What would they be cheaper on? Solar panels are a little bit more effective and they will have a 24/7 coverage if placed in the correct orbit.

However, they would be much harder to cool (space is cold, yes, but heat transfer in vacuum does not work easily and most large structures, such as ISS, require dedicated cooling radiators that take up a large amount of space.) The launch costs would be still very high, maintenance impractical and the large, large surface area of solar panels and radiators would just be primed for being struck by debris.

What orbital data centres are though, is a good dream to sell, a fine way to dismiss environmental concerns of data centres on the ground - “We’re soon going to start putting them in space, but just for now we have to build them on earth. Please approve our requests.”

the short answer is no, general-purpose space datacenters are a non-starter. eg, you're never going to open the AWS console and decide whether you want to deploy a VM to us-east-2 or leo-1.

however, there is a narrow use case for wanting to run more powerful hardware on satellites that would be launched anyway.

for example - you have 2 countries, Alicetopia and Bobistan. they border each other, separated by a big desert, and are on unfriendly terms. their militaries want to make sure they never get surprised by an invasion force attacking them.

Bobistan launches a satellite (or several) that flies over their border region once a day (or more, depending on orbital geometry) and takes pictures (visual-spectrum at least, possibly also infrared, SAR, etc).

those pictures get downlinked and analyzed to answer the question "is Alicetopia building up a military presence on our border to prepare for an invasion?"

this used to be done manually, with people actually staring at imagery to try to find rectangles that looked like tanks. back in the early Cold War days, this was done using physical film that was dropped from orbit, looking for ICBMs. obviously now it's all done with machine learning algorithms.

downlinking those daily images isn't cheap, especially when the steady-state behavior is "nothing interesting here, just a big stretch of desert".

as a result, there's a desire to run a relatively lightweight ML model on the satellite itself, to answer the question "is any of this imagery worth downlinking at all? and if so, is any of it high-priority for downlinking immediately and flagging for human attention?"

for flight safety reasons, you'd want that on a separate GPU/TPU-like processor, so that your rad-hardened CPU that runs the mission-critical parts of the flight software won't be affected by anything that happens with the ML processing.

but that relatively narrow use case definitely doesn't justify the magnitude of the current hype cycle.

The cooling problem is vastly exaggerated, you need around 0.5x the area of your solar panels in radiators.

I think AI inference in space is definitely possible, but it's very unlikely we'll get launch costs cheap enough that they make economical sense.

You can solve all of them far cheaper and easier on land.

https://panthalassa.com/

Data is faster, power is cheaper, cooling is free. There's no reason for it other than to juice spacex stock. It's just another Elon scam to pump stocks. I don't know why anyone wastes breath talking about anything he says.

If your business is selling transport to space, drumming up a new line of business makes sense. If material science keeps progressing, (the metal used in a 2026 car did not exist 10 years ago. Nor did 2026 solar panels.) the values in the equations change. Whether they change enough to justify launching into space remains to be seen. Gigafactory to make batteries for Tesla seemed like an insane investment when it was announced, but makes a lot of sense in hindsight.