The main reason why we can't do that now is because we require models to be digitally reproducible (IMHO, but also read Geoffrey Hinton's mortal computing).

The energy cost come from error correction as much as training algorithms.