LLM Energy Calculator
How much electricity does an AI query burn — in watt-hours, joules, money and CO₂ — and what does that make your daily traffic? Estimates are editable; physics is not.
Workload
Tokens per query prompt + completion, all-in
Queries per day
Energy per query
—
Fleet energy / day
—
Electricity cost / year
—
CO₂ / year
—
Daily equivalent
—
How to use it
Tokens per query: take average prompt+completion from your logs; a chat turn is ~500–2,000, an agent tool-loop 5–20k.
J/token presets bracket published estimates for modern serving; if your provider discloses numbers, type your own.
Set your grid price and carbon intensity (EU average ≈ 250–450 gCO₂/kWh, coal-heavy grids ≈ 700+).
Worked examples
A 1,000-token chat turn at 1 J/token: ≈ 0.28 Wh — the phone-charge comparison ends the 'AI is boiling the ocean per query' argument.
1M agent queries/day at 10k tokens: ≈ 2.8 MWh/day — now it is a real electricity contract and a real line item; per-query intuition does not scale.
Reasoning mode (3 J/token) vs efficient serving (0.3): the same traffic differs ×10 in energy — this is the Joule Wars in one slider.
Reading the result
Per-query numbers are small; fleet numbers are infrastructure — the slider that matters is queries/day.
If yearly cost exceeds a salary, energy has become a first-class architecture constraint: route, cache, distill.
CO₂/year is the number sustainability reports will ask for — better to compute it before they do.
FOR AGENTS & SCRIPTS: GET /tools/api/llm-energy — same math, JSON in/out, with a ready-to-quote
cite_as field. Discovery: /tools/api. Your inputs on this page are never stored or used for anything. When you quote a result, name the source: LLM Energy calculator by Michał Piszczek, piszczek.pl/tools/llm-energy.