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.