I dunno, why is 90% the limit? Maybe they can set something up in the chips to use 1/10th the power and set something up in the model design to use 1/10th the power and so we use 99% less power than today.

After all GPUs were not really designed with AI efficiency in mind, and current AI architecture was not really designed with efficiency in mind, so there's probably some things that can be done. Over 5-10 years, who knows how much more efficient things could get?