Megawatt sounds right for training. But kilowatts for inference. Take a look at tim dettmer’s work (he’s at UW) on int8 to see some of this kind of efficiency work. There’s definitely significant work happening in the open.