For reference, this is based on figures given by Sam Altman, which are worth as much as going to random.org and asking it for a number between 0 and 100 to use as Wh per LLM query.

Scratch that, the latter is probably more reliable.