Berget AI publishes a method to measure CO2 per AI request

The same AI request can emit 8.5 times more carbon in Texas than in Sweden. Swedish inference provider Berget AI measured the gap on identical hardware. The company released an open-source method that calculates emissions for every AI request. It was built with researchers from the Stockholm Environment Institute, WattTime and Klimatkollen.

What Berget AI measured

Berget AI said a typical request to the open model Kimi K3 emits about 0.4 grams of CO2 in Sweden. The same call emits 1.6 grams in Ireland, 2.5 in Germany and 3.2 in Texas. The company said each region ran the request on the same hardware. The result changes with the model, the time of day and whether the prompt is cached.

The release lists the grid intensities behind the numbers. Sweden runs at about 8 grams of CO2 per kilowatt-hour, Ireland at 224, Germany at 337 and Texas at 333. The global average is about 470. For Kimi K3 in Sweden, 89 percent of the per-request figure is from making the hardware. In Texas, 90 percent is from the electricity itself. A Texas call also uses about 6 millilitres of cooling water.

The company corrected an embargoed version of the release. An earlier draft gave 1.2 grams for Ireland, 2.1 for Germany and 3.9 for Texas, and a ten-fold gap. A source review of the grid figures brought those down and set the gap at 8.5 times.

How the method works

The method follows the Green Software Foundation's SCI for AI specification. It is on GitHub under the MIT licence. A calculator at co2.berget.ai runs the method on any model and region. The company said it builds on open-source estimators such as EcoLogits and the Sustainable AI Group's CLEER dashboard. Those two estimate emissions from outside the provider.

Berget AI said it is the first inference provider to publish the per-request figure for its own infrastructure. Customers see their own emissions in the Berget AI console as they use the service. The company excludes training from the scope and said the self-reported figures for that vary by about 50 percent.

What the number does not say

Berget AI reports location-based figures. That uses the average mix of the grid where the servers run, which is the common accounting method. The release said the calculator will add marginal emissions next month. Marginal emissions measure the impact of extra load on the plants that respond to it. The company said marginal figures may change the measured intensity of data centres in Sweden by a lot.

The release gives one model's figures from one provider on one date. Anyone can run the method on their own provider. The company said open-weight models carried 29 percent of tokens on Vercel's AI Gateway in June 2026. That share was 11 percent in April. That is the share where a buyer can choose the region at all.

Source

Berget AI press release: 8.5 times more CO2 in Texas than in Sweden.