Same boundary before comparing
Published figures count different things. Some measure only the AI chips. Others add the host server, idle reserve machines, cooling and power conversion, or the manufacturing of the hardware. TokWatt records all five for every estimate, then scales each value to one boundary, the server plus facility overhead.
chip-only × 1.72 · server-only × PUE 1.13 · already complete × 1
The 1.72 factor is Google's measured ratio between chip energy and its full serving energy. The PUE is the median of the hyperscaler values used in the literature. Values whose boundary is not disclosed are left as published and flagged.
Reference values
Each task has a reference set: the estimates that match its definition, such as the default models of the major assistants for a chat answer. The reference value is the median of that set. The range runs from the lowest to the highest interval the sources report. Tasks resting on a single source are marked.
The billion rule
One billion people using E watt-hours a day draw E ÷ 24 gigawatts on average, around the clock. Every 24 Wh per person per day is roughly one large nuclear reactor.
1e9 people × E Wh/day ÷ 24 h = E/24 GW
Limits
Few developers publish per-query energy, so many rows are models of production systems rather than measurements of them. Agent estimates come from token logs, and the energy of re-reading cached context is the largest unknown. Training, user devices and networks are outside the boundary.