Where the energy actually goes syllabus
The unit is joules
Two numbers make an energy claim interpretable: how much, and of what. Most published claims give you neither, and the argument that follows is about which side's ratio you prefer.
Three boundaries, three answers
The same workload has at least three defensible energy figures:
- Chip only — what the accelerator drew. The smallest number, and the one vendors quote.
- Whole machine — plus CPU, memory, storage, power-supply losses. Commonly 1.5 to 2× the first.
- Machine plus facility — plus cooling and distribution losses, which is what actually leaves the meter.
And then there is embodied energy: everything spent manufacturing the hardware. For a device used lightly, that can exceed everything it will ever draw in operation.
None of these is the right one. The rule is only that a comparison must use the same boundary on both sides, and almost none do.
Take a claim apart
Find any published claim about the energy or carbon cost of AI — a news article, a vendor page, a critical blog post, whichever you already believe.
Write down: the absolute quantity if there is one, the boundary, the baseline it is compared against, and whether that baseline uses the same boundary. Then convert it into watt hours per task.
Most claims will fail at step two. That is the exercise.
Ten times what?
A claim states that a model query uses 10× the energy of a web search. What is the most useful next question?
The claim, converted
Your chosen claim expressed as watt hours per task, or 0 if the source does not contain enough information to convert it. Recording zero is a legitimate and common result.