Researchers from Berget AI, the Stockholm Environment Institute, WattTime, and Klimatkollen have released an open source method for estimating the CO2 emissions of individual AI requests. The method accounts for the carbon intensity of the electricity grid where the inference runs. A typical request to the Kimi K3 model comes to about 0.4 grams of CO2 in Sweden, 1.6 grams in Ireland, 2.5 grams in Germany, and 3.2 grams in Texas on identical hardware. The project includes a web based calculator and a GitHub repository so companies can compare AI suppliers and run their own estimates. The main takeaway is that location can matter more than hardware efficiency. AI buyers that care about emissions need to ask where their inference runs, not just which model they use.
