This is the chapter the thesis lives or dies on, and it is the one I would most like a hostile engineer to attack.
Restating the problem exactly
From the primer in Chapter 1: essentially all electrical power entering a computer leaves it as heat. A 100 kW node is a 100 kW heater. A 1 MW node is a 1 MW heater, which is roughly the thermal output of two hundred domestic ovens running flat out, inside a structure the size of a bus, in a vacuum. There are three ways to move heat: conduction, convection, and radiation. In orbit the second one is gone. There is no air to blow across a fin, no water to evaporate, no ground to sink into. Conduction still works inside the spacecraft, that is how you get heat from the chip to the panel but at the boundary with the universe there is exactly one mechanism, and it is radiation.
The equation, and the one thing it tells you
A surface at temperature T radiates according to the Stefan, Boltzmann law. Net of what it absorbs from its surroundings: q = ε σ (T⁴ − T_sink⁴) where q is watts per square metre, ε is emissivity (how close the surface is to a perfect radiator, good radiator coatings reach 0.85 or better), and σ = 5.67 × 10⁻⁸ W/m²K⁴. Everything about orbital thermal design is contained in that exponent. Heat rejection scales as the fourth power of temperature. Run your radiator ten percent hotter in absolute terms and you reject about 46% more heat from the same panel. That single fact should reorganise how you evaluate this sector.
The number
Take ε = 0.85 and run the equation across plausible radiator temperatures and two sink assumptions, for a twosided panel radiating from both faces:
Radiator surface
Net flux, 250 K sink
Area per MW
Net flux, 200 K sink
Area per MW
40 °C (313 K)
274 W/m²
1,820 m²
386 W/m²
1,300 m²
50 °C (323 K)
336 W/m²
1,490 m²
448 W/m²
1,120 m²
70 °C (343 K)
479 W/m²
1,040 m²
590 W/m²
850 m²
100 °C (373 K)
745 W/m²
670 m²
856 W/m²
580 m²
temp
Figure 5.1 — Why the whole game is running hot
The figure most often quoted in this sector is roughly 1,200 square metres per megawatt. You can now see exactly what that figure assumes: a two-sided panel at around 40, 50 °C with a reasonably cold sink. It is a fair number, not a promotional one but it is not a law of nature either. It is the answer at one point on a curve, and a company that can run its radiators at 70 °C instead of 45 °C needs half the panel. This is the single most useful lens in the sector, so it is worth stating as a rule: Every degree you can run hotter is area you do not have to launch. The thermal problem is therefore not really a thermal problem, it is a question of how hot you are willing to let the silicon get.
Terrestrial data centres run chips cool because cooling is cheap and hot silicon fails sooner. In orbit, cooling is the dominant mass in the vehicle. The economically correct answer is almost certainly to run hotter than any terrestrial operator would tolerate, accept a shorter hardware life, and replace more often. Nobody in this sector talks in these terms yet. I expect the winners will.
The half of the problem the equation does not cover
Radiator area is the easy half. The hard half is getting a megawatt of heat from a few hundred square centimetres of silicon out to a thousand square metres of panel. The chain is a series of temperature drops, and every drop costs you area. Junction to cold plate. Cold plate to coolant. Coolant pumped through a loop. Loop into the radiator root. Root out along the fin, which is never isothermal. If the junction runs at 90 °C and you lose 30 degrees down the chain, you are radiating at 60, not 90, and Figure 5.1 charges you for it. The available technologies, in ascending order of capability and descending order of flight heritage: Heat pipes and loop heat pipes. Passive, extremely reliable, decades of orbital heritage and limited to
hundreds of watts to low kilowatts per line. They do not scale to megawatts. Pumped single-phase loops. Mechanical pumps circulating a fluid. Flight-proven at the tens-of-kilowatts scale on the ISS. Needs pumping power and introduces moving parts. Pumped two-phase loops. The coolant boils, which carries far more heat per kilogram of flow and holds temperature nearly constant. This is the right answer at megawatt scale, and it is the least flight-proven. Managing two-phase flow in microgravity, where vapour does not obligingly rise, is a genuinely hard problem. This is the honest statement of where the sector is: between roughly ten watts and five hundred kilowatts, thermal management is solved with hardware that has flown for decades. Above a megawatt, it is not solved, and the companies claiming gigawatts have not built the intermediate step.
Mass, and the sentence that follows from it
Deployable spacecraft radiators run roughly 5, 12 kg per square metre including structure, plumbing and fluid. At 1,200 m² and the middle of that range, a megawatt of rejection is on the order of 9 to 18 tonnes. Now assemble the whole vehicle.
Figure 5.2 — You are not launching computers. You are launching their cooling
.
A megawatt-class orbital compute node lands somewhere around 27 to 45 tonnes, of which the computers, the only part that earns revenue are roughly six to eight. The revenue-generating payload is about a fifth of what you launch. Everything else is the apparatus required to feed it and cool it. That is not a refutation. Terrestrial data centres are also mostly not-computers: substations, chillers, generators, concrete. It is, however, the number that should be in your head whenever someone multiplies a launch price by a rack weight and declares the economics obvious. At $600 per kilogram, a megawatt node costs $16, 27 million to launch, against perhaps $15, 20 million of computers inside it. Launch is not a rounding error at megawatt scale. It is comparable to the payload.
The complication a cluster introduces
One more effect belongs here, because Chapter 13’s preferred architecture quietly undermines Chapter 5’s arithmetic. Everything above assumes a radiator with a clean view of cold sky. A formation-flying cluster does not have one. If your neighbours sit a few hundred metres away, they occupy part of your radiator’s field of view, and they are not cold sky, they are spacecraft at roughly three hundred kelvin. The effective sink is then a weighted mix: the cold background over most of the hemisphere, and warm hardware over the rest.
Figure 5.3 — A cluster radiates partly into itself
The penalty is modest at realistic spacings and it is not zero. With a tenth of the radiator’s sky filled by neighbours, a panel running at 58 °C needs roughly five to six percent more area; at a fifth of the sky, more than ten percent. Cooler radiators suffer proportionally more, because they have less temperature margin over the sink to begin with, the same fourth-power logic, working against you. Two consequences worth carrying forward. Formation spacing is a thermal decision, not only a control decision, and a cluster design that optimises station-keeping propellant without a view-factor analysis has optimised the wrong variable. And every radiator area figure in this book, including the ones in Chapter 8’s buildup, is a single-vehicle number that should be inflated by several percent for any clustered architecture.
Where heat breaks
Two-phase transport at megawatt scale, the unretired technical risk of the entire sector. Deployment. A thousand square metres of radiator has to fold into a fairing and unfold without a single jam. This is the same mechanism risk as the solar array, doubled. Micrometeoroid and debris puncture. A radiator is a large, thin, fluid-filled target with the largest surface area on the vehicle. Puncture is not hypothetical over five years; it is an actuarial certainty at some rate, which is why panel isolation and loop redundancy are design requirements and not refinements. This is also the chapter that quietly determines what Chapter 16’s insurers will charge. Degradation of optical properties. Emissivity is a coating property, and coatings erode under atomic oxygen and UV. A radiator that loses emissivity loses capacity on exactly the fourth-power curve above.