Berget AI releases open-source MIT-licensed method for calculating CO2 emissions of each individual AI request, developed with SEI, WattTime and Klimatkollen

Berget AI, a Swedish AI hosting provider, published an MIT-licensed method and an npm package (@bergetai/co2-calculator 1.1.1) for estimating CO2 emissions per AI request. It follows the Green Software Foundation's SCI for AI spec. A public calculator is at co2.berget.ai, and customers can see per-request figures in the Berget console. Berget's own example puts a typical Kimi K3 request at about 0.4 g CO2 in Sweden, 1.6 g in Ireland, 2.5 g in Germany and 3.2 g in Texas on identical hardware, which is the 8.5x gap in the headline. The figures include hardware manufacturing but exclude training. They are location-based averages, not marginal estimates, which are promised next month.

Buyers, especially of open-weight models where they choose where requests run, get an open and checkable way to compare regions and, in principle, providers. The clearest takeaway is that grid location can matter more than almost any other factor for inference emissions. Sustainability and compliance teams could use the code as a starting point for reporting. Impact is limited for now because this is one small provider's method, and it only matters beyond Berget if other providers adopt it or regulators reference it.

The 8.5x headline is Berget's own number, and it flatters Berget's Swedish data centres. It is a location-based average, not a measure of the emissions a single extra request actually causes. Berget also corrected earlier embargoed figures that claimed ten-fold, which shows how sensitive the result is to method. Training emissions are left out, so this is not a full footprint, and no independent audit of the figures has been reported.

1. 01Whether the promised marginal emissions estimates arrive next month, and how far they move the Sweden vs Texas gap from 8.5x.
2. 02Whether any larger provider, such as Vercel AI Gateway or a major cloud, publishes per-request emissions using the same SCI for AI method or the Berget package.
3. 03Independent review of the GitHub repo and methodology, for example by the Green Software Foundation or academic researchers, and early npm download and fork numbers.