The environmental footprint
small per prompt, large in aggregate
AI uses energy and water. A climate negotiator who uses it should be able to say how much. Your own use is small. Total demand is large, and growing.
Measured, per prompt
energy
carbon
water
Google's measurement of a typical text prompt to Gemini. It includes idle servers and data-centre overheads, which most estimates leave out. Google, arXiv:2508.15734, August 2025.
01Your own use is small
One exchange uses about as much as running a laptop for a few seconds.
A full day of heavy use is still less than a short car trip.
Polly shows this figure for every account, and publishes how it works it out below.
02Total demand is large
- Data centres are adding electricity demand faster than grids are switching to clean power.
- New data centres are going up in places already short of water.
- The companies that publish the small per-prompt figures have seen their total emissions rise since they started building these systems.
03What intelligence costs to run
- on the frontier
- beaten by a model on the frontier
- closed model, placed from its price
- one model at different efforts
Up and to the left is better. A filled dot is on the frontier: no other measured model is both smarter and lighter to run. An ✳ is a closed model placed from its price, a rough estimate that cannot set the frontier. Hover or tap any dot for its figures.
Scores: Artificial Analysis Intelligence Index, 27 August 2026. Energy: The ML.ENERGY Benchmark v3.0.
How this is worked out
- Energy is measured, not guessed. Artificial Analysis counts the tokens each model writes per question. Energy per token is based on 33 real meter readings from the ML.ENERGY benchmark.
- Reasoning effort is the setting you control. Each model appears once per effort level tested. Kimi K3 uses 4,197 tokens on low and 25,474 on max, and both are on the frontier.
- How a model is run matters more than which one you pick. Depending on the server setup, the same model can use ten times more energy.
- Price is not a guide. Cost and energy move together (r = 0.74), but the price per unit of computing varies 60 times over. Qwen3.6 35B A3B reasoning and Ring-2.6-1T cost about the same, yet one uses 19 times the energy of the other.
- Who is missing. If a lab does not say how much of its model each token uses, we cannot work out its energy. Every model on this chart publishes its weights.
The full method, and the closed models placed from their price, are in chapter 13 at ai4cop.org/read#environmental-footprint.
04The calculation, shown
Energy per token
Water and carbon
Reading your prompt costs about a quarter as much as writing the answer. Re-reading a prompt already seen costs a tenth. The three are counted separately. Polly's model uses 18 billion parameters per token. A typical exchange uses about 0.10 Wh and 0.11 mL of water, just below Google's figure above.
05What this leaves out
Training the model. The water used to make the electricity. Your own device. The carbon from building the hardware.
Take every figure on this sheet as correct to within a factor of two, not as an exact measurement.
06What to do about it
- 1
Turn reasoning off when you do not need it. With it on, the same model can use twenty to forty times as many tokens.
- 2
Ask providers for their figures, including water, and which power grid serves your requests. Fair questions for a supplier, and few people ask them.
- 3
Publish your own figure. A delegation that reports the footprint of its own AI use can fairly ask others to do the same.