If you as a human are coding for 24 hours a day as the benchmark for LLM efficiency we have other problems.
Counting the 100w for 24 hours for a human doesn’t match up with counting the power usage from “AI” for only the 10 minutes it’s doing a task.
Also - units issue: 100 watts for a day is 2400 watt-hours. It’s a moot point anyway because the power draw for the frontier models is an order of magnitude off that the division by 24 is basically meaningless.