What you describe needs continued investment into datacenters, which will be a no go when the bubble pops. The hardware doesn't last forever (avg. life expectancy is somewhere around 6 years iirc), and running that hardware also isn't free at all.

Add to that that you can nowadays run Cyberpunk 2077 on medium detail with 60 FPS on a freaking Android Phone with FEX and Proton, and personal computing is far from dead, regardless of what they want.

Edit: I'd rather expect mainboards that make HBM available for consumer PCs to pop up, and that hardware for personal inference drops in price dramatically.