Carbon footprint
How much does our AI use emit, and where should it run?
From requests to kilowatt-hours and kilograms of CO₂e, with the lever that changes the answer by an order of magnitude: hosting region.
- ≈ 2 min
- Free calculation · AI commentary on request
- You leave with: kWh and kg CO₂e, by hosting region
The workload and where it runs
The same workload emits 6.7 times more outside France than in France. For ordinary use, the hosting choice weighs more than the model choice.
The per-token figures are uncertain by at least a factor of two; the ranges say so. The grid intensity is solid — and it dominates.
The footprint
The figures above are computed and free. If you also want them read back to you — what they mean, what is specific to your sector, where to start — that is produced on request.
Region by region
| Line | Estimate |
|---|---|
| Recurring | |
Electricity per month 115,520k tokens × 0.3 Wh × PUE 1.2. | €42 €13 — €139 |
Emissions per month — FR grid 55 gCO2e/kWh. France: nuclear and hydro dominated. Among the lowest-carbon grids in the world — a genuine argument for EU-hosted inference. | €2 €0 — €13 |
Emissions per month — EU grid 250 gCO2e/kWh. EU-27 average. | €10 €2 — €44 |
Emissions per month — US grid 370 gCO2e/kWh. United States average; varies widely by state. | €15 €4 — €62 |
Emissions per month — WORLD grid 480 gCO2e/kWh. World average. | €20 €5 — €78 |
The assumptions
Every figure above rests on these. The ones marked “assumption” have no external source: they are modelling choices, labelled so they can be argued with rather than mistaken for measurements.
- Energy per 1,000 tokens 0.1–0.8 WhAssumption
- Mid-sized models. The central figure is in line with independent measurements of open-weight models of this class. Uncertain by at least a factor of two: providers do not publish this figure.
- PUE × 1.2Published
- Power usage effectiveness. Hyperscale sites report ~1.1; the fleet average is nearer 1.5.
- Grid intensity 55 gCO2e/kWhPublished
- France: nuclear and hydro dominated. Among the lowest-carbon grids in the world — a genuine argument for EU-hosted inference.
- Scope inference onlyAssumption
- Model training, hardware manufacture and water cooling are not counted. For ordinary use, inference is the part your hosting decision influences.
What the calculation does, what the AI does
Every figure on this page comes from a deterministic algorithm running in your browser: the same input always gives the same result, and nothing is sent to produce it. AI runs only afterwards, to write the commentary from those figures. It produces none of them — the format it fills has no numeric fields — and any sentence quoting a value the calculation did not produce is removed before display.
Where to go next
See all 19 tools →- Roadmap
In what order, and over how long?
≈ 3 min · Quarterly phases and what each one delivers - Training sizer
How many sessions, over how long, for what budget?
≈ 3 min · Session count, real duration and the full budget - Freed capacity
The hours saved — how many people is that, and what does it become?
≈ 2 min · Hours saved as FTEs, and the three ways to read them - Cost calculator
What does our AI transformation actually cost, and when does it break even?
≈ 3 min · A line-by-line budget, the return and the break-even month - Pitfall radar
What is going to go wrong in our programme, and what will it cost?
≈ 2 min · Detected pitfalls, their weighted cost and the fix for each - AI usage policy
Our teams already use AI: what rules do we give them, in writing, this week?
≈ 4 min · A complete AI usage policy, ready to sign