OK, I see I should have used "trillions of dollars worth of hardware capable of performing not-so-precise floating point calculations at the exaflops level" as the basis for my question. The hyperscalers are making a singularity-or-nothing bet on this stuff; there are no other uses for this much hardware of this type. If there were, such hardware would already be in place.

(IMHO, the fact that AI models tolerate absurdly-low precision means that we're doing something wrong and wasteful in our entire computational approach to the field, but that's neither here nor there.)