🔬 Space Is Freezing. Cooling a Computer There Is the Hardest Part.

Space Is Freezing. Cooling a Computer There Is the Hardest Part. — Woody Magazine

Last Friday, in the largest public offering in history, SpaceX listed on the Nasdaq — and tucked into the fanfare was a single schematic. It showed a satellite called AI1, and its job is not to relay signals but to run artificial-intelligence workloads. A data center, in orbit. The intuition almost writes itself: space sits hundreds of degrees below zero, so surely it's the perfect icebox for a rack of overheating chips. It is exactly backwards. The hardest problem an orbital data center has to solve is getting rid of the heat.

The difficulty isn't temperature; it's transport. On Earth, a fan's draft or a loop of chilled water physically carries heat away — air and water are the couriers. A vacuum has no couriers. The only exit left is radiation: a warm object shedding energy as infrared light, glowing itself cool at wavelengths the eye can't see. And as cooling goes, radiation is painfully slow.

So how does radiative cooling actually work?

The chain runs like this. An AI chip dumps hundreds of watts onto a die the size of a fingernail. A cold plate pressed against it takes up that heat; a pumped coolant loop, or a web of heat pipes, ferries the heat out to a broad radiator panel; and only there does the surface finally throw it into space as infrared. The International Space Station works the same way, rejecting heat through ammonia loops and rotating radiator wings turned away from the sun.

The thing to grasp is that the problem isn't where the heat goes, but how fast. Deep space is cold enough — about −270 °C, a mere 2.7 kelvin above absolute zero. But radiation itself is slow, and in low orbit it's slower still: the panel also catches infrared glowing off the Earth and sunlight bouncing back, so its destination is never as frigid as that headline number implies.

The rate is governed by a law from the 1880s, Stefan-Boltzmann, which holds that radiated power climbs with the fourth power of a surface's absolute temperature. The practical upshot: a slightly hotter radiator sheds dramatically more heat from the same area. Holding a single 700-watt Nvidia H100 at 20 °C takes nearly 3 square meters of radiator; let the chip run at 85 °C and that falls to about 1. Which is why the industry is racing to build chips that tolerate running hot — though push too far and lifespans collapse, so 60 °C tends to be the truce.

Fig. 1 — A chip concentrates heat on a fingernail-sized die. A coolant loop carries it to a broad radiator, which radiates it to space as infrared. The radiator must point at the shade, away from the sun — in Musk's phrase, "the best radiator never sees the sun." (Solar panels do the opposite, facing the sun; that's a separate flow.)

Slow means large, and large means heavy. The ISS sheds 70 kilowatts of waste heat through radiator wings weighing some seven metric tons. Scale that hardware to a modest one-megawatt data center (a thousandth the size of a terrestrial hyperscale campus), and the radiators alone come to around 100 tons. That is roughly ten times the weight of the computers doing the actual work.

~10×

For a 1 MW orbital data center, the radiators needed to dump its heat may weigh about ten times as much as the computers themselves (~10 tons), on ISS-class estimates. Cooling is mass — and mass is launch cost.

Space is vast. But is there room to unfurl all that metal?

Start with the arithmetic. A one-megawatt radiator runs somewhere between four tennis courts and a hockey rink — call it 1,200 to 1,600 square meters. A gigawatt-class campus, the kind now being planned on Earth, is a thousand times larger. Multiply it out, crudely, and the radiator alone would cover 1.6 square kilometers — more than 200 football fields. "Space is big," you might say. But the constraint isn't empty volume. It's three other things.

First, unfurling. A rocket fairing is a narrow cylinder a few meters across. Hundreds or thousands of square meters of radiator have to fold inside it and then deploy in orbit — which is why engineers are developing origami-style panels, in the spirit of the James Webb Space Telescope, that pack tight and unfold huge.

Second, mass. One analysis puts the radiators for a 10-megawatt facility at 200 to 1,000 tons. At that weight, the economics may not close no matter how cheap launch becomes.

Third, real estate. The orbits that receive continuous sunlight are limited, and low Earth orbit is already crowded: roughly 15,000 satellites and 30,000 tracked pieces of debris circling at 7 kilometers a second. SpaceX's long-range plan would add up to a million data-center satellites — tens of times the entire current population.

Starlink alone performed 145,000 collision-avoidance maneuvers in the first half of 2025. A sprawling radiator is, by definition, a bigger target. And the runaway chain reaction of debris striking debris — the Kessler syndrome that NASA's Donald Kessler warned of in 1978 — is no longer hypothetical. Floating thousands of square meters of structure into that traffic is a liability all its own.

Fig. 2 — A radiator folds origami-tight into the narrow fairing and unfurls in orbit. A single one-megawatt unit already spans a hockey rink; a gigawatt campus would need a thousand times more. The harder question isn't finding empty room — it's the collision-and-debris burden of lofting giant structures into an already-crowded orbit.

So why go up at all?

Because Earth can no longer feed two of AI's appetites: power and cooling water. U.S. data centers drew 183 terawatt-hours in 2024 — more than 4 percent of the country's electricity, roughly the entire annual demand of Pakistan. By 2030 the IEA expects that to swell 133 percent, to 426 terawatt-hours. The hunger runs so deep that Microsoft has agreed to revive a reactor at Three Mile Island — the site of America's worst nuclear accident, in 1979 — restarting a separate unit that had closed in 2019, to feed its AI. Orbit, by contrast, offers sunlight with no clouds and no night: the same panel yields several times more energy than it would on the ground.

Underneath all of it sits one number: the cost of launch. Radiators and solar arrays are heavy, and in space a kilogram is money. For most of the space age, lofting a kilogram to orbit ran about $10,000. Reusable rockets cut that by an order of magnitude, into the low thousands — and Starship aims lower still, in the low hundreds. "Land the first stage and fly it again" is a single sentence, yet no one pulled it off for half a century, and that one feat reopened the ledger on plans long shelved as too expensive. The dream of making power in space isn't new: the engineer Peter Glaser proposed it in 1968, and the U.S. Department of Energy and NASA studied and shelved it in the 1970s as technically sound but economically hopeless. What changed wasn't the idea. It was the foundation beneath it.

And the field is crowded with more than Musk. Google has promised prototype "Project Suncatcher" satellites for 2027; Nvidia-backed Starcloud trained its first AI model in orbit in December 2025. Yet the simplest objection still stands: if the radiators outweigh the computers tenfold, and a 10-megawatt rig needs hundreds of tons of them, why not just lay solar panels across a desert? Even SpaceX, in its own listing prospectus, calls the venture "early-stage" and warns it "may not achieve commercial viability."

The contest won't be won by smarter chips. It will be decided by how much heat a single square meter can shed.

So the fate of the orbital data center rests on neither imagination nor the cleverness of silicon. It rests on how fast hot metal can glow itself cool — the slow law that Stefan and Boltzmann wrote down in the nineteenth century. Whether AI truly moves to orbit will turn on how many tons of radiator a falling launch price can lift. Building the new machine is a genuine feat; so was collapsing the cost that makes it thinkable at all. Both are remarkable. But the final grade, as always, is handed out by thermodynamics.

THE TAKEAWAY

Space is cold, so cooling a computer there sounds easy. It's the reverse: in a vacuum, heat can leave only by radiation, and that rate is chained to the fourth power of temperature. A single megawatt already needs a hockey-rink-sized radiator, and folding that slab, lofting it, and unfurling it into a crowded orbit is what "cooling" really means up there. The real wall isn't power or radiation. It's thermodynamics — and the only reason we get to test it now is that launch finally got cheap.

Sources

이 블로그 검색

Daily Woody

🔬 Space Is Freezing. Cooling a Computer There Is the Hardest Part. — Woody Magazine, Jun. 15, 2026

Space Is Freezing. Cooling a Computer There Is the Hardest Part.

댓글

댓글 쓰기

이 블로그의 인기 게시물

Daily Woody Economy | 2026.04.30 (목) — FOMC 8:4 분열 표결, Powell 시대 끝

📚 Tank Day Was Never Just About a Tumbler — Woody Magazine, May 19, 2026

Daily Woody | May 8, 2026 — Han Gets 15 Years; Yoon's Bench Goes Next