Swedish infrastructure provider Berget AI has released an open-source method for estimating the carbon emissions of individual AI inference requests, giving customers a way to track results in the company’s console rather than relying on a broad industry average.
The company developed the approach with researchers and specialists from the Stockholm Environment Institute, WattTime and Klimatkollen. It combines measurements from Berget AI’s infrastructure with information about electricity-grid emissions and the manufacturing footprint of computing hardware. The method follows the Green Software Foundation’s SCI for AI specification and has been published under an MIT licence.
Berget AI used the calculator to compare a typical request to the open Kimi K3 model on identical hardware in several regions. Its location-based estimates put the request at about 0.4 grams of carbon dioxide equivalent in Sweden, 1.6 grams in Ireland, 2.5 grams in Germany and 3.2 grams in Texas. On those figures, the Texas result is roughly 8.5 times the Swedish result.
Those numbers are estimates issued by the company, not an independent audit of every deployment. Berget AI also corrected figures circulated in an earlier embargoed release after reviewing the grid-intensity inputs. The company said request-level results can vary with the model, time of day and prompt caching, which is why its method calculates each request separately.
The current version uses average emissions for the regional grid where servers operate. Berget AI plans to add marginal-emissions estimates, which attempt to measure the effect of adding electricity demand, and warned that this change could materially alter comparisons. The method includes hardware manufacturing but does not yet cover the water and energy used to manufacture chips.
The release is intended to make provider comparisons more transparent as open-weight models can be placed in different data centres. Berget AI argues that customers choosing where inference runs should be able to see the carbon implications alongside cost and performance.
The company said customers can now monitor their own emissions per request, while other providers can adopt or inspect the published methodology. Its public calculator applies the method across models and regions. The project also builds on external estimation tools, including EcoLogits and the Sustainable AI Group’s CLEER dashboard, while using operational data available to the infrastructure provider.
For buyers, the main practical conclusion is that geography can change the estimated footprint even when model and hardware remain constant. The figures should still be read with their accounting assumptions in view, particularly the use of average grid intensity and the method’s incomplete coverage of chip production.
