One of the latest points of "what if" discussions surrounding AI advancements and their need to absorb every watt of computing power revolves around putting datacenters in space for power and cooling benefits, and itvsounds like a grand idea on its face.
But going by the thoughts and online calculator of Andrew McCalip from Varda Space Industries, his personal opinion is that the economics of such a proposition are a proverbial kick in the teeth, and should make any sensible observer ask "why?". A technical deep-dive by Andrew Côté from Hyperstition Incorporated essentially concludes the same.
On the face of it, the idea has merit — solar power is plentiful when you don't have to deal with that annoying day/night cycle, and space is pretty darn chilly, minus some "minor" engineering challenges. Even with extremely optimistic projections, plus the hope that reusable rockets will continue to evolve, and ignoring many associated costs, McCalip's math puts a SpaceGPT installation at over three times the cost of its earthbound equivalent: a baseline $51.1 billion for space, compared to $15.9 billion on terra firma.
His online calculator is extremely detailed, letting one configure 15 parameter sliders to try and get a feel for how good (or bad) the idea is. Said parameters range from adjusting the launch cost to Low Earth Orbit (LEO), satellite size, GPU failure rates, as well as comparable earthbound figures.
It should be noted that McCalip's math only accounts for the bare logistical differences of a launch versus a standard build, with a lot of assumptions about engineering challenges, and he remarks as much. McCalip goes as far as calling the notion "FOMO and aesthetic futurism", and that "people are using back-of-the-envelope math, doing a terrible job of it, and only confirming whatever conclusion they already want."
That's a shot across the bow at the (literal) pie-in-the-sky ideas of using LEO satellites to host datacenters. It's also easy to understand the frustration of those who wish that datacenter buildouts would move faster. For terrestrial data centers, tons of time is spent navigating legal and literal ground, worrying about water supplies, ecological and carbon output concerns, and so on and so forth.
Solar power is always available, so its effective price can approach near zero, helping with lowering the operational cost. There's no regulation regarding land (or barely any regulation at all), and no need to worry about passing transmission lines either. Cooling looks easy on the surface level, with space radiators being extremely efficient. Plus, with the advancements coming from SpaceX et al are reportedly hovering around $1000 per kg to launch into LEO.
Stay On the Cutting Edge: Get the Tom's Hardware Newsletter Get Tom's Hardware's best news and in-depth reviews, straight to your inbox. Contact me with news and offers from other Future brands Receive email from us on behalf of our trusted partners or sponsors
Even though McCalip's analysis already puts the kibosh on the financial side, Andrew Côté's exceedingly deep dive into the engineering difficulties faced by such an enterprise delivers the killing blow. That's also an excellent read , as the list of things that can go wrong with setting up a datacenter in space can be reasonably summed up with "yes." Did you know, for example, that AI accelerators tend to up and die a lot when used at full tilt, and can't exactly be replaced in 5 minutes in space?
In these times where AI market valuations are climbing ever higher, it's arguably refreshing to see some down-to-earth thinking from those in know. Do check out McCalip's economic analysis and Côté's thoughts on the engineering challenges to be faced. Maybe one day the numbers will add up, but for now, it seems AI computing is best performed on good old terra firma.
... continue reading