The CO₂ of your AI use

Your footprint is worked out with one formula: tokens × weight × energy per token × data-centre overhead × grid × hardware. Weight says how much energy reading a token takes compared with writing one. The result is in CO₂e: CO₂ and the other greenhouse gases, counted as CO₂.

Your tokens are counted. The other numbers are not known exactly. For each of them, the list of assumptions below gives a low, a typical and a high value.

• Low and high are the lowest and the highest value we take seriously. The page never uses a value outside them.
• Typical is the middle. We treat a lower value and a higher value as equally likely.

The page works out your footprint 10,000 times. Each time it uses a different combination of values, taken from a fixed pattern that spreads the combinations evenly. For every assumption you have not set yourself, half of the 10,000 runs use a value below typical and half a value above it. Positions near the middle of each slider are used most often. Positions near the two ends are used least often. That shape is our choice. No data can settle it.

The energy values for the model sizes move together: a run that is high for one size is high for all. Every other assumption varies on its own.

The page sorts the 10,000 results and cuts the list into 100 groups of 100. Each dot in the chart is the middle of one group. The middle estimate lies between group 50 and group 51, so half of the results are below it and half above. The likely range leaves out the 5 lowest and the 5 highest groups, so 90 of 100 results land inside it.

The extreme range is the formula with every assumption you have not set at its low value, and again with each at its high value. No result can fall outside it.

Nothing here is random. The same data always gives the same result.

The range is only as good as that list. If a real value lies below our low or above our high, your real footprint can lie outside the range too.

“Saves roughly A to B” under a tip means this. If the days counted here had been the same except for this one change, your footprint would have been A to B lower in 90 of 100 possible outcomes. The tips are ordered by the middle of each saving, which can differ from the order of the ranges.

The car figure counts only what leaves the exhaust: about 160 g of CO₂ per km for a new petrol car in Europe in real driving.

Giving money does not undo emissions. We follow the idea of a contribution claim from a 2024 paper by the University of Oxford: what you give is a contribution to climate work, and your estimate stays what it is.

A cost range under “Three ways to contribute” means this. In at least 90 of 100 possible outcomes, funding that many kilograms with this provider would cost between the two figures.