Space is cold. That makes cooling a data centre in orbit sound easy: point the servers at a 3 kelvin void.
The trouble is that cold is a sink with nothing to carry heat to it. A vacuum has no air or water, so every watt a server consumes becomes waste heat with one escape route: infrared radiation. That route is weaker than terrestrial conduction and convection, and it costs surface area, mass and launch.
That second payment is the physics tax: the geometric penalty of shedding heat by radiation alone. Once you see the tax and the Stefan-Boltzmann law behind it, you can test any vendor’s number against a radiator-area-per-megawatt figure. The economics of orbital data centres are being priced now, and a model that treats a 3 K sky as free cooling is wrong before the first launch.
Why is cooling in space so much harder than cooling a terrestrial data centre?
On Earth, waste heat leaves a chip by conduction into a heatsink, then by convection into moving air or chilled liquid. Facilities measure that overhead as PUE, which sits around 1.09 to 1.4 in efficient sites.
In orbit there is no atmosphere, so both mechanisms vanish and only infrared radiation remains, shedding less heat per square metre than moving air or water. The design problem turns from rejecting heat cheaply into finding enough surface area, and the work falls on the spacecraft itself, with the orbital environment layered on top.
Why doesn’t the cold of space make orbital data centre cooling free?
Space sits around 3 kelvin, but that cold is a property of the sink, and nothing carries heat to it. The facility has to emit the heat itself as infrared, and the radiators that do the emitting are the part you pay for.
A vacuum insulates like a thermos: it defeats conduction and convection. One white paper estimates a two-sided radiator held at 20°C emits about 633 watts per square metre, roughly a thousandth of what water cooling removes per square metre.
That is the confusion behind “free cooling”. The sink is free and cold, but the mechanism that reaches it, radiator surface, carries a cost in area, mass and launch. Sunlight, Earth albedo and infrared loading add to that load. Emission is the only route left, so here is how it works.
How does radiative cooling actually work in a vacuum?
Any warm surface emits infrared photons that carry energy away, no medium required, which is why radiation keeps working where conduction and convection cannot.
The rate is set by the Stefan-Boltzmann law, P = εσAT⁴. Power scales with emissivity, area, and the fourth power of temperature. Area is one lever; temperature is the other, and doubling absolute temperature raises emission sixteenfold.
Emissivity is set by the radiator’s coating. AZ-93 white paint, a common spacecraft coating, has an emissivity of about 0.92, and ultraviolet light and atomic oxygen degrade those coatings over a mission’s life. That surface is a deployable radiator, hardware you have to launch from Earth.
How does the Stefan-Boltzmann law set the radiator area a 1 MW orbital data centre needs?
At realistic temperatures each square metre sheds little, so area has to be large. A single 700 W GPU held at 60°C needs about 1.4 square metres, and a 32-GPU rack, once you add CPUs, memory and networking, draws about 40 kW and needs an 80 square metre radiator.
SpaceX’s AI1 is the most concrete public datapoint: roughly 160 square metres of deployable radiator for a 250 kW payload. Scale that toward a megawatt and you land around 640 square metres.
Published benchmarks span roughly 400 to 2,500 square metres per megawatt. The spread is all assumptions: the 20°C radiator from earlier is one point on the curve, temperature, emissivity and margin all move the number, and running hotter buys area back up to the ceiling set by chip leakage and error rates. Radiators and solar arrays together are often 65 to 70 percent of satellite mass, which is how the heat problem becomes a launch cost problem.
What is the “physics tax” of computing in space?
The physics tax is the cumulative penalty of rejecting heat by radiation alone: extra radiator mass, more complex thermal management, constrained power budgets, and a ceiling on compute density. It is the answer to the question people actually ask: what’s the catch?
The ISS needed about 325 square metres to remove 70 kW, at a historical cost of $340 to $500 million. That is the shape of the problem: heat has to be radiated, and radiation charges for surface area.
Follow the chain: waste heat becomes radiator area, then mass, then drag and station-keeping, then launch cost, then ownership cost. That chain is what drives the 20-year cost premium covered elsewhere.
What does the physics tax mean for whether orbital compute actually makes sense?
Radiative cooling is solvable. The tax is structural: it sets a floor on cost and a ceiling on compute density. That is why parity hinges on two numbers, radiator area per megawatt and launch cost per kilogram.
The contrarian view matters. Radiators are a modest share of total mass next to the solar arrays a facility needs anyway, and cooling is unlikely to be what kills the idea. Where a terrestrial site reports PUE, an orbital site reports radiator mass per megawatt, and both are permanent overhead.
For your business the test is simple: both numbers have to fall together. Launch cost could drop toward $100/kg, and radiator area per megawatt has to come down with it. The full cost comparison and the build-versus-buy piece cover the details.
Cooling is the load-bearing constraint on orbital data centres, and the physics tax is why orbital compute carries a premium. Every watt is paid for twice: once in power, once in radiator area, and that second payment never goes away. See what orbital data centres are and the wider cluster for where this sits.
Frequently Asked Questions
Is space actually cold, or is it just empty?
Deep space sits near 3 kelvin, so the background is genuinely cold, but cold is a property of the sink, not of the surroundings a spacecraft feels. The vacuum is also an almost perfect insulator, which is why heat cannot drain away by touch. A surface in sunlight actually runs hot, and the only cooling route left is radiating infrared from its own warm surfaces.
Can an orbital data centre dump heat into Earth’s shadow to cool down?
Not directly. Earth’s shadow removes incoming sunlight, but a radiator still sheds heat only by emitting infrared, so the shadow cannot absorb the heat for it. Passing through eclipse helps by cutting solar loading, which lets surfaces run cooler, but the vacuum still limits rejection. The real cost of eclipses is repeated thermal cycling across the hardware.
Would running the chips hotter shrink the radiators?
Yes, dramatically. Because radiated power scales with the fourth power of absolute temperature, a modest temperature rise cuts the area a radiator needs by a large factor. The catch is that hotter silicon leaks more current and raises error rates, so operating temperature is capped by reliability and power efficiency, not by thermal ambition.
Could you just vent coolant into space instead of using radiators?
Open-loop cooling does exist: spacecraft can boil off water or ammonia through sublimators to dump heat. But that approach consumes a finite supply of coolant and cannot run for years. A megawatt-scale data centre must close the loop and reject heat continuously, which means radiators rather than consumables.
Do the solar arrays add to the heat the radiators must reject?
Yes. Solar arrays convert only a fraction of the sunlight they capture into electricity; much of the rest becomes heat that the thermal system must manage. Arrays and radiators together are often 65 to 70 per cent of a satellite’s mass, which is why the thermal problem and the power problem are really the same problem.
What happens if a radiator is struck by orbital debris?
A puncture can disable a panel or, worse, breach a coolant loop and take out the whole thermal system. Designs manage this with redundancy: multiple independent radiator panels, isolation valves and protective standoff layers. Losing some surface area cuts rejection capacity, so spacecraft carry margin, which adds mass and cost.
Do radiator coatings lose performance over a mission?
They do. Ultraviolet light, atomic oxygen and thermal cycling gradually degrade the emissivity of a radiator’s surface, so it radiates less heat at the same temperature as the years pass. Engineers choose tough coatings and then add design margin up front, which means the radiator is oversized from day one to stay adequate at end of life.
Why do two orbital data centres with the same power need different radiator areas?
Radiator area depends on more than wattage. Operating temperature, surface emissivity, coating degradation, orbit, view factors to Earth and the Sun, and how much design margin is carried all move the number. That is why published benchmarks for orbital compute span roughly 400 to 2,500 square metres per megawatt: they are not describing the same assumptions.
How much launch mass does all that radiator area add?
A great deal. Radiators and solar arrays together are often 65 to 70 per cent of a satellite’s mass, and every extra kilogram must be lifted to orbit. Thermal hardware therefore lands directly in the launch budget and the twenty-year cost of ownership, which is why radiator area per megawatt sits alongside dollars per kilogram as a decisive figure.
Do orbital data centres still use liquid cooling internally?
Yes. Chips still need a pumped coolant loop or heat pipes to move heat from the silicon to the radiator surface, because the vacuum only changes the final step, not the internal one. Inside the sealed spacecraft, conduction and pumped convection work normally. The physics tax begins at the moment that heat has to leave the outer surface.
How can I check whether a vendor’s cooling claim is credible?
Ask two questions: what radiator area per megawatt does the design assume, and at what operating temperature? Then reconcile the answer against the Stefan-Boltzmann law. A claim that leans on free cooling from the cold of space is a red flag, because the sink is free but the surface area and mass that reach it are not.
Does the physics tax mean orbital data centres can never make sense?
No. The tax is structural, not fatal. Radiative cooling works, radiators are a tractable engineering problem, and their mass is comparable to the solar arrays a facility needs anyway. The tax still sets a floor on cost and a ceiling on compute density, so parity depends on both radiator area and launch cost falling together.
