So commercial artificial intelligence operations (OpenAI, Anthropic and so on) require some high-performance processors with high power and cooling requirements.
Now there is a trend of talking about putting such data centres in space. SpaceX and some others are betting the farm (and probably your funds) on this.
A few months ago, I was seeing articles like this one, saying the cooling would be a limiting factor because there’s no convection up there. Plus: solar power is still not as abundant, and radiation will complicate matters.
But this upbeat article (MSNBC) from this week doesn’t even mention cooling as a factor.
So which is it? Has anything changed with Moore’s law or the laws of physics? Or were/are the doomsayers wrong?
Cooling would still be a huge problem. Even if the launches were free (they’re not), that alone would be enough to make space datacenters more expensive than terrestrial ones. Anyone talking about putting datacenters in space is either an idiot, or is trying to sell something to idiots (or both).
Elon Musk just became a trillionaire at least partly based on that idea, and he’s no idiot.
The idea is that, when the AI takeover happens, the data centers will have laser and anti-missile systems in place to destroy any feeble human effort at rebellion. LOL
Reminder that this is FQ.
Then he’s selling things to idiots.
Pure speculation on my part, but I think some of the appeal of orbital AI is to (eventually, potentially) be free from earthly laws and regulations. I don’t know if it involves lasers, though.
It probably wouldn’t help, though, since the most burdensome regulations would be against things that simply aren’t possible in space. A space-based datacenter won’t be using excessive water, and it won’t be polluting the air with its power generation, and so forth. And it’d be much cheaper and easier to just bribe some jurisdiction on Earth (any jurisdiction, and there are some where the bribes would be really cheap) to not regulate you.
His actual job, regardless of what various PR and image consultants have convinced us, is to sell things. And he appears to be pretty good at that.
If his talk about other technologies had come to pass, we’d have already had fully autonomous vehicles and self-sustaining colonies on Mars for several years now.
His current talk about installing more than the total terrestrial compute capacity into orbit every year starting within the next 2 or 3 years is flat out fantasy. Aside from the operational costs or cost to launch to orbit or even launch capacity (which are certainly not negligible), there is insufficient fabrication capacity worldwide to build even a fraction of that much equipment on an annual basis and we aren’t capable of getting up to that point for several years, even assuming the entire world cooperates on it (again, not a realistic assumption itself).
So, yes, ‘selling to idiots’ is a pretty good bet if you are using his predictions as a basis for practicality or likelihood
Aside from the question of cooling, how good is the communications with people on the ground? Most data centers have fiber connections but how good would the connection be?
Presumably communications would be through Starlink or equivalent. So, Gigabit or somewhat higher level would be achieveable but it’s going to be 2 or 3 orders of magnitude less than the big data centers here on the ground.
Bandwidth is definitely another limiting factor that would have to be considered (on top of launch costs, maintenance, etc). If sufficiently large volumes of data need to be processed, it may actually be faster to send a mission up with hard drives. Even down here, physically connecting a drive and then physically transporting it to another location is often faster than trying to transfer the data over the network. As we used to say, a 747 has better bandwidth (but higher latency).
This cannot be discounted. Space-qualified hardware for satellites is significantly more expensive than “normal” industrial grade or even military grade. And is normally 20 years behind current state of the art. Plus you don’t get to replace parts that go bad if they’re on a damn satellite.
Although as datacenters go, AI datacenters have a fairly high ratio of computing resources to bandwidth used. A common AI use case is a fraction of a kilobyte of text input, then a whole bunch of computation on that text, and then a fraction of a kilobyte of text output.
Altogether, though: It would be possible to design and build a space-based datacenter. But even if you did that, if you then, instead of launching it, just put the whole system on a ship and parked it in international waters, that would still give you all of the same benefits, for a fraction of the cost. Nobody talks about floating datacenters, because they’d be completely impractical, but they’re still less impractical than orbital ones.
How big would the solar panels have to be, even just to power the hardware (leaving cooling out of it for now)? Data centers are giant power hogs, right?
Sort of big power hogs. A lot of the power they use down here is for the cooling, rather than to directly power the computation.
So, less than maybe we would think but still a lot. Note this isn’t a real, practical proposal yet but closer to pie in the (literal) sky prospect for which the engineering is, at best, in the ‘concepts of a plan’ stage.
Lifecycle management of these things is going to be stupid simple. You don’t maintain it. You don’t repair it. You don’t upgrade it. You use it until it breaks and then (of you’re a responsible orbital citizen) de-orbit it or kick it to a graveyard orbit and launch another.
The problem is that, AFAIK, AIs improve with time (assume they’re allowed to continue training) and you would lose that as soon as enough of the hardware of the satellite punks out. It seems like a waste to discard a mature AI just because you literally sited its hardware in space. Persistence is not an option when you’re only one hardware error away from unplanned forcible decommissioning.
Oh, yeah, there will be no second-hand market for used AI datacenter hardware, either, so I guess we continue to enjoy ridiculous scarcity and prices on computing hardware.