For basically any “space datacenter” scenario, imagine putting that same thing in a vast desert instead. You’ll find it’s easier and an order of magnitude cheaper.
If it is a double-sided radiator panel, the physical panel area could be about half:
2,530 m² of panel, about 50 m × 50 m, assuming both sides radiate effectively.
Also temperature matters enormously so
At emissivity 0.85:
Radiator temp
Area for 100 MW
300 K
~256,000 m²
500 K
~33,200 m²
800 K
~5,100 m²
So the answer is about 5,000 m² (lol this is like “a football field” on each side) at 800 K, but balloons to absurd levels like hundreds of thousands of m² if you are trying to dump room-temperature waste heat which there would be a significant amount of. That is for a single small data center at current power needs. In the US alone data centers use 176 TWh (https://www.congress.gov/crs-product/R48646), so there is no chance we are going to be migrating a significant portion of it into space.
80C (still very hot for datacenter hardware coolant) is 350K. And there are other challenges, like effects from being in LEO, or proximity to wherever the solar array is.
There’s tons of spam about “solving” this after the Tech Bro boom, but I don’t really buy anything I’ve seen. Nothing but a bunch of lead (or the Earth’s atmosphere) is going to stop fat gamma rays or extremely fast nuclei.
Yeah, the temperature was an estimate for the nuclear reactor that would be needed lol, I tried to explain that most of the datacenter would be closer to room temperature which would require absurd sizes of radiators
AI doesn’t actually need that much. I’m pretty sure entire models like GLM 5.2 or Deepseek v4 are trained (and served) on a much, much smaller scale than a 100MW cluster.
But if that’s the case… why even invest in orbital data centers in the first place?
Why not desert ones? Why all this cash there instead of actually improving LLM architectures!?
There are so many nested levels of absurdity here. It’s all just total mania, with zero punishment to those doing the funding because they are too rich to feel any consequences.
It’s possible, but not economical.
For basically any “space datacenter” scenario, imagine putting that same thing in a vast desert instead. You’ll find it’s easier and an order of magnitude cheaper.
Yeah, maybe not impossible, but I mean extremely unlikely. I found a thread on reddit that had examples and a spreadsheet https://www.reddit.com/r/IsaacArthur/comments/11kp7s4/how_large_of_a_heatradiator_would_a_spacecraft/
To run a data center in space you would need some kind of reactor producing around 100 MW. If rejecting 100 MW at 800 K
A= 100,000,000 / 0.85×5.670374419e−8×800
The number is the Stefan-Boltzmann constant (σ) https://physics.knox.edu/OnlineHW/zTest-PhysicalConstants.html
A≈5,065 m²
So roughly:
5,100 m² of radiating surface
That is a square about:
√(5065) ≈71 m per side
If it is a double-sided radiator panel, the physical panel area could be about half:
2,530 m² of panel, about 50 m × 50 m, assuming both sides radiate effectively.
Also temperature matters enormously so
At emissivity 0.85:
So the answer is about 5,000 m² (lol this is like “a football field” on each side) at 800 K, but balloons to absurd levels like hundreds of thousands of m² if you are trying to dump room-temperature waste heat which there would be a significant amount of. That is for a single small data center at current power needs. In the US alone data centers use 176 TWh (https://www.congress.gov/crs-product/R48646), so there is no chance we are going to be migrating a significant portion of it into space.
800K is 526C! You can’t run a datacenter at that.
80C (still very hot for datacenter hardware coolant) is 350K. And there are other challenges, like effects from being in LEO, or proximity to wherever the solar array is.
And this is just one of MANY ridiclous engineering challenges. Another great example is that GPUs, memory, and SSDs get random bit flips in orbit, and the issue gets worse with smaller lithography: https://www.itpro.com/server-storage/high-performance-computing-hpc/367323/hpes-supercomputer-helps-iss-astronauts
There’s tons of spam about “solving” this after the Tech Bro boom, but I don’t really buy anything I’ve seen. Nothing but a bunch of lead (or the Earth’s atmosphere) is going to stop fat gamma rays or extremely fast nuclei.
Yeah, the temperature was an estimate for the nuclear reactor that would be needed lol, I tried to explain that most of the datacenter would be closer to room temperature which would require absurd sizes of radiators
To be fair, 100MW is pretty big.
AI doesn’t actually need that much. I’m pretty sure entire models like GLM 5.2 or Deepseek v4 are trained (and served) on a much, much smaller scale than a 100MW cluster.
But if that’s the case… why even invest in orbital data centers in the first place?
Why not desert ones? Why all this cash there instead of actually improving LLM architectures!?
There are so many nested levels of absurdity here. It’s all just total mania, with zero punishment to those doing the funding because they are too rich to feel any consequences.