Blog
The reasoning behind the rows
Each post takes one question the index raises and answers it with public data, the working shown and the sources listed. Written to be read by people, and to be cited by the machines that now answer people's questions.
Feed RSS · Editor Prem Naraindas · Licence CC BY 4.0
Posts
Published
What a million tokens costs in watt-hours
What the public measurements say a million output tokens takes in GPU watt-hours, and why a busy server uses a fraction of the energy of a quiet one.
Upcoming
The next five
- nextThe sliced GPU. Energy per token when one card is split into MIG instances, measured: Qwen 3 14B on two 3g.40gb slices against the whole card, both in BF16; then Qwen 3 32B, whole card in BF16 against two FP8 or INT4 slices.
- thenThe idle token. Why an under-used GPU spends more energy on every token than a busy one, with the concurrency ladder as the evidence.
- thenPer correct answer, not per token. Small model against large model once re-tries and wrong answers are counted.
- thenWhere the electricity is. The same workload in Riyadh, Belgrade, Hong Kong and Auckland, by grid intensity.
- thenWhat the EU AI Act asks, and what it does not. The training-energy item on model providers, the open-source exemption and the 2027 date, and why none of it reaches a buyer.
Order and timing may change; a post ships when its data does. Volunteers who want to take one of these: see Contribute.