I am by no means an expert but I would imagine the answer to this question depends a lot on what you mean by 'a model'. Do you mean 'all of (e.g.) OpenAI's compute over an hour'?
Humans run at about 100 watts. A 4090 GPU alone is 450 watts.
Humans run at 100 watts while performing simple tasks, the figure approaches 550 watts while performing demanding cognitive or demanding manual tasks.
Hence why I bring up potential energy savings is this other HN thread: