The course started from one question: how much does a sentence sent to a model consume. The answer is a chain of calculation where every link has its source and its boundary.
The journey, in short
Writing costs more than reading. The prompt is read in one pass, the reply is written token by token: in the reference measurement of chapter 4, an output token costs about 35 times an input token. Context then adds to every token written. On a coding agent step with 126,180 tokens of context, it is 68% of the energy (chapter 12).
To go from watt-hours to carbon, you still have to pick a PUE, an electricity emission factor and an accounting method (chapter 13).
What will not move
The figures in this course will age fast. The mechanics much less: as long as models write token by token, a long context will weigh on every token produced. Faced with a new figure, the questions stay the ones in the course: what it covers, what hardware it was measured on, which electricity factor it uses.
The parameter sheet lists the data still missing, with the protocol that would fill it. If a figure has moved since, the feedback button on the site sends it to me.