There's two scenarios going forward that I can see.

Either there's a massive reduction of inference and training costs due to new discoveries and then those big hardware hoarders have no moat anymore or nothing new is discovered and they won't be able to scale forward.

Either way it doesn't sound good for them.