2,024 · Last reviewed 2026-05-24
AI Footprint Index
One AI prompt is invisible — but it draws real electricity, evaporates real water to cool the servers, and emits real CO₂. This index translates the measured cost of common AI tasks into comparable units, then shows what a light, medium and heavy AI habit adds up to over a year. Every figure is inference-only (using the model, not training it), water is on-site scope-1, and the numbers are estimates drawn from the sources below.
Lightest task
Gemini prompt (measured)
0.24 Wh
of energy per use
Heaviest task
Reasoning ("thinking") prompt
7.6 Wh
of energy per use
Energy per AI task
| Rank | Task | Energy | Water | Carbon |
|---|---|---|---|---|
| 1 | Gemini prompt (measured) | 0.24 kWh | 0.4 L | 113 g |
| 2 | Short AI chat | 0.30 kWh | 0.5 L | 141 g |
| 3 | GPT-4o query | 0.42 kWh | 0.8 L | 198 g |
| 4 | Long document (~10k tokens) | 2.50 kWh | 4.5 L | 1,178 g |
| 5 | AI image generation | 2.91 kWh | 5.2 L | 1,369 g |
| 6 | Reasoning ("thinking") prompt | 7.60 kWh | 13.7 L | 3,580 g |
Figures are per 1,000 uses, so the numbers stay readable. Water uses the 1.8 L/kWh industry-default cooling factor; carbon uses the world-average grid (471 gCO₂/kWh, 2024).
What a year of AI use adds up to
Three stylised users, each with a fixed monthly mix of tasks, projected over a year and translated into everyday activities. A heavy daily user’s annual AI footprint is real but still modest next to a single long-haul flight or a year of driving — the point is comparison, not alarm.
| User | Energy / year | Water / year | Carbon / year | In everyday terms |
|---|---|---|---|---|
| Light user | 0.2 kWh | 0 L | 0.1 kg | ≈ 17 phone charges · 0 showers · 0 km driven |
| Medium user | 2.8 kWh | 5 L | 1.3 kg | ≈ 230 phone charges · 0 showers · 5 km driven |
| Heavy user | 22.5 kWh | 41 L | 10.6 kg | ≈ 1,878 phone charges · 1 showers · 43 km driven |
How the index is built
Each task carries a cited per-use energy figure: a short chat 0.30 Wh, a measured Gemini prompt 0.24 Wh, a GPT-4o query 0.42 Wh, a long document 2.5 Wh, an image 2.907 Wh, and a reasoning prompt 7.6 Wh. Energy is converted to water by multiplying kWh by a 1.8 L/kWh water-use-effectiveness factor (on-site, scope-1), and to carbon by multiplying by the world-average lifecycle grid intensity of 471 gCO₂/kWh. The personas apply a fixed monthly task mix, multiply by twelve, and express the result as phone charges, showers and kilometres driven using cited everyday anchors.
Notes & limitations
These are estimates. Per-task energy varies widely by model, prompt length, hardware and data centre, and measurement boundaries differ (full-system energy is roughly 2.4× GPU-only energy). Closed frontier models are not directly measured. Water shown is on-site cooling only and excludes the larger off-site (scope-2) water used to generate the electricity. Carbon depends on the local grid, which is far cleaner in some regions than the world average used here. Use the figures as orders of magnitude for comparison, not precise accounting.
Sources
- Epoch AI — How much energy does ChatGPT use? (0.3 Wh chat, 2.5 Wh long input)
- Hugging Face / Luccioni — AI Energy Score v2 (reasoning ≈ 7.6 Wh/query)
- Luccioni, Jernite & Strubell — Power Hungry Processing (image ≈ 2.907 Wh)
- Google — Measuring the environmental impact of AI inference (Gemini 0.24 Wh)
- OWID — Carbon intensity of electricity (lifecycle, 2024)