you're correct. this video was a pretty good visual exploration of some of it: https://www.youtube.com/watch?v=_qpdUNMt2yg

my personal understanding is that in-orbit compute is perfectly practical up to some obvious limits like the ones you describe. a few reasonably sized clusters up there (tens of kilowatts) doing high priority processing jobs paid by the flop is a great idea. localized compute on existing satellites already does some of this but some earth observation company being able to rapidly scale up image processing for an hour is a great option to have. the really big stuff is just a fantasy.