A standing premise first: this page is about where heat can go, not about any power source, and no magnetic fusion device has yet produced more fusion energy than the heating put into its plasma. Nothing below depends on one. With that stated, the question is the first one anyone asks a fusion company that points at the Martian dirt. Sunlight in orbit comes without clouds or dust, there is no land to buy, and the chips are already heading up. Why not put the computing in orbit? Because of where the heat goes.
In sunlit orbit the Sun delivers about 1,361 watts to every square metre,1 continuously across the sunlit arc of every orbit, without a cloud, without a dust season. Land is free. Permits are someone else’s problem. If a data centre runs on sunlight, orbit looks like the obvious place for it. We agree with all of that. It is the next sentence where the argument usually goes wrong.
A spacecraft is not cooled by its surroundings — there are no surroundings to carry heat away. No air, no convection, nothing to blow the warmth off a hot chip. The only exit for heat is radiation: a surface glowing infrared into the dark. And radiated power rises with the fourth power of temperature, which means a cool radiator barely works at all.
Per square metre of radiator with a typical spacecraft coating:3 ≈ 369 W at 20 °C, ≈ 776 W at 80 °C.15
Raising the panel from room temperature to 80 °C more than doubles what it sheds. In vacuum, radiator temperature is the strongest lever there is. The question in orbit was never “is it cold enough.” It is “how big a glowing panel, how hot, and pointed where.”
A radiator sheds heat only to what it can see. From 550 kilometres up, a panel facing down sees the Earth across 85 percent of its view: a warm planet glowing its own infrared, and reflecting sunlight, straight back into the panel.1 Run that panel hot, at 80 °C, and it nets about 549 watts per square metre. Run it near room temperature, at 20 °C, and it keeps about a quarter of its emission: 142 watts — nearly useless.15 That is why orbital compute must run hot. It is a requirement, not an optimisation.
A flat plate also radiates from its upper face into deep space, so a two-sided panel at 80 °C sheds about 1,325 watts per square metre of plate. Turning the panel edge-on to the Earth, so that each face sees it across only about a quarter of its view, helps less than it sounds: about 1,414 to 1,445 watts per square metre of plate, 5 to 7 percent more, because the upper face was already looking at deep space.15 Even so, the areas climb with the load. A 100-kilowatt AI rack needs about 75 square metres of panel (75–110, depending on attitude); a megawatt needs 750 to 1,100. With panels, loops, pumps, structure and micrometeoroid shielding, that comes to about 9 tonnes of radiator per megawatt of heat (5–19, depending on panel design and orientation).15 Power was free. Getting rid of it is the bill.
One honest limit: the panel masses behind those tonnes are engineering estimates we have not yet checked against a primary source. The heat fluxes above do not depend on them.
In June we wrote that the thin Martian atmosphere was enough to convect — to blow heat off a radiator the ordinary way — and we called the surface “cooling-rich”. The coefficient we used was 3.5 to 9 watts per square metre per kelvin. When we restated the paper as a technical memorandum, we first wrote down, dated and sealed, the result that would prove it wrong. Then we recomputed it from textbook flat-plate physics with measured properties of carbon dioxide at Martian pressure.1415 That result came in.
| Surface and air | Coefficient, W/m²K | Reading |
|---|---|---|
| 1 m radiator, still air | 0.26 | Radiator scale600 Pa CO₂, 210 K, panel 60 K warmer |
| 1 m radiator, 4 m/s wind | 0.52 | Radiator scale |
| 1 m radiator, 25 m/s dust-storm wind | 1.24 | Radiator scalethe windiest case still stays near one |
| 2 cm element, 4 m/s wind | 3.6 | Sensor scale |
| 4 mm element, 4 m/s wind | 8.7 | Sensor scalewhere our June figure lives |
Our June figure was real physics on the wrong object. Values of 3.5 to 9 belong to elements a few millimetres to a few centimetres long, such as the rods of a wind sensor.12 For a radiator a metre across, the air carries ten to twenty times less. Three independent results agree at that scale: a computational study of a 0.65-metre rover plate in 15 m/s wind gives 1.4 to 1.8,10 a full-scale model of a radioisotope generator in still Martian air gives 0.16 to 0.24,11 and Curiosity’s measured surface energy budget limits the heat the ground loses to the air to 18 to 38 watts per square metre.9 Taken at a field value of 0.5, convection adds about 72 watts per square metre to a panel at 80 °C (29–143): about 9 percent of what it sheds, not the half we had assumed.15
Forcing the air through a finned heat exchanger with fans does work: with millimetre passages, about 930 watts per square metre of footprint, for fan power of about 1 to 4 percent of the heat moved.13 But that is a machine you power, not a gift from the planet. It has not been demonstrated at scale under Martian conditions, measured pressure losses at low density have run higher than the correlations predict, and what dust does to the passages is not quantified.
Now set the same panel face-up on the Martian surface. It sees only the sky — cold, thin, with no planet in it. On a clear night that sky radiates like a body at 150 to 165 kelvin; in a global dust storm, at 200 to 225 kelvin, inferred from how much the surface warmed at night during the 2018 storm.8 A panel at 80 °C nets 739 to 751 watts per square metre on a clear night and 648 to 696 in a global dust storm — 1.36 and 1.18 times the 549 of the orbital panel facing Earth.15 By day, with the Sun screened, it still nets about 724; left in the Sun, about 614.4
The difference is geometry. On Mars the warm ground, the Sun and the sunlight reflected off the ground all sit in the lower half of the panel’s view, where a screen can block them. In orbit the Earth fills the half the panel radiates into, and nothing can block it. The ground itself is no help: dry regolith conducts so poorly that a kilowatt would need about 51 square metres of contact, and the soil around it warms up within hours.7 It is a slow buffer, not a heat sink.
| Sunlit orbit | Mars surface | |
|---|---|---|
| Power | Rich1,361 W/m², continuous in the sunlit arc | Poor≈ 590 W/m² clear-sky peak; real night; dust storms |
| Radiating face | 549 W/m²Earth in view, cannot be screened | 739–751 W/m² clear night648–696 in a dust storm; ground and Sun can be screened |
| Per square metre of plate | 1,325 W/m²two faces; 1,414–1,445 held edge-on | ≈ 796 W/m²one face plus the air |
| Passive air cooling | None — vacuum | About 9 % of rejection |
| Powered air cooling | None | ≈ 900 W/m² of footprint with fansnot demonstrated at scale |
| Heat into the ground | None | Negligible |
| Binding limit | Radiator area, temperature, mass | Power through night and dust |
One more number keeps the case at its correct size. An orbital panel radiates from both faces; a Martian panel lying face-up radiates mostly from one. Per square metre of plate, orbit sheds 1.7 to 2 times more heat, depending on how the orbital panel is held.15 The surface does not make radiators lighter per watt, and we no longer call it cooling-rich. Its sky is cleaner, not richer: it can be screened, it does not depend on pointing away from a planet, and it leaves the option of powered air cooling that orbit does not have.
The surface’s real weakness is power. Sunlight there peaks at about 590 watts per square metre on a clear day,4 the night is real, and the 2018 global dust storm cut solar array output by 96.7 percent.5 A power source that does not care about the sky closes that gap: fission today, which NASA selected for crewed Mars surface power in December 2024;6 fusion only if it becomes the lighter option at large scale, which we examine, conditionally, in Coming Back. We do not fight solar in orbit; that is solar’s turf, and we say so plainly. We work the place where the sky is clean and only the power is missing. That is the case at its correct size, and it is still the reason FUNNY points at the dirt.
Everything here is steady state: no eclipse or attitude transients, no coating wear, no micrometeoroid design. The coating properties and panel masses are estimates. The Martian sky temperatures are inferred from reported surface warming, not measured directly. No one has measured convection on a metre-scale plate on Mars; our coefficients come from textbook correlations with an allowance for turbulence. The fan-cooled exchanger is a modelled design. Nothing here says when either architecture is built, or that any power source works. The memorandum has had no external review.
Download the technical memorandum — FUNNY-TM-2026-003 Rev C, Heat Rejection for Dense Compute in Sunlit Low Earth Orbit and on the Martian Surface (PDF, 10 pages) → · the calculation script (Python) (with its convection module)
Every number sourced, calculated or marked as an estimate; every table produced by the published script; every open source archived. The source numbers on this page are the memorandum’s own reference numbers. It corrects five points of the June white paper, the largest being Martian convection; its revision record lists all five. The original June 2026 white paper remains available unchanged, for the record.
What changed, October 2026. This page was rewritten in the format of our other articles, with numbered and graded sources. Its numbers now follow the memorandum’s central inputs, NASA’s reference values for Earth: the orbital panel nets 549 and 142 W/m² at 80 and 20 °C, where the earlier text gave 545 and 137 from the white paper’s inputs. Revision C of the memorandum adds a panel held edge-on to the Earth. The finding and its September correction are unchanged.
A standing premise of this page: no magnetic-confinement device has yet demonstrated net fusion gain; nothing here depends on one, and no claim is made about any power source. Stated June 2026, revised September 2026, rewritten October 2026.
Every number above traces to a source or to our own arithmetic, and the memorandum shows the arithmetic. The numbers here match the memorandum’s reference list. Click a number in the text to land here.
Full derivation, assumptions and sensitivity: FUNNY-TM-2026-003 (download above). Not externally reviewed.