Inference energy estimator

Estimates of AI's footprint vary by a factor of a thousand depending on assumptions. This calculator makes the assumptions visible so you can argue about them. Nothing you enter here leaves your browser.

Scenario

Using EU average: g CO2e/kWh. Edit any preset's value in the Assumptions table below.

Assumptions

Every number below is a default you can overwrite. Sources are given so you can judge, and argue with, each one.

Editable defaults used in the calculation. DOI: 10.1145/3630106.3658542 (Luccioni, Jernite & Strubell, 2024, Power Hungry Processing).
AssumptionDefaultSource
Luccioni, Jernite & Strubell 2024, measured range for classification tasks.
Luccioni, Jernite & Strubell 2024, measured range for short text generation.
Luccioni, Jernite & Strubell 2024, measured range for long text generation.
Luccioni et al. 2024 report a mean of about 2.9 Wh per text-to-image query, with a range across models spanning more than an order of magnitude.
Extrapolationnot directly measured by Luccioni et al. 2024; scaled up from their per-pixel image-generation figures.
Estimateorder-of-magnitude figure from the chapter statement that a short clip costs roughly a hundred image generations.
Approximatehydro-dominated grid; approximate 2023 figure from public electricity-map data.
Approximateapproximate 2023 figure from public electricity-map data.
Approximateapproximate 2023 figure from public electricity-map data.
1.8 Rarely disclosedoperators rarely publish this figure; 1.8 L/kWh is a commonly cited illustrative value, not a specific operator's disclosure.
Approximatetypical smartphone battery capacity, for comparison only.
Approximatetypical electric-car energy consumption, for comparison only.
Result
—

What to notice: the energy total depends only on the per-query watt-hours and the number of queries; the grid preset and water figure change only the CO2e and water outputs, never the energy. Try coal instead of Norway to see how much the same energy use can cost the climate.