The Physics Tax on Orbital Data Centers: Why Space Computing Is Mostly a Cooling Problem

Jensen Huang told the crowd at Nvidia GTC in March that "space computing, the final frontier, has arrived." The pitch has since hardened into a real spending category. SpaceX absorbed xAI and is sketching a constellation of orbital data centers. Google announced Project Suncatcher with Planet, aiming to fly two satellites carrying Tensor Processing Units by early 2027. The startup Starcloud filed an FCC proposal for an 88,000-satellite constellation.

The sales deck writes itself: abundant solar power, free cooling in the cold of space, and no earthquakes, floods, or protesters showing up at the gate. Andrew Cavalier, an aerospace analyst at ABI Research, took the physics seriously and arrived at a much less flattering picture in a feature for IEEE Spectrum. His central finding: running an Nvidia H100 in orbit for a year costs at least an order of magnitude more than running the same chip in a terrestrial data center. And the bottleneck that drives that number is not the rocket. It is heat.

What's claimed versus what physics allows

Start with the line that gets repeated most often, that space is a free cooling solution. Space is cold, yes, but it is also empty. The two heat-removal mechanisms data centers rely on, conduction and convection, both need matter to carry energy away. In a vacuum, both are gone. The only path left is thermal radiation, and radiation is governed by the Stefan-Boltzmann law: the power a surface can shed scales with its area times the fourth power of its temperature.

For a systems architect, that fourth-power term is unforgiving in a specific way. You cannot meaningfully crank temperature without cooking the silicon, so the only knob you actually control is area. Cavalier calls the result a "physics tax": every additional watt you need to reject demands more radiator surface that you paid to launch and now have to point at deep space.

The numbers make the tax concrete. A single 700-watt H100, held at a GPU-friendly 60 °C while facing a 3-kelvin background, needs roughly 1.4 square meters of radiator. Scale that to a normal AI rack, about 32 GPUs across four server boards plus CPUs, memory, and networking, and you are dissipating around 40 kilowatts. Cooling that one rack in vacuum takes an 80-square-meter radiator, roughly a pickleball court. A 100-megawatt facility, modest by current hyperscale standards, would need on the order of 2,500 of those.

What's actually new in the engineering response

The interesting part is not the complaint, it is how the industry is trying to wriggle out from under the area requirement. Two approaches are getting real attention.

The first is origami-style deployable radiators, descendants of the folding sunshield on the James Webb Space Telescope. High-conductivity composite panels pack into a tight volume for launch, then unfurl into large, lightweight thermal wings once on orbit. This buys area without buying proportional launch volume, which is the right thing to optimize.

The second is stranger and more interesting: liquid-droplet radiators. Instead of carrying a rigid structure at all, the spacecraft sprays a stream of coolant oil straight into the vacuum. Each droplet radiates from its full surface as it drifts across an open loop, and a collector catches the fluid downstream and pumps it back. It reads like science fiction, but once heat loads climb into the megawatts, exposing bare droplets may be the only way to dodge the mass penalty that rigid panels impose.

Limitations the brochures skip

Cooling is only the first of three compounding problems, and each one degrades over time rather than staying fixed at the optimistic launch-day value.

Radiation eats the radiator. In low Earth orbit, ultraviolet light and atomic oxygen chemically attack thermal coatings. Over a typical five-year lifespan the surface loses emissivity, and Cavalier's model shows the required area per chip climbing from about 1.4 to nearly 2.0 square meters just to hold the same temperature. That is a 40 percent increase in radiator mass you must launch up front, plus extra atmospheric drag, simply to survive the coating's decay.

Radiation also eats the silicon. This is the part that should worry anyone planning to run real workloads. Genuinely radiation-hardened processors exist and are reliable, but they are slow and expensive, and no rad-hard chip can run a modern large language model. To get useful compute density, these constellations have to fly the same commercial H100s and TPUs found in terrestrial racks, which are soft targets. Cosmic rays flip memory bits and can trigger latch-ups that destroy logic. The mitigation borrowed from edge-computing practice, including systems on Artemis II and the HPE Spaceborne Computer on the ISS, is brute-force redundancy: run the same calculation on three or more nodes, compare answers, and reboot whichever one disagrees. That works, but it means a meaningful slice of your launched compute exists only to check the rest of it.

The power side mirrors the heat side. Sunlight in orbit delivers a generous 1,361 watts per square meter, but space-grade panels degrade 1 to 3 percent per year, and the area math is brutally symmetric. Generating power runs around 400 W/m² and rejecting the resulting waste heat runs around 450 W/m², so every square meter of solar array demands roughly another square meter of radiator. The radiator is not a coating you tack onto an existing surface, it is a structural co-equal of the power system. Keeping panels aimed at the sun, radiators aimed at the void, and antennas aimed at Earth all at once requires high-torque attitude control with plenty of failure modes, in an environment where sending a repair crew is not an option.

Where the economics actually close

None of this means orbital compute is pointless. It means general-purpose AI training and inference, the headline use case, is the wrong fit. The applications that justify the physics tax are the ones where the data is already in orbit and the cost of moving it down is the real constraint.

The clearest example is Earth observation. Hyperspectral and synthetic-aperture-radar satellites can generate hundreds of terabytes per day, and the radio-frequency downlink pipes plus ground stations cannot absorb that volume of raw data. Processing on orbit and sending down only the extracted insight turns an impossible bandwidth problem into a tractable one.

The second is collision avoidance. There are over 17,000 satellites up there, most in low Earth orbit, and the Kessler-syndrome risk of a cascading collision chain is no longer theoretical. SpaceX reports that Starlink executes an avoidance maneuver every two minutes on average, with onboard AI flagging threats but most of the heavy computation still happening on the ground. In a megaconstellation, that ground loop does not scale. The observe-orient-decide-act cycle has to collapse from minutes to milliseconds, onboard, and the standard satellite flight computer cannot run the probability models that requires. The economic case for space compute, then, is not to relocate the cloud. It is to put high-performance processing next to the sensors that are already there.

The pattern underneath

A fair objection is that all the input numbers improve over time: launch cost, chip cost, panel efficiency. Cavalier's model already uses optimistic figures, including a $44-per-kilogram Starship launch cost, and he grants that hardware will get cheaper. But the binding constraint is not launch price. Radiators and solar arrays already eat 65 to 70 percent of satellite mass, space-grade photovoltaics cost orders of magnitude more than terrestrial panels, and today's high-efficiency cells lean on germanium substrates whose supply is concentrated in China. A shift to radiation-tolerant perovskite cells could move the math, but that is five years out at best.

The deeper point connects orbit back to the ground. The same wall that limits a satellite, how much heat you can reject per unit of hardware, is the wall slowing down hyperscale facilities in Northern Virginia, where power density and cooling, not chip availability, increasingly gate new construction. The constraint that matters for compute at scale has quietly stopped being the silicon and started being the thermodynamics. Space just makes that lesson impossible to ignore, because in a vacuum there is nowhere for the heat to hide.