

Definitely a possibility! But dealing with “only being a normal profitable company” is a very different problem to “oops, we were selling $10 for $5 and VCs have stopped giving us money to burn, and people are using self hosted models too”, which is the possible outcome for the big AI labs.

Your point about poinitng (ha!) is incorrect, its pretty trivial to maintain pointing at the target. Hubble achived 7mas pointing accuracy over extended periods (thats ~0.000002degrees) with technology more than 30 years out of date. That gives you ~1.2m accuracy from geostationary orbit, which seems fine.
The real point is getting a mirror which is large enough and perfect enough into orbit is completely infeasible. As you rightly say, the maximum potential power it can provide is equal to solar insolation time its area.