The Thermos Trap: Why Cooling AI in Space Comes Down to Launch Cost

Everybody owns a vacuum thermos. Coffee you pour before work is still hot at lunch. That kitchen object is at the center of the loudest argument against AI data centers in orbit, and I think it points the wrong way.

On November 2nd, 2026, a post on X went viral. Jan Broersen, a professor of philosophy, logic and AI at Utrecht University, took aim at Elon Musk's claim that hardware is easier to cool in space. "The opposite is true," he wrote. "A vacuum is the best possible isolator. In space, you can only dissipate heat through radiation." His conclusion: orbital data centers will not happen in our lifetimes.

He is right that radiation is the only way out. Where I think he goes wrong is assuming that way out is narrow.

The replies

Shaun Maguire, a Sequoia partner with a physics doctorate from Caltech, said the real hurdle is lifting mass on Starship. Once you can loft that mass cheaply, he wrote, cooling is passive: a light, high conductivity aluminum structure with an ammonia loop, shedding heat by radiation that scales with temperature to the fourth power.

Musk replied on November 3rd: "Yeah, much easier to cool in space." He said about 1,300 watts per square meter is easily achievable for an aluminum radiator pointed away from the sun, because that panel mostly sees cosmic background radiation about 3 degrees above absolute zero. Solar panels in orbit generate about 250 watts per square meter, he said, so they need about five times the radiator's area, and most of the satellite's mass is solar panels.

Dr. Philip Metzger, director of the Stephen W. Hawking Center for Microgravity Research and Education at the University of Central Florida and a cofounder of NASA's Swamp Works, went straight at the thermos: "Vacuum is definitely not the best possible insulator. We fill vacuum Dewars with granular material like perlite to stop the radiative heat transfer between inner and outer jackets." At GPU temperatures, he added, radiative transfer grows enormously. Musk's reply: "Yup, T^4 scales super fast."

What Earth charges for cooling

Every watt that enters an AI chip leaves as heat. One Nvidia H100 pulls up to 700 watts. Stack 72 modern AI processors in one rack, the kind of liquid-cooled rack filling xAI's Colossus in Memphis, and that cabinet throws off roughly 120 kilowatts. On the ground, that heat goes from cold plates into liquid, then through chillers that burn extra electricity, then out of cooling towers that evaporate water.

The Uptime Institute's 2025 survey of 681 data centers put average power usage effectiveness at 1.54. For every 100 watts reaching the chips, another 54 go to cooling and power overhead. Top hyperscale buildings run near 1.1. Lawrence Berkeley National Laboratory's December 2024 report says U.S. data centers directly consumed about 66 billion liters of water in 2023, more than triple the 2014 figure. At the average site rate of 0.36 liters per kilowatt hour, a 100 megawatt cluster running flat out evaporates around 860,000 liters a day.

Towns are saying no. Memphis paused data center expansion for 135 days after an October 2026 council meeting. New York's Executive Order 62 paused permits for new sites of 50 megawatts and up. One tracker counted 527 local moratoriums in force as of November 3rd, and a September NBC poll found 69% of Americans oppose a new data center near them. That pressure is what pushes the cooling debate toward orbit.

Three ways heat moves

Heat leaves a hot object by conduction (grabbing a hot skillet), by convection (blowing on soup), or by radiation, where every warm object sheds energy as electromagnetic waves, mostly infrared. In a vacuum, conduction and convection to the outside drop to zero. Google's Project Suncatcher team said it plainly in September 2026: "In space, there's no air flow. In a vacuum, you can only diffuse heat via radiators."

The error is treating radiation as weak. Look at how a thermos is built. If the vacuum gap stopped heat by itself, the walls would need no special finish. Instead the inner walls are silvered. Without that mirror, infrared light crosses the gap and cools your coffee.

Metzger's example from NASA cryogenics makes it concrete. Liquid nitrogen is stored at 77 kelvin, minus 196 Celsius, in a double-walled vacuum tank called a Dewar. Even that cold, one square meter of dark surface radiates about 2 watts across the vacuum, so engineers pack the gap with perlite, a crushed volcanic rock, just to block the photons.

Temperature to the fourth power

The Stefan-Boltzmann law sets the rule: radiated power equals emissivity times a constant times area times absolute temperature in kelvin, raised to the fourth power. Raise a radiator's kelvin temperature 10% and its output jumps 46%. Raise it 20% and output more than doubles.

Now compare that nitrogen tank to a radiator at a GPU coolant temperature of 350 kelvin, which is 77 Celsius. That is about 4.5 times warmer on the kelvin scale. Raise that ratio to the fourth power and radiated heat climbs about 427 times. The same dark square meter that leaked 2 watts now puts out about 851. A cold thermos barely glows in infrared. A hot GPU radiator is an infrared floodlight.

Space barely shines back. The cosmic microwave background sits at 2.725 kelvin and returns about three millionths of a watt per square meter. The moon shows the effect at scale. NASA's Diviner instrument measured the lunar equator falling from about 390 kelvin at noon to 95 kelvin before sunrise, and shadowed craters reach 25 kelvin by radiating into space alone. Vacuum blocks air currents. Light crosses it freely.

The space station objection

The sharpest version came on November 3rd from Tom Atos: "The ISS requires 7,000 kg of radiators to dissipate just 70 kW. Meanwhile, a single rack of modern AI processors generates roughly 120 kW." A 1 gigawatt orbital data center, he added, would need the radiator equivalent of 14,000 space stations.

That is a fair challenge, and the station figures hold up. NASA's manual for the station's external active thermal control system describes two pumped ammonia loops of 35 kilowatts each feeding six main radiator wings. Each eight-panel wing measures about 23.3 by 3.4 meters and weighs 1,123 kilograms. Together they give 475 square meters at 6,736 kilograms. That works out to 147 watts per square meter and 96 kilograms per kilowatt.

So how does Musk claim almost nine times that rate? Three differences.

Temperature. The station is a home with astronauts in shirt sleeves, so its loops run cool. A 1997 NASA thermal design study lists an average radiator surface temperature of 288.1 kelvin, 15 Celsius, at 0.90 emissivity. A data center has no crew, and GPU coolant loops run at 60 to 80 Celsius.

Two faces. A flat wing radiates from front and back at the same time. A critic posting as Doug argued that a panel at 80 Celsius radiates only 880 watts per square meter, a figure that fits one face. Run the law for a two-sided panel at 0.90 emissivity facing deep space: 828 watts per square meter at 300 kelvin, 1,213 at 330, and 1,300 at 336 kelvin, which is 63 Celsius. Musk's figure is the textbook output of a two-sided radiator at an ordinary coolant temperature. At 350 kelvin it reaches about 1,534, over ten times the station's rate.

Materials. The station's shuttle-era honeycomb panels weigh about 14 kilograms per square meter. StarCloud CEO Philip Johnston told Ars Technica in July 2026 that those radiators are "expensive and heavy," and that two thirds of his team builds light radiators.

Coatings and pointing

On November 2nd, Andrew Côté made a good point in reply to Maguire: emissivity plus reflectivity equals 1 at a given wavelength, so the worst possible radiator is a polished mirror. That is Kirchhoff's law, and bare polished aluminum reflects infrared instead of emitting it. My read is that Musk's "aluminum radiator" means aluminum structure under an optical coating like NASA's white Z-93.

The trick is wavelength. Sunlight from a 5,800 kelvin star peaks near 0.5 micrometers. A 336 kelvin radiator glows around 10 micrometers. So a coating can look white to sunlight and close to black in thermal infrared. StarCloud's thermal engineering note cites one that absorbs 9% of incoming sunlight while emitting 92% in infrared.

Pointing matters too. At 400 to 650 kilometers, a face aimed at Earth soaks up roughly 200 watts per square meter of planetary infrared. Google's Suncatcher and SpaceX's designs both use a dawn-dusk sun-synchronous orbit, riding the day-night line so the sun stays off to the side nearly all the time. Google says a panel there produces up to eight times the annual energy of the same panel at mid-latitude on Earth. Solar wings face the sun. Radiator wings sit edge-on at 90 degrees, both faces toward deep space.

Radiator versus solar array

NASA's TSIS-1 instrument measures sunlight at Earth orbit at 1,361 watts per square meter. At 18.4% efficiency, an array makes 250 watts per square meter, and every one of those watts ends up as heat. A shaded radiator at 63 Celsius dumps 1,300. Divide and you get 5.2.

For 1 megawatt, that means 4,000 square meters of solar, a square 63 meters on a side, against about 770 square meters of radiator, 28 meters on a side. Under StarCloud's conservative case, 20 Celsius with one face in sunlight netting 633 watts per square meter, the radiator grows to 1,580 square meters, still well under half the array. At 1 gigawatt, it is 4 square kilometers of solar against 0.77 of radiator. ## Hardware already flying

StarCloud-1 launched on November 2nd, 2025 aboard a Falcon 9 on the Bandwagon-4 mission. The 60 kilogram satellite carried the first Nvidia H100 flown in space. Co-founder Adi Oltean said it completed the first on-orbit training run with Andrej Karpathy's nanoGPT and ran Google's Gemma. A year later, Johnston said it is still operating, with two 8 kilowatt, 450 kilogram satellites planned for 2027 rideshares.

Google announced Suncatcher on November 4th, 2025. Its reference design puts 81 satellites within a 1 kilometer radius at 650 kilometers. In a 67 MeV proton beam at UC Davis, its Trillium TPU v6e survived a shielded five-year mission dose of 750 rad with zero permanent failures, and its high bandwidth memory showed irregularities only above 2,000 rad. On October 1st, 2026, a Planet-built 1 kilowatt prototype carrying four Trillium chips launched on SpaceX's Transporter-18. On October 30th, during a pass over Svalbard, engineers sent its first prompt to Gemma 4. It answered with a haiku: "Blue marble below / data streams in silent grace / starlight feeds my core."

Who is lining up next

In China, ADA Space and Zhejiang Lab launched 12 satellites of the Three-Body Computing Constellation in mid-2025 and 8 more on October 14th, 2026, toward a 2,800 satellite plan. Axiom Space launched two orbital data center nodes in January 2026. Jeff Bezos said at Italian Tech Week in October 2025 that giant training clusters "will be better built in space because we have solar power there 24/7." Blue Origin filed Project Sunrise with the FCC for up to 51,600 satellites.

SpaceX filed on January 30th, 2026 for up to 1 million satellites. SpaceX's filing says launching 1 million metric tons a year at 100 kilowatts per ton adds 100 gigawatts of AI compute annually. In June, Musk held up a draft AI1 satellite model at Bastrop that scales Starlink V3 hardware to roughly 120 to 175 kilowatts of sustained compute. On September 28th, Starship Flight 14 reached orbit and deployed 26 Starlink V3 satellites.

The hard parts that are real

Honesty cuts both ways. Radiating from the wing is passive. Moving heat from a thumbnail-sized chip to a 160 square meter wing is plumbing: cold plates, heat pipes or pumped ammonia that cannot freeze or leak when a micrometeoroid hits. The station took radiator face sheet damage in September 2008 and later vented ammonia from one of its loops. Researcher Robert Graham noted on November 3rd that protecting fluid lines across thousands of satellites is real engineering work.

Nobody flies to 650 kilometers to swap a dead GPU. AI chips turn over every one to two years. Satellites fly five years or more.

Then launch cost, the biggest hurdle of all. Maguire put it directly: once you can get huge amounts of mass into space cheaply, "it is genuinely easier to cool things in space." A dedicated Falcon 9 lists around $74 million for up to 22,800 kilograms to low Earth orbit when expended, roughly $3,250 per kilogram. Rideshare runs near $6,500.

The $200 per kilogram line

Google's Suncatcher team ran that learning curve in their November 2025 paper. Their finding: the math flips once launch to low Earth orbit falls to about $200 per kilogram, which they project by the mid-2030s. SpaceX targets $100 to $200 per kilogram and below with a fully reusable Starship. At $200, amortized launch cost roughly matches what a ground data center pays for electricity alone, about $570 to $3,000 per kilowatt per year.

Back to the thermos. A vacuum traps heat when the walls are cold and mirrored. Coat a two-sided radiator so it is black in infrared, warm it to 63 Celsius with GPU coolant, and aim it at 3 kelvin space. The fourth power does the rest: 1,300 watts per square meter into the universe, no cooling tower, no water permit. My read is that physics sides with Musk on cooling. Whether orbital compute scales comes down to Starship's price per kilogram.