What your AI costs the planet

A million output tokens are roughly ten novels' worth of text. Generated by a small open model in France, they cause about 120 grams of CO2e — one kilometer of driving, less than a cup of coffee. Generated by a large closed frontier model in the United States, our estimate puts them at roughly 17,000 grams — about 140 kilometers. More than a hundred times as much, for the same amount of text.

Where the spread comes from

The arithmetic behind our estimates is short: tokens × energy per token × carbon intensity of the electricity where the model runs. The last two factors vary enormously. Energy per token grows with the model's size — more precisely, with the number of parameters doing work on each token. Between the smallest and the largest model in ownAI, that is a factor of about twenty. Then comes the grid: the US electricity mix is four to seven times more carbon-intensive than the French or Swiss one. Multiplied, the two produce the hundredfold spread.

For scale: a single answer stays tiny — a few grams even on the largest model, less than a hundred meters of driving. But tokens add up. Agents working on their own, long documents, a full workday — answers quickly become millions of tokens.

Seeing and choosing

ownAI keeps a running estimate of your emissions. You find it in the app behind the credits icon at the top right, and in your dashboard under "Climate Impact" — together with the receipts for the carbon removal (more on that below).

The most effective decision is the choice of model. European-hosted open models are by far the cleanest per token — and usually the cheapest, too. A small model is enough for many everyday tasks; the large one can be picked deliberately when it is needed. And a local model on your own machine runs on your own electricity — how clean that is, your power contract decides.

What we do

For the usage that runs through ownAI Credits, we estimate the emissions every day and buy the permanent removal of at least that much CO2 from the atmosphere through Ecologi: biochar and enhanced rock weathering. Removal costs about seventeen times as much as avoidance certificates — the difference is that the CO2 is actually taken back out of the air, rather than someone else promising to emit less in the future. This is paid from our margin; it does not affect your prices.

Because no provider discloses how much energy a request actually uses, we estimate from token counts and published energy models. Wherever a number is uncertain, we pick the assumption that raises the estimate: the grid's actual carbon intensity instead of an operator's renewable claims, deliberately high size estimates for models whose makers disclose nothing.

What is still missing

This account is incomplete, and that belongs here. It is an estimate, not a measurement. It covers only the electricity of inference — the manufacturing of chips and servers, the water consumption of data centres, and the training of the models are not included. Requests through your own API keys or local models never touch our proxy and do not appear in this account. And the 17,000 grams from the opening rest on a deliberately high size estimate; the true value is probably lower.

We are working on making the estimate more complete, and we update the methodology as better data becomes available. The full calculation — every constant, every source, every limitation — is public: carbon-methodology.md. Our public removal receipts: Ecologi profile.