My thinking is most aligned with yours relative to all the other replies. Short term tulip bulb mania pricing on vidya game hardware is annoying and making me delay purchases. And my concerns for long term are about what all that heavy compute will be used for.

All the grug-brained politics and popular panic hysteria about power and water and heat for AI seems to me to be a transient thing, at least with regards to inference. The power requirements for inference will drop to practically zilch once models get etched onto dedicated hardware and the dynamic state requirments become optimized.

I want to know what the real plans are for all the general purpose compute and memory they are building everywhere. If inference is going to wind up being cheap or free, then that general purpose compute will in the best case be left unused and available for lease like dark fiber was after the over-buildout of the 2000's.

In the worst case, all that compute will be used for probably something awful like surveillance.