Data Centres In Space
Created/Modified: 2025-12-14/2025-12-15
Data centres in space are an implausible idea.
This doesn't mean they won't work. A lot of implausible ideas work, or at least work well enough to let someone claw their way into office or make some money.
Nor are they implausible because they seem like a stupid idea at first glance to a naive observer, because how an idea looks to a naive observer is unrelated to how plausible it is, and every observer is naive when an idea is new.
Our intuition is terrible at judging what will work and what won't. Me, I judged them a dumb idea based on a rough in-my-head estimate of the radiative cooling numbers. What did you judge them a dumb idea on?
My rough in-my-head numbers said data centres in space were a dumb idea, but how dumb? Not as dumb as I thought, but still... pretty dumb.
The numbers are interesting. This is how physicists and engineers--as opposed to techbros with philosophy or political science degrees--think about things. Maybe if techbros took more STEM courses they wouldn't think ideas like data centres in space are good ones. Yeah, that's the problem: techbros take too many humanities courses, not enough STEM courses. That's what I hear, anyway. A lot of people are saying.
The problem is one of energy.
Solar power in, radiative power out.
Solar power comes from the sun. To get it continuously in space you need to either be in solar orbit or polar orbit. Solar orbit is probably a non-starter: you want your data centre to be near to Earth, and the Earth-Sun radial Lagrange points, which are the only place you could put a satellite in solar orbit that stays put relative to the Earth, is 1.5 million kilometres away, which would mean a light-speed round-trip time-delay of 10 seconds. Maybe not a killer problem, but definitely a pain.
Polar orbits are ones that are go around the Earth at a high tilt relative to the equator. If you see a fast-moving "star" just after sunset headed north or south, that's likely a spy satellite in a polar orbit. Even polar orbits require some satellite manoeuvring, because the plane of the orbit will stay constant as the Earth moves around the Sun, so the orbit has to be tweaked to keep your (huge) solar panel array illuminated.
Back in the '70s there was a push to power the Earth with solar panels in space, and "We don't know how to orient large structures in space" was one of the many arguments against it. We know a lot more now, thanks to the decades of experience with the International Space Station, but we definitely don't know how to point (or build) solar arrays of anything like the required size.
And what is the required size?
As I said, before writing this piece I started out with a rough idea of the cooling problem posed by data centres in space, because it looked like the most obvious issue, so I'll run through those numbers first before coming back to solar panels.
The reason why no one can hear you scream in space is because there is no atmosphere to speak of, so cooling is radiative.
Radiative cooling is governed by the Stephan-Boltzmann law, which is:
R = σ*T4where σ, the Stephan-Boltzmann constant, is 5.67-8 W/m2K4: 57 nanowatts per square metre per kelvin-to-the-fourth-power.
Space is "cold" in the sense that if you have a surface that's not exposed to sunlight, it's probably exposed to the microwave background radiation, the left-over light from the Big Bang. This has a temperature of about 4 K, which may as well be zero so far as radiative cooling is concerned. It generates about 14 micro-watts of incoming radiant energy from the universe at large. That fourth power in the temperature does a lot of work for us here.
The sun, on the other hand, produces about 1361 W/m2 at Earth's orbit. This is called the "solar constant", although it's not precisely constant. We'll use that number when we get to talking about power requirements.
Data centres run on chips called "graphics processing units", or GPUs, which are a technology originally developed for video game rendering. A modern GPU like NVIDA's Blackwell runs at just under 100 C (the boiling point of water) and consumes around 1 kW. These are rough numbers, which is all we need, and they are generous numbers: 1 kW is on the low side, 100 C is on the high side. 85 C is a more realistic number.
Because kelvin is temperature relative to absolute zero, and celsius is temperature relative to the freezing point of water, we have to add 273 to our 100 C to get 373 K for the temperature, because water freezes at 273 K.
Plugging in the numbers, this gives us 1098 W/m2 for 100 C and 931 W/m for 85 C, so I'm going to call it 1 kW/m2, which works out nicely: one square metre of cooling is required for every chip, round figures.
What we want is a sense of the scale of these things, so... how many chips are in a data-centre? According to Microsoft, their latest and greatest has "hundreds of thousands of GPUs" and according to OpenAI they are planning a "100,000 GPU data centre".
So let's take them at their word and estimate a hundred thousand GPUs for our data centre in space. And if you can read that without hearing it as IN SPAAAACE you're, well, a different kind of person than I am.
A hundred thousand GPUs requires a hundred thousand square metres of radiator surface--which would be insulated on the side facing the sun and have the radiative side constantly pointed off into the depths of space--to provide a hundred kilowatts of cooling. That's a square over 300 m on a side, which is somewhere in the vicinity of the typical modern shopping mall complex, including the parking lot. This is a rough comparison: the world's largest shopping mall is in Edmonton, and the mall complex itself, independently of its world's-largest-parking lot, is has almost five times more surface area than the required radiative area for a typical modern data centre IN SPAAAACE [sorry, couldn't resist].
That factor of five is an interesting figure, because mass-produced solar panels are 15-20% efficient, and the solar constant--the power available from the Sun--is a bit over 1 kW per square metre, so a space data centre needs about five times the area of solar panels as cooling panels.
Or: the solar panels for a typical data centre in space would have about the same surface area as the West Edmonton Mall, not including the parking lot. Half a million square metres, round figures.
The good news, I guess, is that if we make the back side of the solar panels our radiator surface, we can run our chips at any temperature we want. The basic design of a data center in space is going to be a sandwich: solar panels on one side, radiative surface on the other, chips in the middle.
The chips will probably be in a compact configuration, not spread out in a thin layer, because 30 metres is 100 ns propagation delay at the speed of light, and even more for typical conductors, where signals travel slower than light. Keeping the chips physically close to each other matters a lot in terms of performance, but you need to feed power cables in and have some kind of liquid cooling system pulling heat out and distributing it across your radiative surface, which, again, needs to be pointed away from the sun all the time.
It would be kind of exciting to see what happened if the orientation control failed, and the full power of the sun fell on the dark radiative surface instead of the solar panels, pumping all that energy back into the cooling system. I'm thinking it would probably be a bit like this:
There are actually devices called "thermo-siphons" that only let heat flow one way through a cooling system, but they don't work in zero gravity, and if you spin the data centre to get artificial gravity, it points the wrong way: not in toward the centre but out toward the rim, and you can't put your chips out at the rim because they'd be too far apart from each other.
Building a data centre on Earth costs billions. Building anything in space costs at least hundred times as much, even if we knew how, which we don't. Learning how will cost additional billions. The first prototype, slated to go up in the next year or so, has a mass of 60 kg. The real thing, even at the modest scale I've scoped out here, will be more like six million kilograms, assuming around 10 kg per square metre, once we've built all the infrastructure required to create a structure that covers the better part of a square kilometre and strong enough to be continually reoriented so it faces the right way all the time.
Launch costs are still comfortably north of $1000 per kilogram, but let's say they come down to that, so launch costs alone would be $6 billion for the lousy little system I've outlined here. The AI fantasists are projecting solar panel structures that are 4 kilometres on a side, for a total area of sixteen million square metres.
At 1 kg per square metre that's a launch cost alone of $16 billion. And that's the insanely low bare minimum. It represents about 250 Falcon heavy launches. So far there have been ten.
Component cost is another thing: I used a 15-20% efficiency figure for solar panels above, but satellite solar panels are closer to 30% efficient, but we're not talking a dollar or two per watt the way we are on Earth. Either the solar panels used are going to be orders of magnitude more expensive, or require much larger area, higher launch cost, etc.
On the good side, this might lead to much cheaper high efficiency panels for terrestrial use. Maybe "AI" will solve climate change by making efficient solar panels ultra-cheap!
GPU chips have the same problem as solar panels: cosmic radiation will damage and degrade ordinary commercial chips, which already only have a lifetime of a few years. Many electronic components have commercial and military versions, with MILSPEC chips being required for aerospace applications. They are typically five to ten times more expensive than commercial grade, but for all practical purposes they never fail.
Conclusion: for once, your intuition is correct! Data centres IN SPAAAACE are a dumb idea. They require the building of vast structures we don't know how to put together, don't know how to keep pointed in the right direction, would be hideously expensive to launch, and without enormous leaps in component robustness would be failing out within a decade.
Could they still happen? Sure. People do stupid things all the time. Sometimes they even work.
But as this analysis shows, although cooling is a big issue, it's far from the only, or even the biggest, one.
So my intuition was wrong about that.
Which my intuition or "common sense" or whatever you want to call it usually is.
How about yours?