Radiators in space are extremely inefficient because there's no conduction.
Also you have huge heat inputs from the sun. So you need substantial cooling before you get around to actually cooling the GPUs.
EDIT: people continue downvoting and replying with irrelevant retorts, so I'll add in some calculations
Let's assume
1. cheap 18% efficient solar panels (though much better can be achieved with multijunction and quantum-cutting phosphors)
2. simplistic 1360 W/m^2 sunlight orthogonal to the sun
3. an abstract input Area Ain of solar panels (pretend its a square area: Ain = L ^ 2)
4. The amount of heat generated on the solar panels (100%-18%) * Ain * 1360 W / m ^ 2, the electrical energy being 18% * Ain * 1360 W / m ^ 2. The electrical energy will ultimately be converted to computational results and heat by the satellite compute. So the radiative cooling (only option in space) must dissipate 100% of the incoming solar energy: the 1360 W / m^2 * Ain.
5. Lets make a pyramid with the square solar panel as a base, with the apex pointing away from the sun, we make sure the surface has high emissivity (roughly 1) in thermal infrared. Observe that such a pyramid has all sides in the shade of the sun. But it is low earth orbit so lets assume warm earth is occupying one hemisphere and we have to put thermal IR reflectors on the 2 pyramid sides facing earth, so the other 2 pyramid sides face actual cold space.
6. The area for a square based symmetric pyramid: we have
6.a. The area of the base Ain = L * L.
6.b. The area of the 4 sides 2 * L * sqrt( L ^ 2 / 4 + h ^ 2 )
6.c. The area of just 2 sides having output Area Aout = L * sqrt( L ^ 2 / 4 + h ^ 2 )
7. The 2 radiative sides not seeing the sun and not seeing the earth together have the area in 6.c and must dissipate L ^ 2 * 1360 W / m ^ 2 .
8. Hello Stefan-Boltzmann Law: for emissivity 1 we have the radiant exitance M = sigma * T ^ 4 (units W / m ^ 2 )
9. The total power exited through the 2 thermal radiating sides of the pyramid is then Aout * M
10. Select a desired temperature and solve for h / L (to stay dimensionless and get the ratio of the pyramid height to its base side length), lets run the satellite at 300 K = ~26 deg C just as an example.
11. If you solve this for h / L we get: h / L = sqrt( ( 1360 W / m ^ 2 / (sigma * T ^ 4 ) ) ^ 2 - 1/4 )
12. Numerically for 300K target temperature we get: h/L = sqrt((1360 / (5.67 * 10^-8 * 300 ^ 4)) ^ 2 - 1/4) = 2.91870351609271066729
13. So the pyramid height of "horribly poor cooling capability in space" would be a shocking 3 times the side length of the square solar panel array.
As a child I was obsessed with computer technology, and this will resonate with many of you: computer science is the poor man's science, as soon as a computer becomes available in the household, some children autodidactically educate themselves in programming etc. This is HN, a lot of programmers who followed the poor man's science path out of necessity. I had the opportunity to choose something else, I chose physics. No amount of programming and acquiring titles of software "engineer" will be a good substitute for physicists and engineers that actually had courses on the physical sciences, and the mathematics to follow the important historical deductions... It's very hard to explain this to the people who followed the path I had almost taken. And they downvote me because they didn't have the opportunity, courage or stamina to take the path I took, and so they blindly copy paste each others doomscrolled arguments.
Look I'm not an elon fanboy... but when I read people arguing that cooling considerations excludes this future, while I know you can set the temperature arbitrarily low but not below background temperature of the universe 4 K, then I simply explain that obviously the area can be made arbitrarily large, so the temperature can be chosen by the system designer. But hey the HN crowd prefers the layers of libraries and abstractions and made themselves an emulation of an emulation of an emulation of a pre-agreed reality as documented in datasheets and manuals, and is ultimately so removed from reality based communities like physics and physics engineering, that the "democracy" programmers opinions dominate...
So go ahead and give me some more downvotes ;)
If you like mnemonics for important constants: here's one for the Stefan Boltzman constant: 5.67 * 10^-8 W / m^2 / K ^ 4
thats 4 consecutive digits 5,6,7,8 ; comma or point after the first significant digit and the exponent 8 has a minus sign.
Space is not empty. Satellites have to be boosted all the time because of drag. Massive panels would only worsen that. Once you boosters are empty the satellite is toast.
Building this is definitely not trivial and not easy to make arbitrarily large.
I am highly skeptical about data centers in space, but radiators don't need to be unshaded. In fact, they benefit from the shade. This is also being done on the ISS.
The JWST operates at 2kw max. That's not enough for a single H200.
AI datacenters in space are a non-starter. Anyone arguing otherwise doesn't understand basic thermodynamics.
On Low Earth Orbits (LEOs), sure, but the traces of atmosphere that cause the drag disappear quite fast with increasing altitude. At 1000 km, you will stay up for decades.
On Earth, you can vent the heat into the atmosphere no problem, but in space, there's no atmosphere to vent to, so dissipating heat becomes a very, very difficult problem to solve. You can use radiators to an extent, but again, because no atmosphere, they're orders of magnitude less effective in space. So any kind of cooling array would have to be huge, and you'd also have to find some way to shade them, because you still have to deal with heat and other kinds of radiation coming from the Sun.
It's easier to just keep them on Earth.
The whole concept is still insane though, fwiw.
people heavily underestimate radiative cooling, probably because precisely our atmosphere hinders its effective utilization!
lesson: its not because radiative cooling is hard to exploit on earth at sea level, that its similarily ineffective in space!
This is precisely why my didactic example above uses a convex shape, a pyramid. This guarantees each surface absorbs or radiates energy without having to take into account self-obscuring by satellite shape.
A very high end desktop pulls more electricity than the whole JWST... Which is about the same as a hair dryer.
Now you need about 50x more for a rack and hundreds/thousands racks for a meaningful cluster. Shaded or not it's a shit load of radiators
https://azure.microsoft.com/en-us/blog/microsoft-azure-deliv...
for a reasonable temperature (check my comment for updated calculations) the height of a square based pyramidal satellite would be about 3 times the side length of its base, quite reasonable indeed. Thats with the square base of the pyramid as solar panel facing the sun, and the top of the pyramid facing away, so all sides are in the shade of the base. I even halved my theoretical cooling power to keep calculations simple: to avoid a long confusing calculation of the heat emitted by earth, I handicapped my design so 2 of the pyramidal side surfaces are reflective (facing earth) and the remaining 2 side triangles of the pyramid are the only used thermal radiative cooling surfaces. Less pessimistic approaches are possible, but would make the calculation less didactic for the HN crowd.
for a 4 m x 4 m solar panel, the height of the pyramid would have to be 12 m to attain ~ 300 K on the radiator panels. Thats also the cold side for your compute.
for a 4 km x 4 km solar panel the height of the pyramid would be 12 km.
The larger you make the area, the more solar energy you are collecting. More shade = more heat to radiate. You are not actually making the problem easier.
for a target temperature of 300K that would mean the pyramid height would be a bit less than 3 times higher than the square base side length h=3L.
I even handicapped my example by only counting heat radiation from 2 of the 4 panels, assuming the 2 others are simply reflective (to make the calculation of a nearby warm Earth irrelevant).
These people are all smoking crack.
Here's a big one: you can't put radiators in shadow because the coolant would freeze. ISS has system dedicated to making sure the radiators get just enough sunlight at any given time.
Here's a big one: you can't put radiators in shadow because the coolant would freeze. ISS has system dedicated to making sure the radiators get just enough sunlight at any given time.
[0] https://developer.nvidia.com/deep-learning-performance-train...
Click the "Large Language Model" tab next to the default "MLPerf Training" tab.
That takes 16.8 days on 128 B200 GPU's:
> Llama3 405B 16.8 days on 128x B200
A DGX B200 contains 8xB200 GPU's. So it takes 16.8 days on 16 DGX B200's.
A single DGX (8x)B200 node draws about 14.3 kW under full load.
> System Power Usage ~14.3 kW max
source [1] https://www.nvidia.com/en-gb/data-center/dgx-b200
16 x 14.3 kW = ~230 kW
at ~20% solar panel efficiency, we need 1.15 MW of optical power incident on the solar panels.
The required solar panel area becomes 1.15 * 10^6 W / 1.360 * 10^3 W / m ^ 2 = 846 m ^ 2.
thats about 30 m x 30 m.
From the center of the square solar panel array to the tip of the pyramid it would be 3x30m = 90 m.
An unprecedented feat? yes. But no physics is being violated here. The parts could be launched serially and then assembled in space. Thats a device that can pretrain from scratch LLaMa 3.1 in 16.8 days. It would have way to much memory for LLaMa 3.1: 16 x 8 x 192 GB = ~ 25 TB of GPU RAM. So this thing could pretrain much larger models, but would also train them slower than a LLaMa 3.1.
Once up there it enjoys free energy for as long as it survives, no competing on the electrical grid with normal industry, or domestic energy users, no slow cooking of the rivers and air around you, ...
this system would not be given such an orbit. Its trivial to decrease the cooling capacity of the radiators: just have an emissivity ~0 shade (say an aluminum foil) curtain obscure part of the radiator so that it locally sees itself instead of cold empty space. This would only happen during 2 short periods in the year.
The design issues of the ISS are totally different from this system.
Nobody said sending a single rack and cooling it is technically impossible. We're saying sending datacenters worth of rack is insanely complex and most likely not financially viable nor currently possible.
Microsoft just built a datacenter with 4600 racks of GB300, that's 4600 * 1.5t, that alone weights more than everything we sent into orbit in 2025, and that's without power nor cooling. And we're still far from a single terawatt.
a different question is the expected payback time, unless someone can demonstrate a reasonable calculation that shows a sufficiently short payback period, if no one here can we still can't exclude big tech seeing something we don't have access to (the launch costs charged to third parties may be different than the launch costs charged for themselves for example).
suppose the payback time is in fact sufficiently short or commercial life sufficiently long to make sense, then the scale didn't really matter, it just means sending up the system described above repeatedly.
Either it does or it doesn't make financial sense, and if it does the scale isn't the issue (well until we run into material shortages building Elon's Dyson sphere, hah).