The second is thermal, and it is the harder of the two, because you can price a five-year replacement cycle, but there is no equivalent workaround for heat. It has no validated answer at scale, and it is the one that decides the shape of the satellite, the choice of orbit, and how much of your payload does anything useful at all.
Space is a flask (not a fridge)
In space, nobody can hear you scream. For precisely the same reason, nobody can take your heat away either. Sound needs a medium to travel through, and so does convection.
This is the part that trips people up, because space is famously cold - a few degrees above absolute zero, colder than anywhere your engineers will ever design for. But cold is not the same as cooling…
…you are not putting your Data Centre somewhere cold, you are putting it somewhere empty. There is no air to blow over anything, no water to carry anything away, nothing to hand the problem to.
Which is the whole trick of the flask. Nobody engineers a Thermos to be warm. They engineer it to be empty, and let the emptiness do the work.
So, how do you get heat out of a flask?
There is exactly one exit left. Heat can leave as infrared radiation, and radiated power per square metre follows the Stefan–Boltzmann law:
Two things in that equation are yours to set. The first is ε, emissivity - how willingly a surface radiates. And here your flask is actively working against you, because the silvered lining inside it exists specifically to reflect infrared back in and stop radiation doing what convection can't.
So the job is to build the anti-Thermos: a surface engineered to emit as hard as physics allows.
Good news - we're rather good at that. Modern radiator coatings reach an emissivity of about 0.85, and the theoretical ceiling is 1.0. You can't beat perfect, and we're already most of the way there.
Which leaves temperature. And the temperature is raised to the fourth power.
Double a radiator's temperature and it sheds sixteen times as much heat from every square metre. So for a fixed quantity of waste heat, the hotter you can afford to run the radiator, the less radiator you have to build.
Here is that trade, with everything else held constant:
|
|
|
Radiator area for the same job
|
300 K (27 °C) - where AI silicon actually runs
|
|
|
|
|
|
|
|
|
|
|
800 K (527 °C) - where no chip survives
|
|
|
Idealised comparison; real radiators also contend with Earth's infrared and albedo loads.
Run it hotter, in other words, and the area collapses, and with it the mass, and with it the launch bill. Now read the bottom row again - the physics is offering a 98% discount to anyone who can run their processors at 527 °C…
…The thing is, nobody can physically do that.
And that is the catch that everything else in this article is a consequence of: temperature is the master variable, and you don't get to pick it.
An AI accelerator's waste heat is low-grade…
…It leaves at roughly the temperature the silicon runs at, somewhere around 300-350 K, and there is no dial. To lift it somewhere more useful you would need either semiconductors that tolerate far higher junction temperatures than any current process node does, or a heat pump, which draws power, which makes more heat, which you then also have to reject. In simpler terms, that is a down escalator you are paying to ride upwards.
The fourth power, then, is generous to the hot and merciless to the tepid. AI silicon is tepid. It is stuck in the least forgiving corner of the curve, and no amount of clever architecture gets it out.
That is a strong claim, so it deserves testing.
The industry has made two serious attempts to design its way out of exactly this…
…one very large, one very small, and both are worth exploring…
Serious design attempt #1 - one very large machine
Start with the cleanest version of the problem. A single 100 MW platform, parked in a geostationary-class orbit where Earth's infrared and reflected sunlight drop to single-digit watts per square metre and your radiator sees something close to a genuinely empty sky.
Even with the environment doing its best for you, rejecting 100 MW at a realistic 300-350 K needs somewhere between 0.13 and 0.24 km² of single-sided radiator. That is a square 350 to 490 metres on a side, weighing anywhere from several hundred to well over a thousand tonnes.
Nothing of that size fits inside a payload fairing, so you can't launch it.
You would have to assemble it up there, which promotes in-space construction from an interesting future capability to the load-bearing assumption of your entire business case.
You would also have created the largest object in orbit with absolutely no ability to duck.
Attempt one isn't wrong, it's just the wrong shape. Note, though, that the physics hasn't moved an inch…
Serious design attempt #2 - a great many small ones
(…Which is why the industry has converged on swarms)
Hundreds of modest satellites flying in tight formation, stitched together with free-space optical links. Google's illustrative design is 81 satellites inside a 1 km radius, in a dawn-dusk sun-synchronous orbit at about 650 km chosen for near-permanent sunlight. Each unit fits inside an existing fairing and needs no kilometre-scale truss, no deployment mechanism of unprecedented size, and no in-space assembly at all. It is a genuinely elegant answer to the structural problem, which is exactly what it was designed to solve.
It is not an answer to the thermal one though, because the thermal one comes with you.
Cost out a single satellite. Take one built from in-production, space-qualified parts, delivering around 15-16 kW of usable compute. Add communications, bus and power-management losses, and you have roughly 23 kW of heat to get rid of. At a practical net rejection of about 300 W/m², that calls for some 75 m² of radiator, massing 260-370 kg.
Which brings us back to the number this article opened with.
If between 45% and 65% of your spacecraft is now radiator, you haven’t launched a computer with a cooling system, you’ve launched a cooling system with a computer on board.
Low Earth orbit then adds a few refinements that geostationary is spared. At 650 km you cross from sunlight into eclipse and back roughly sixteen times a day, thermally cycling every heat pipe, panel and joint dozens of times more often than a GEO platform would ever see.
Sixteen sunrises a day sounds romantic right up until you are a solder joint. Earth's infrared and albedo loads are stronger and more variable down there. And satellites flying this close spend part of their time radiating at one another rather than into clean, empty sky…
…aka: eighty-one flasks in a picnic basket, each keeping the others warm.
None of it is disqualifying, but all of it makes the same penalty worse.
How much would this cost, and how much of it has to go right?
At today's launch pricing, orbital power costs roughly $13,500-17,200 per kW per year. US terrestrial Data Centres pay $570-3,000. That is an order of magnitude, and it is the number the sceptics reach for.
Two clarifications before anyone screenshots it. It is annualised - a satellite launches once and works for about five years, so its one-off launch cost is spread across that life to sit fairly beside a recurring bill. And it compares power cost, not total cost of ownership: the orbital figure leaves out the bus, insurance, ground operations and deorbit, just as the terrestrial figure leaves out servers, buildings and staff.
The question being tested is whether launched power can get into the same postcode as grid power, (not whether it has already moved in).
The bridge across that gap is a learning curve. Launch pricing has historically fallen about 20% for every doubling of cumulative mass launched. Extend the line and LEO reaches roughly $200/kg by the mid-2030s, at which point orbital power lands near $810/kW/year - comfortably inside the terrestrial band.
That line is being asked to do a great deal of learning.
- It assumes something like 180 Starship-class launches a year, sustained.
- It assumes total launched mass grows roughly a thousandfold.
- It assumes the 20%-per-doubling curve holds all the way down instead of flattening out as the industry matures, which is what learning curves usually do.
- And it low-key assumes an efficient satellite, while we have just spent two sections establishing that up to two-thirds of it may be radiator.
That last point closes the loop. Adding the radiator takes the satellite from roughly 575 kg to 945 kg - a two-thirds increase in launch mass before a single extra watt of compute is delivered. At $810/kW/year, orbital power is still inside the terrestrial band, but most of the margin has gone, and it is the thermal design that spent it.
Heat rejection, therefore, is one of the conditions that has to hold for it to work at all.
The same flask, 650 kilometres lower
None of this makes orbital compute science fiction. Google plans to launch two prototype satellites with Planet by early 2027, and the first data-centre-class GPU is already operating in orbit aboard Starcloud-1. The concept is being tested, properly, by serious people.
But the reason it is worth an engineer's attention has very little to do with space.
Every advance in Data Centre cooling over the past decade has been a way of prising that lid open a little further.
- Hot-aisle containment
- Free cooling
- Direct-to-chip liquid
- Immersion
…each one a better answer to the same inconvenient fact, which is that AI silicon generates enormous quantities of heat at a temperature too low to be worth much to anyone.
On Earth we have never had to confront that squarely, because there has always been something to hand the problem to. Air, then water, then somebody else's water.
That is the only real difference between the two cases.
A terrestrial Data Centre is a flask we have been holding open for thirty years. Orbit screws the lid back down, takes away the air and the water, and leaves the physics standing there with nothing on, which makes it an unusually clarifying place to think from, even if nobody ever builds the thing.
So the interesting question goes from "can we put a Data Centre in space?" to: "at what cost, and against which terrestrial baseline does it actually represent an improvement?"
That conversation is well worth having, provided it stays grounded in physics, honest about its assumptions, and clear about which constraint is genuinely still open.
Low-grade heat rejection is that constraint. It is unsolved in orbit. It is also unsolved on the campus you are permitting this quarter, and down here, at least, you can still take the lid off.
You’re probably ready for a coffee now.
