Cooling is relatively easy, you just need radiators which are passive, and essentially reduce to a launch cost penalty. You are right that they can't be serviced, but that is missing the point of orbital data centers. The whole point is that you can build hundreds of thousands of these in a factory and launch them in a scalable manner. The power, cooling, etc. comes for "free". In the long run, as the cost of the chip, launches, etc. goes down, orbital data centers will scale better terrestrial ones.

As a side note, I don't understand why I keep seeing these wrong arguments on HN repeatedly. Like everything mentioned in this thread can easily be fact checked. Radiative cooling is solved, launch costs are going down, so power costs will pay themselves back very quickly, etc.

You can argue about specifics, like chips will get more sophisticated + power efficient and fabrication will be the true longterm bottleneck, or SMRs/fusion could reduce energy bottlenecks, but talking about cooling as if convective cooling is the only option is just nonsensical.

It’s not an unsolved physics problem. Every satellite in space has to deal with it and even the ISS deals with it by having massive radiator arrays that face perpendicular to the sun.

The problem with data centers in space is one of materials science and engineering: how to make radiators large enough and effective enough to cool it while also being economically feasible, both in terms of construction and getting them up there in the first place.

We can make a space data center right now. It would just be terrible and expensive.

It's so, so cheap to buy tap water and dump it on the heat exchange.

GPUs depreciate super fast. It might last 5-7 years but it's already outdated at 2-3

Also space has more radiation

5 years is a Starlink's typical lifetime. Data center satellite lifetimes will probably be shorter. Demise sooner, replace more often. GPUs get more energy efficient every year and leaving the slower, hungrier chips up there much longer than 3 years seems wasteful given the cheap cost of launch.

I think this could be done at an interesting scale even on Falcon 9 alone. If Starship does even 20% of its early design goals, it'll beat Falcon 9 and we could see orbital servers being demised and replaced every 3 years, maybe even 2, for ones with abnormally high failure rates.

Now, whether or not this will all make money in the end has a lot to do with what's going on down here on terra firma and how long it takes to get useful capacity into orbit.

(It's taken 7 years to get Starlink capacity enough for serving 10M customers. Verizon FiOS did 10M in 5 years. AT&T Fiber took 4-5 years to deploy to 10M. So, space isn't a lot slower than terrestrial.)

uhh no I dont think the road vehicle is harsher than a rocket launch

Ars had an article that cited some HPE testing on the international space station that said regular hardware is _probably_ fine up to about 5 years

Definitely not definitive but it's plausible current hardware could survive with minimal modification

The cooling is not an issue for their current designs. Look at their AI1 Satellite specs. They clearly have the cooling figured out. The thing is that it's not a datacenter, it's a single rack. A single rack that weighs multiple tons.

You can figure out the weight of the thing based on the total power output, and "power to weight ratio" from SpaceX's own diagrams. Then look up how much it costs to launch per ton, and even look up what they are projecting it will cost with Starship. Even if they get costs down, it's still astronomical. I just can't figure out who would pay that much money to put a rack into space. There's no way the power savings are worth it. Unless you have some niche where you need your workload in space, I can't see the value at all.

The same way that xAI has compute on the ground and there's such demand that the compute exists so it will be sold. I think SpaceX can conceivably sell the space compute easily. This goes with the business strategy of Elon which is just sell services to himself. He sells rocket launches to himself with the excuse of creating a space datacenter. The same way he sells cybertrucks to spacex because the public won't buy them. I'm pretty confident it's definitely cheaper to put the compute in Texas, but might be viable enough to sell himself rocket launches and make his company worth more to leverage for more ventures. I'm not a business person I don't really get how it's all so valuable if the only demand is itself.

Training, no. Latency is too high.

Inference the latency becomes trivial.

Other things, I suspect latency is too high again.

For reasons already outlined, I have doubts about their math. Targeting $250/kg payload costs is ambitious for a rocket that has not yet successfully reached orbit or proven cost-effective.

Even if we do somehow succeed at affordably dumping tons of GPUs into orbit, what do we do about the Kessler Syndrome?