Space Is a Thermos,
Not a Freezer.
Everyone selling AI data centers in space is right about the power and wrong about the heat. In a vacuum, the only way to get rid of a gigawatt of waste heat is to glow — and the physics of glowing is brutal.
The pitch is almost too good to argue with
A pitch is circulating right now, and its appeal is obvious. You take the data centers that are eating the world’s electricity, and you move them to orbit. Up there the Sun never sets, so the panels run around the clock. There’s no land to buy, no water to drain, no permitting delays. And space is cold — about 2.7 kelvin, roughly −270 °C, colder than anywhere on Earth — so cooling, the thing that makes data centers such a nightmare to build, should come for free.
The fantasy now has hardware. In November 2025, a startup called Starcloud put a 60-kilogram satellite into orbit carrying five GPUs — including the first data-center-grade Nvidia H100 ever flown — and trained a small model in space. It’s valued at around a billion dollars now, the fastest company in Y Combinator’s history to get there. SpaceX, in the S-1 it filed in May 2026, named “orbital AI compute” as a core growth story and floated a target of launching 100 GW of it a year. Jeff Bezos told CNBC that gigawatt-scale data centers in space are coming within ten to twenty years — though he couldn’t resist mocking Elon Musk’s timeline: “If you want it to be six years, say it’s three.”
I think the energy argument is basically right. The cooling argument is backwards — and it’s the one that decides whether any of this works.
Here’s the whole post in one sentence: space is not a freezer but a thermos — the device we invented to stop heat from moving — and that one fact is the hardest, most underrated problem in the orbital-compute pitch.
The cold of space is a rounding error
Nearly everyone gets the same thing wrong here, and it’s the most natural mistake to make: space is cold, therefore cooling is easy. The first clause is true. The second does not follow.
Here’s the one equation worth carrying through the whole piece. A warm object radiates heat at a rate proportional to the fourth power of its temperature — the Stefan-Boltzmann law. Fourth power. That exponent is the entire story: double a surface’s temperature and it sheds sixteen times the heat.
A radiator running at a survivable 77 °C (350 K) throws off about 850 watts per square meter. The 2.7 K background of deep space throws back about three-millionths of a watt per square meter. The ratio is roughly one part in 280 million. In other words, how cold space is barely enters the equation at all. What sets your cooling rate is how hot your own radiator runs — 350 to the fourth power utterly swamps 2.7 to the fourth. The cold of space is a rounding error.
A vacuum is a wonderful thing to put around coffee, and a terrible thing to put around a megawatt.— Vish Nandlall, telecom engineer, on orbital data centers
So why does cooling in space feel like it should be easy? Because on Earth we almost never cool by radiation. We cool by convection and conduction — we move heat into air or water and let that fluid carry it away. A vacuum has no fluid. Conduction and convection simply stop working, because there’s nothing to conduct or convect into. A satellite, as one ex-NASA thermal engineer put it, “is an excellent thermos bottle.” That’s not a flaw you can engineer around. That is what a vacuum is — the same reason the flask in your kitchen keeps coffee hot for hours: the emptiness around it kills every heat path except the slow one.
In orbit, that slow one — glowing, radiating infrared light, the same heat you feel off a campfire — is the only path you have left. And a gigawatt data center is, before it’s anything else, a gigawatt heater: essentially every watt of electricity it draws comes back out as heat. Your only way to get rid of it is to make a giant surface glow.
There’s exactly one lever that helps, and everyone reaches for it: run the radiator hotter. Because of that fourth power, pushing a panel from 300 K to 400 K cuts the area you need to about a third. The trouble is the lever is mostly bolted down. Silicon stops working reliably much past 100 °C, and the second law of thermodynamics insists your radiator be colder than the chip it’s cooling. So you’re stuck running radiators in the lukewarm range, where they’re least effective.
The real problem isn’t temperature. It’s flux.
Once you accept that you can only cool by glowing, the central problem of orbital computing comes into focus — and it has nothing to do with how cold space is. It’s about flux: heat per unit of area.
A modern AI chip is a ferocious little concentrator of heat. An Nvidia H100 burns about 700 watts under a postage stamp of silicon — roughly 86 watts per square centimeter. A radiator glowing at a survivable temperature sheds about 0.06 watts per square centimeter. Line them up:
- A GPU die: ~86 W/cm²
- A space radiator: ~0.06 W/cm²
That’s a gap of more than a thousand to one. The chip dumps heat about a thousand times more densely than a radiator can shed it — which means, unavoidably, your cooling surface has to be roughly a thousand times the area of your chips.
And here’s the part the “go to space” pitch quietly skips: that ratio is identical on the ground and in orbit. Stefan-Boltzmann doesn’t care what altitude you’re at. The chip concentrates heat into a postage stamp; the radiator can only release it across a tennis court — on Earth, in orbit, anywhere. Space doesn’t fix the physics. It just hands you somewhere to unfold the tennis court.
How much radiator does a gigawatt need?
This is the whole cooling problem in one equation. Drag the dials and watch the fourth power do its work: nudge the temperature up and the radiator collapses; let it drift toward room temperature and it explodes.
The sobering calibration: the entire International Space Station — the largest structure humanity has ever flown — rejects about 70 kilowatts through ~420 m² of radiator. A single rack of Nvidia’s GB300 systems draws roughly twice that. The whole space station, thermally, can’t keep half a server rack alive. (A purpose-built hot radiator does better than the station’s cold ammonia loop — but “better” still lands you at square kilometers and thousands of tons per gigawatt.)
The sky you’re dumping heat into isn’t empty
There’s a second assumption hiding in the pitch: that a radiator in space sees nothing but cold, empty sky in every direction. In low Earth orbit — which is where every one of these demos actually flies — that’s just false.
The Earth is right there, filling roughly a third of your view, and it is not cold: it glows with about 240 watts per square meter of its own infrared, plus the sunlight it bounces back at you. A radiator face pointed at Earth isn’t dumping heat into the void; it’s being warmed by a planet. And the Sun — the entire reason you came up here — is a 1,360-watt-per-square-meter blowtorch that will happily cook any radiator that turns to look at it.
This creates a genuinely funny geometric problem. Your solar panel wants to stare straight at the Sun, to collect every watt it can. Your radiator wants to do the exact opposite: turn edge-on to the Sun and hide from it, so it isn’t soaking up the very heat it’s trying to shed. The two halves of your spacecraft want to point in directions ninety degrees apart.
The orientation war
The panel wants the Sun; the radiator wants to duck it. Rotate the spacecraft and watch power and cooling fight. Then flip to the fix everyone actually uses — and see what it costs.
Why this matters: bolted to one surface, the Sun you came to harvest is the Sun that cooks your radiator — there’s no angle where both win. The real fix, the one the space station uses, is a separate motor that holds the radiators edge-on while the panels track the Sun. It works. It’s also one more giant joint, spinning for years, that nobody can fly up and repair.
On Earth, you rent a planet-sized radiator for free
Step back and ask why cooling feels free on the ground, because the answer reframes the whole comparison. When a terrestrial data center blows hot air out of a cooling tower or pipes warm water into a river, the air and the water are not the heat sink. They’re a conveyor belt. They carry the heat to the actual sink — the planet — and the only way the planet itself gets rid of heat is the same way your orbital radiator does: by glowing into space.
Here’s the thing about that planetary radiator. The Earth radiates about 122 petawatts to space — 122 million billion watts — from a surface of 510 million square kilometers, and it does it for free: zero mass cost, gravity-anchored, pre-installed. A ground data center doesn’t avoid radiative cooling. It just rents Earth’s enormous radiator and lets the atmosphere shuttle heat over to it.
So going to space doesn’t remove a step from cooling. It removes the planet. You give up the free conveyor belt (no air, no water) and the free radiator (now you build and launch your own), and in exchange you get cheaper electricity. That’s the actual trade. Not “free cooling” — the opposite of free cooling. (The Sun really is worth maybe five times more up there: constant and unfiltered. But that’s an energy win, not a cooling win, and the two keep getting quietly swapped.)
This is where the gap between the demo and the dream gets hard to wave away. Today: one 60-kilogram satellite, one H100, a toy model. The vision: SpaceX’s 100 gigawatts a year. To loft that much hardware would take something like 450 times the entire mass the world launched into space in all of 2025. Amazon’s cloud chief, Matt Garman, put it flatly: there “are not enough rockets” to launch a million satellites, and it isn’t economical.
To its credit, the community building this is fairly honest about where the argument actually sits. Strip away the noise and the whole debate reduces to a single number: the cost to put a kilogram into orbit. It needs to fall by roughly 10×, to around $200/kg, for orbital compute to pencil out — somewhere in the mid-2030s, if Starship delivers. Notice what that means. The physics isn’t really in dispute anymore. Cooling is hard, energy is good, and the open question is logistics. As one analysis put it, going to space optimizes the cheapest part of a data center — the electricity — while wildly inflating the most expensive part: the hardware, and the rockets to get it there.
You can beam the power up. You can’t beam up a technician.
Suppose you solve all of it. You build your square kilometer of radiator, you gimbal it away from the Sun, you get the launch cost down. You still have to keep the thing running for years — and this is where orbital cooling stops being merely expensive and starts being genuinely dangerous.
Cooling is the one subsystem on an orbital data center that is both the biggest physical target and the most fragile. It’s square kilometers of thin panel — the largest thing on the spacecraft — which makes it the most likely thing to be punctured by micrometeoroids and orbital debris. And it’s plumbing: pumps, valves, pressurized coolant in long thin tubes, every one of which can wear out or fail. The space station’s radiators have already been hit. On the ground, you send a technician. In orbit, over a multi-year mission, a puncture in a coolant line isn’t a risk you might run into. It’s a schedule.
You can deliver energy to a satellite many ways — that’s the easy part. You cannot deliver a pair of hands. And the asymmetry compounds: a ground GPU that goes obsolete in five years gets pulled, resold, and recovers 60–70% of its value. An orbital GPU is fused to its radiator and its solar wing, slowly degraded by radiation, and at end of life it burns up in the atmosphere. You launch a fortune and you get back nothing.
That’s the real shape of the orbital cooling problem. It isn’t that physics forbids it — it doesn’t. It’s that physics makes the cheapest input, power, a little cheaper, while making the most expensive input — hardware you can never touch again — far more expensive.
A gigawatt in orbit is a small star
I’ll end on the part I think is genuinely beautiful, because it turns the whole problem into a piece of physics rather than an opinion.
Strip away the silicon and the marketing and ask what an orbital data center actually is, thermodynamically. It’s a body that takes in concentrated energy and radiates it back out as heat. That is the definition of a star. A small, artificial, infrared star, whose fuel is a power cable and whose glow is its own waste heat.
You can even compute its color. A one-sided surface basking in sunlight at Earth’s distance from the Sun, shedding heat only from its lit face, settles at about 120 °C. That isn’t a coincidence I’m reaching for — it’s the same Stefan-Boltzmann balance that fixes the temperature of a hypothetical Dyson sphere, the megastructure a civilization might wrap around its star to capture all of its energy. Your data center’s radiator and an alien Dyson sphere glow at the same temperature, set by the same equation, separated only by scale.
What temperature does sunlight alone set?
A sun-facing surface that can only cool by glowing settles at a temperature fixed by one thing: its distance from the Sun. Your data center’s radiator and an alien Dyson sphere obey the same equation. Drag the collector and find the temperature.
The coincidence that isn’t one: at Earth’s distance a sun-facing collector settles near 120 °C — whether it’s your data center’s skin or a star-swallowing Dyson sphere. Same law, scaled eighteen orders of magnitude. (Glow color is illustrative; a real 400 K surface radiates in the infrared, invisible to the eye.)
And here’s the kicker inside the kicker. In 1960, Freeman Dyson published a short paper in Science proposing that we could detect advanced alien civilizations precisely by their waste heat — the unavoidable infrared glow of computation and industry at scale. The same thermodynamics that makes cooling the hard problem of building a data center in space is the thermodynamics that would make that data center, scaled up far enough, visible across the galaxy.
You can encrypt the data. You can hide the model weights. You cannot hide the heat. Anything that thinks, at scale, has to glow — and getting rid of that glow, not finding the power to feed it, is the problem that will decide whether the computers ever really move to space.