Comment by mike_hearn
18 hours ago
It sounds hard but it shouldn't not make sense.
1. Solving cost of launching mass has been the entire premise of SpaceX since day one and they have the track record.
2. Ingress/egress aren't at all bottlenecks for inferencing. The bytes you get before you max out a context window are trivial, especially after compression. If you're thinking about latency, chat latencies are already quite high and there's going to be plenty of non-latency sensitive workloads in future (think coding agents left running for hours on their own inside sandboxes).
3. This could be an issue, but inferencing can be tolerant to errors as it's already non-deterministic and models can 'recover' from bad tokens if there aren't too many of them. If you do immersion cooling then the coolant will protect the chips from radiation as well.
4. There is probably plenty of scope to optimize space radiators. It was never a priority until now and is "just" an engineering problem.
5. What mass manufacture? Energy production for AI datacenters is currently bottlenecked on Siemens and others refusing to ramp up production of combined cycle gas turbines. They're converting old jet engines into power plants to work around this bottleneck. Ground solar is simply not being considered by anyone in the industry because even at AI spending levels they can't store enough power in batteries to ride out the night or low power cloudy days. That's not an issue in space where the huge amount of Chinese PV overproduction can be used 24/7.
> There is probably plenty of scope to optimize space radiators. It was never a priority until now and is "just" an engineering problem.
It's a physics problem, as others pointed out, but even if we take it as another "just an engineering problem", have a look at the Hyperloop. Which is similarly just a long vacuum tube, and inside is like an air hockey table, not that big a deal, right?...
Musk's companies never tried to make the hyperloop, they never even started on it. SpaceX is a bit different.
so spacex worked on an orbital data center?
1 reply →
> There is probably plenty of scope to optimize space radiators. It was never a priority until now and is "just" an engineering problem.
Well, it's a physics problem. The engineering solution is possibly not cost efficient. I'd put a lot of money that it isn't.
That bit reminded me of someone who wanted us to design a patch the size of a small postage stamp, at most 0.2mm thick, so you could stick on products. It was to deliver power for two years of operation, run an LTE modem, a GNSS receiver, an MCU, temperature and humidity sensor and would cost $0.10. And it would send back telemetry twice per day.
'A mere matter of engineering'.
9 replies →
Not only is it not cost-effective, it's pointless (in this context).
Radiators works almost just as well on Earth. Convection and conduction more than make up the difference.
What makes you so sure? SpaceX already has thousands of 6 kW networking racks flying around in LEO and they dissipate their heat just fine, and are plenty cost-effective. You think they can't do any better than that with a new design specifically optimized for computing rather than networking?
Probably, but they likely can't do better than we can do on Earth. Networking in space offers specific advantages that are not easy to replicate on Earth. Data centers in space don't have clear advantages beyond easily debunked ideas about cooling and power.
1 reply →
The solar panels on the newest satellites can deliver 6kW but the power that satellite actually uses is less. The satellite is only using 300W[1] during the dark phase of it's orbit when it can use it's entire mass to cool down. Is that limit because of the battery or is it because the satellite needs to radiate all the heat it acquired from the other half of the time in the sun?
[1] https://lilibots.blogspot.com/2020/04/starlink-satellite-dim...
2 replies →
3. There are WAY more things to get corrupted on a computer system than tokens. And non-determinism does NOT mean it’s tolerant to faults. Random values are intentionally introduced at the right moment for LLMs.
“just an engineering problem”
Sounds a bit like that Dilbert where the marketing guy has sold a new invisible computer and is telling the engineers to now do their job and actually make it.
> It sounds hard but it shouldn't not make sense.
It does not make sense.
The question isn't "can you mitigate the problems to some extent?", it's "can you see a path to making satellite data centers more appealing than terrestrial?"
The answer is a flat out "no," and none of your statements contradict this.
Terrestrial will always be better:
1. Reducing the cost of launches is great, but it will never be as cheap as zero launches.
2. Radio transmissions have equally high bandwidth from Earth, but fiber is a better network backbone in almost every way.
3. Radiation events don't only cause unpredictable data errors, they can also cause circuit latch-ups and cascade into system failure. Error-free operation is still better in any case. Earth's magenetosphere and atmosphere give you radiation shielding for free, rad-hard chips will always cost more than standard (do they even exist for this application?), and extra shielding will always cost more than no shielding.
4. On Earth you can use conduction, convection, AND radiation for cooling. Space only gets you marginally more effective radiation.
5. Solar is cheaper on the ground than in space. The increase in solar collection capability per unit area in space doesn't offset the cost of launch: you can get 20kW of terrestrial solar collection for around the price of a single 1U satellite launch, and that solar production can be used on upgraded equipment in the future. Any solar you put on a satellite gets decommissioned when the inference hardware is obsolete.
And this ignores other issues like hardware upgrades, troubleshooting, repairs, and recycling that are essentially impossible in space, but are trivial on the ground.
> Solving cost of launching mass has been the entire premise of SpaceX since day one and they have the track record.
They have to solve for it being cheaper to launch and operate in space vs building and operating a datacenter with its own power generation on Earth.
My lord a sensible comment her. A hearty upvote.
It really isn't. It's plainly incorrect and ignorant of the actual problems.
I have no expertise is this area, so I'm not getting into whether or not this idea makes sense.
That being said, this statement strikes me as missing the point:
> Solving cost of launching mass has been the entire premise of SpaceX since day one and they have the track record.
As I understand it, SpaceX has a good track record of putting things into space more cost effectively than other organisations that put things into space.
That is not the benchmark here.
It doesn't matter if Musk can run thousands of data centres in space more cost effectively than (for example) NASA could. It matters whether he can do it more cost effectively than running them on earth.
The cost of "launching" mass on Earth is not zero, though.
I didn't suggest that it was.
I don’t think that statement was missing the point. As you point out, what matters is the total cost of ownership of the system. The cost of launching mass into space today isn’t the only reason terrestrial data centers are more cost effective today, but it’s the main one. If you make it cheap enough to send giant solar arrays and radiators to space, the other costs of operating in space may start to look like a small price to pay to eliminate the need for inputs like land and batteries.