The same AI request emits 8.5 times more CO2 in Texas than in Sweden

Berget AI, a Swedish AI provider, released an open-source method to calculate CO2 emissions for each individual AI request, developed with SEI, WattTime and Klimatkollen. Using the Kimi K3 model, the release estimates about 0.4 g CO2 per request in Sweden versus 3.2 g in Texas, on identical hardware. The method uses location-based grid averages, includes hardware manufacturing emissions, and follows the Green Software Foundation's SCI for AI specification under an MIT licence. The release also includes a correction to earlier grid intensity figures.

Key takeaways

• A typical Kimi K3 request is estimated at 0.4 g CO2 in Sweden and 3.2 g in Texas.
• The method calculates each request separately rather than applying an average, and is published under an MIT licence.
• Location-based figures use average grid mix where servers run and include hardware manufacturing emissions.
• Marginal emission rates, which may significantly change Swedish data-centre figures, are planned for the calculator next month.

Why it matters

Buyers rarely receive per-request emissions data from AI providers. A transparent, open method could let buyers compare providers and regions on measured figures rather than estimates made from outside.

Read with care

Source claims not independently checked; estimates use average grid intensities, exclude training emissions, and the release itself notes that marginal-rate methods may change results substantially.