzmann law is about to ruin a lot of venture theses. In vacuum, the only way to shed heat is radiation, and that forces every orbital data center designer to the same uncomfortable math: an 80-kilowatt AI rack needs a radiator roughly the size of a pickleball court, and every chip you add requires roughly another square meter of radiator area, forever. That is the central argument of an IEEE Spectrum analysis by an ABI Research aerospace analyst, and it reframes every recent orbital-compute announcement from Nvidia, Google, Starcloud, and SpaceX as a radiator-design problem wearing a GPU-launch costume.
The physics starts with a single number. The Stefan-Boltzmann law says a surface in vacuum can only dump heat by glowing in the infrared, and the energy it sheds scales with the fourth power of its temperature. At a 60 °C operating point, a 700 W H100-class chip needs about 1.4 m² of radiator surface, according to the same analysis. Stack 32 of those chips into a standard 80 kW AI rack, add CPU, memory, and networking, and you need roughly 80 m² of radiator per rack, which the author helpfully notes is about the size of a pickleball court. Scale to 100 MW, a reasonable data-center neighborhood, and the radiator count climbs to at least 2,500 court-sized panels, plus the structure, attitude control, and launch mass to hold them in the right orientation.
The other shoe is the near 1:1 solar-to-radiator budget. The solar constant at Earth's distance is about 1,361 W/m², but panels actually deliver closer to 400 W/m² after thermal and packing losses. Radiators in low Earth orbit face the sky and deep space, where the same physics hands back about 450 W/m² of rejection capacity. Those two numbers are nearly matched, which means every square meter of solar generation you add to feed more compute requires roughly another square meter of radiator, on the opposite face of the spacecraft, in the right attitude, in eclipse-free orbits. You cannot trade launch mass for cooling. You trade launch mass for cooling.
Then comes the time tax. LEO radiator coatings degrade under ultraviolet and atomic-oxygen exposure, and the analysis puts the end-of-life penalty at roughly 40%, the radiator area per chip growing from about 1.4 m² to 2.0 m² over a typical five-year satellite life. That is forty percent more radiator mass, drag, and launch volume, launched up front, just to keep the same chip cool on its last day in orbit. Solar panels quietly take a similar hit, losing 1 to 3 percent of their conversion efficiency per year from radiation damage. The hardware that makes the kilowatt does not survive as long as the kilowatt.
Silicon survives the orbit worse than the panels do. Ionizing radiation flips bits in commercial memory, and rad-hard processors are too slow to run modern AI workloads. Operators' only realistic option is to fly commercial GPUs, Nvidia H100s and Google TPUs, and accept that some fraction of them will misbehave, then paper over the errors in software. The analysis notes that the same trio-voting and cluster-redundancy pattern is already used on Artemis II flight computers, SpaceX vehicles, and HPE's edge servers on the ISS. It works. It also means launching three chips to do the work of two, which the radiator math then has to cool and the launch budget then has to lift.
The TCO line, by the analyst's count, is brutal. Even at an optimistic $44/kg Starship launch cost and $0.20/kWh terrestrial power, putting one GPU in space and running it for a year costs at least an order of magnitude more than running the same GPU on the ground. That is the figure that should sit next to every orbital-compute press release, because the same chips in a Phoenix warehouse are competing against cheap power and cheap cooling air, and the orbital version has to pay the full radiator and radiation tax on top.
The industry response is not denial but acceleration. Nvidia used its GTC conference in March, as the analysis reports, to launch a "space computing" initiative, with Jensen Huang calling orbital data centers "the final frontier" of accelerated computing, and to show a Starcloud module built around a Vera Rubin generation Space-1 computer. Google followed with Project Suncatcher, planning two TPU-equipped satellites with Planet in a dawn-dusk sun-synchronous low Earth orbit by early 2027, a configuration that keeps the satellites out of Earth's shadow almost continuously and gives the radiators a stable deep-space view. Starcloud filed with the FCC for an 88,000-satellite orbital data center constellation, and SpaceX, having acquired xAI, has disclosed plans to put data centers in orbit. None of these projects have produced an energy-balance sheet, and the analysis suggests that the ones that succeed will be the ones that treat radiator geometry and duty cycle as the design driver, not the line item.
There are a few mission profiles where the cost premium genuinely closes. Preprocessing Earth-observation data on-orbit, where downlink bandwidth is the real bottleneck, can save more bandwidth than the radiator costs. Real-time hypersonic-missile detection and tracking wants a vantage point that orbital sensors already provide and a latency budget that ground processing cannot. Active collision avoidance, as low Earth orbit fills with tens of thousands of new satellites, needs compute close to the tracking radar. The analysis lists these as the niche applications that could plausibly justify orbital compute at unit economics the bulk market cannot.
The design space that opens up, once heat rejection is treated as the primary constraint, is concrete. Orbits with high eclipse duty cycles force radiators large enough to ride out the dark side. Dawn-dusk sun-synchronous orbits like the one Google picked for Suncatcher give nearly continuous power but trade thermal stability for it. Higher altitudes, like medium Earth orbit, see less atomic oxygen and less drag, but lose the proximity advantage that makes on-orbit Earth observation worth the trouble. Radiator geometry, whether flat panels, deployable wings, or pumped loops that route heat to a single large surface, is the engineering decision that will sort winners from also-rans in the next eighteen months. The press releases will keep counting GPUs. The serious operators will be counting square meters.