Comment by AntiUSAbah
6 hours ago
Of course it works, the question is how this would look like and if its financial feasable.
You make a H100, ship it to a space dock, load it onto a rocket (rocket requires fuuel, the rocket, etc.) send it up, deploy it, monitor it live 24/7, have means of adjusting its orbit, if it breaks, its immediade full loss, otherwise it will degenerate faster in space than on earth, now it needs a high speed up/downlink to do anything reasonable which also requires a base station. The base station has to track this satelite.
One H100 costs 40k, consumes 700 Watt peak and need probably at a minimum 5 square meter of area for cooling and solar.
The colossus datacenter from musk has 250.000 of these.
Now you have to track 250.000 single satelites, you have to coordinate the communication between the, up and downlink to earth.
250.000 * 5 square meter of area.
This alone increases the potential debris in space.
And this is ONE 300 MW Datacenter replacement. ONE.
It’s very easy to overestimate the difficulty of cooling things in space, unless you actually run the numbers. So please follow along as Scott Manley runs the numbers: <https://www.youtube.com/watch?v=FlQYU3m1e80>.
Basically a Starlink v3 satellite has an estimated power budget of 20kW. Add in the heat absorbed from the environment (both directly from sunlight and reflected off of the Earth) and you’ll find that it must reject about 22kW of heat. That’s a fair amount, but at 65°C it can radiate it all away just using it’s own surface area! No radiator required at all!
Of course the power density of computer racks has been going up over the years. If you want to reach 100kW per satellite then they will need a modest radiator, but nothing extravagant. It would still be smaller than the solar panels, and far smaller than the ones on the ISS. And don’t forget that because radiated heat goes up as the fourth power of temperature, raising the temperature of the system by even a small amount raises the radiation emitted by a lot. If you design the system to run hotter you can minimize the size of the radiator. Most chips these days are designed to max out at 100°C to 110°C without lasting damage, although running them at that temperature 24/7 may reduce their lifespan. There will be some sweet spot in the middle.
And it turns out that a Starlink v3 already has a volume somewhat larger than a 48U rack. You talk about launching 250k satellites in order to have 250k GPUs in orbit, but that’s ridiculous. A real compute swarm will be hundreds or thousands of satellites each equivalent to a whole rack of GPUs.
But you’re not wrong to be skeptical. The economics might not work out even if the cooling is easy enough. It’s just that rejecting the idea takes a lot more than back–of–the–envelope calculations.
I'm not rejecting the basic idea in itself. There is nothing in this idea which we as humans can't do today. No issues here. Its just so much more expensive than just doing it in a dessert and putting fibre and solar panels and batteries there.
The Starlink v3 doesn't exist yet in space, it also needs Starship apparently and Musk said it will have the size of a Boeing 737 fully deployed. So it will not be small and its not proofen yet.
A rack with 48u will either have 12 or 24 GPUs which equals to 9kW or 17kW. Than its not 250k satellites for a 'small' 300MW DC but only 25k. Still a very crazy number.
I would love to see all of this scifi stuff happening. Spaceship in space, travel gates, dyson sphere but there is just no current breakthrough in our society which would indicate that this makes sense.
In my opinion, we as a society will have to get rid of capitalism first before we will do the next step and just because Musk needs a story to sell to keep his construct alive, doesn't mean its the right time.
Everything you wrote is some definition of hard, but all doable. None of this is purely in the territory of 'known' impossible(like FTL travel).
Now different people have different points where they quit when things get hard.
This is true for even everyday things in life. Quitting triggers exist for people at various points in the ladder. The end of ladder and path both exist, its upto you to decide if you wish to continue climbing, or give up and quit.
Your mileage may vary.
My problem is not the doing thing but the economy of it.
We are nowere near any resource limitation on planet earth for AI Datacenters.
Musk sells this story because he has Starship which needs payload to make financial sense. The payload doesn't exist so he inventes DC in Space.
Its the same thing as SpaceX buying Tesla Cybertrucks.
His old colossus datacenter is a 300MW Datacenter he now rents out to Anthropic because he doesn't even need his own compute. Colossus DC is probably 10x cheaper than his whole Space AI DC Story and will be for a long time.