Apparently the whole “space datacenters” idea is a thing that some people are buying into.

Again.

I felt compelled to comment.

Now, I haven’t actually gone out and done the numbers or anything, or thought about it very hard or anything, but I have two big concerns that I suspect may not have been fully taken into account by people advocating for this “solution”.

(there are other concerns, but these are the two that immediately leap into my mind)

1. Cooling. Space isn’t really exactly “cold” like the popular conception says. Getting rid of heat is actually a fairly big problem, and it’s very easy to overheat things: In space, there’s no conduction of convection – the only way to dissipate heat from your GPU cluster is via radiation. This is why the ISS has huge radiators on it to dissipate the heat they generate. Otherwise things would get pretty toasty – by which I mean “unsurvivable” – pretty quickly, since every watt of heat generated would just add up.

Radiation is the least efficient way of dissipating heat. It also requires a lot of surface area for the heat to radiate from. That’s why those radiators on the ISS are so big.

Science fiction tends to completely ignore these facts, or at best it’ll handwave some magical heat dissipation solution that’s about as practical in the real world as inertial dampeners. But the reality according to our current understanding of physics would look quite different to what you see in the movies: Those huge fusion engines or matter/antimatter reactors your ship uses are going to generate a fucktonne of heat. And that’s going to require some huge radiators if you don’t want to literally cook your crew surprisingly quickly. In reality those star destroyers would look very different because they’d need a ton of radiators. In fact, heat is one of the prime reasons why stealth in space is so extremely difficult, bordering on impossible if your ship is crewed and you want it to stay that way.

And it’s not just about crewed vessels, either. A machine might be able to run a lot hotter, but electronic components tend to not like excessive heat either. Plastics and solder will melt a long time before your superstructure does.

Fun fact: Datacenters, and especially processors and GPUs, generate a lot of heat.

Sure, it’s possible to dissipate the heat from a bunch of GPUs with radiators, but I strongly suspect that for a concentrated cluster of GPUs, like you’ll want to have for, say, “AI” datacenters, will require a LOT of surface area for radiators – far more per square meter of usable area than the ISS, which isn’t filled to the brim with heat-generating processors.

Did I mention that convection isn’t a thing in space? So of course your CPU/GPU fans are useless there. You’ll need to use some more complex solution like heat pipes and/or liquid cooling. Also don’t forget to account for the microgravity environment when you design that liquid cooling system,

And that’s before we start talking about surface area required for solar cells to power these GPUS – which are also notoriously power-hungry.

TL;DR: I haven’t done the math, but I suspect your GPU-cluster-in-space-datacenter is going to have much more area and weight used on radiators and solar panels than it will on GPUs.

2. Lifecycle. I think you might be surprised how long commodity computer hardware doesn’t last. And even server-grade hardware that’s designed to run 24×7 doesn’t last all that long, really. A hard disk has an average lifespan of something like 5 years IIRC.

And that’s in an environment where it’s shielded from a ton of radiation by earth’s atmosphere and ozone layer. In space it will last even less time due to the cosmic rays and the solar flares and whatnot.

Unless you spend a ton of money hardening it, of course.

Fun fact: when the rest of the world was using gigahertz circa-pentium3 machines, the fastest radiation hardened machine you could buy was something on the order of a 486. A lot of the nasa stuff was running on radiation-hardened 386 machines for a LONG time after the 386 was totally obsolete on earth. I’m pretty sure the space shuttle used a rad-hardened 386 right up until its last flight in 2011. I’d be unsurprised to learn that there are still 386s running on the ISS.

I did a quick search and found some more modern examples, such as the a radiation-hardened ARM chip, the SAMRH71. Ooh yay, ARM! A modern architecture! Just what we need, right?

It runs at 100Mhz. with 1MB RAM and 128K of flash. Ooo such impressive stats! I’ll leave it as an exercise to the reader to compare those numbers with the requirements for any machine learning model currently being spruiked by the industry.

I also found the RAD5500, which is a 64bit powerpc chip that can run up to the blinding speed of 450mhz! Wow! With only 6 of those you could almost do as many FLOPS as a single core of my 12-core laptop!

I didn’t look very hard. I’d be unsurprised if there are faster rad-hardened machines out there. But I can pretty much guarantee that they’re going to at least an order of magnitude slower than the cutting edge, and they’ll cost at least an order of magnitude more than the equivalent server-grade hardware.

But there are other options! The cubesat approach to radiation hardening is to do nothing – you just stick a mobile phone in there and hope that it doesn’t get fried by cosmic rays on day one.

Fun fact: cubesats have a much higher failure and DOA rates than most other satellites. There’s probably no correlation there, though. Could just be a coincidence.

Alternatively, you could just put two or more identical not-hardened machines up there, and use the redundant systems and have things like voting mechanisms to do a sort of “software radiation hardening”. I seem to recall reading about experiments with that approach that indicated it worked. And it’s only ~double the cost! And double the space! and double the weight to launch into space, and double the heat and power requirements! And still just as susceptible to things like solar flares, except now there’s 2+ machines for the flare to destroy.

But really we can set all that aside, because even if we manage to come up with a perfect solution to the radiation issue, there’s still the other issue I see: Hardware obsolescence. Nobody wants to be using last year’s GPUs. Particularly in a field moving as rapidly as LLMs are.

…so, what’s the plan? you’re going to spend a ton of money launching these big GPU clusters into space, and a ton of money and weight (which means more money) on radiators and solar panels, and then just let all that hardware die in a couple of years because the chips are obsolete? (Assuming, of course, that the radiation environment doesn’t get to them first)

I haven’t run any numbers, so I’m not prepared to say it’s impossible. But I’m, let’s go with “extremely skeptical” of your business model. I’d be super fascinated to see your numbers on things like heat dissipation and ROI versus expected GPU lifetime.

You DID run those numbers before running your mouth off, right? I assume you did, otherwise you’d just be engaging in nonsense science fiction speculation. That, or maybe you’re just ignorant of some basic facts about operating in space. So I’ll be super keen to see that spreadsheet you definitely must have. Can’t wait till you release it!