Reasoning has an energy cost. Report energy per task.
Research finds long, reasoning-style queries use many times the energy of typical ones, and that efficiency gains can offset much of it.
The International Energy Agency projected in April 2025 that global data-centre electricity use will roughly double by 2030, to about 945 terawatt-hours.
A 2025 study by Oviedo and colleagues estimated a median of 0.31 watt-hours per query for large models on current hardware, rising 13-fold to 3.91 watt-hours when queries are about fifteen times longer than typical, as with extended reasoning. The same study found 8 to 20-fold energy reductions available across models, serving software and hardware.
ESS view: organizations reporting the footprint of AI should measure energy per completed task, including reasoning time, rather than relying on estimates of training energy alone.