For sake of argument, let's assume that the six-figure (seven, maybe?) price tag on the hardware was no longer a factor and it was possible to train the models locally, I think the sources of the content everyone is trying to train their local model against would quickly shut down the inundation of traffic they're receiving from the hundreds of thousands of individual computers all trying to build their own "unlimited" model.

The computing requirements enforce the current reality that the training of models will be centralized.

This places a larger ethical burden on those central entities, IMO.