Power used is important but so is speed. For example right now I am sure the GPT cluster consumes power in the order of megawatts but you have to remember it serves thousands of users simultaneously.
Yes, so energy use per task. I thought that was clear.
Optimizing energy use per task at will be a waste of time at this point of time, this will come later. The metric used by OpenAI seems to be more useful right now.