The Pulse
Berget AI’s 8.5× Test Shows Where AI Runs Matters
Berget AI says a typical Kimi K3 request produces an estimated 0.4 grams of CO₂ in Sweden and 3.2 grams in Texas on identical hardware. Its request-level estimate includes hardware manufacturing emissions but excludes model training and the
AI.info Team · Written with AI · fact-checked ·
On October 21, Berget AI released an open-source calculator that estimates the carbon emissions of individual AI requests, then used it to compare the same Kimi K3 workload in Sweden and Texas. The company’s estimate: 0.38 grams of CO₂e per request in Sweden and 3.2 grams in Texas, an 8.5-fold difference on identical hardware. The announcement, developed with Stockholm Environment Institute, WattTime and Klimatkollen, also corrects figures in an earlier version of the release.
One Kimi K3 request, four regional estimates
Berget’s example uses a B300 node in a shared deployment and a request with 5,000 input tokens and 387 output tokens, weighted to reflect a typical day’s traffic. The company estimates roughly 4 watt-hours of energy in both locations; the carbon totals differ because the electricity supply and data-center conditions differ. Its rounded figures are 0.4 grams in Sweden, 1.6 in Ireland, 2.5 in Germany and 3.2 in Texas.
The grid-intensity figures used in the release are about 8 grams of CO₂e per kilowatt-hour for Sweden, 224 for Ireland, 337 for Germany and 333 for Texas. Berget says its request-level estimates also account for hardware manufacturing, which makes the contrast less extreme than comparing grid intensity alone: manufacturing accounts for 89% of the Sweden estimate, while electricity accounts for 90% of the Texas estimate. The Texas request also carries an estimated six milliliters of cooling water.
The release revises its own headline result
Berget’s October 21 correction says an earlier version distributed under embargo put the Ireland, Germany and Texas estimates at 1.2, 2.1 and 3.9 grams, respectively, and described Texas as ten times Sweden. A review of the grid-intensity inputs led the company to revise those figures to 1.6, 2.5 and 3.2 grams, bringing the Texas-to-Sweden comparison to 8.5 times. The Sweden estimate did not change.
That revision matters because the comparison depends on inputs, not just a meter reading attached to each API call. Berget says the estimate varies with the model, time of day and whether part of a prompt is cached. Its published result is a location-based estimate using the average electricity mix for the region where the servers run, not a direct measurement of the emissions caused by one specific request.
Open models make location a buyer’s choice
The company argues that per-request figures could help buyers compare providers and choose where to run open-weight models. Such models can be served by different data-center operators, unlike models tied to a single provider’s infrastructure. Vercel’s analysis of its AI Gateway routing data found that open-weight models accounted for 29% of tokens in June 2026, up from 11% in April, giving Berget’s location-based comparison relevance beyond its own service.
“With this method our customers can follow their emissions request by request and compare providers on data instead of marketing.”
Andreas Lundmark, CEO of Berget AI
The tool is available under the MIT license and follows the Green Software Foundation’s SCI for AI specification, according to Berget. The company says it is the first inference provider, to its knowledge, to calculate and publish emissions for every request on its own infrastructure.
Training is excluded; chip manufacturing’s water and energy are not yet counted
The calculator includes hardware manufacturing emissions but excludes model training from the published comparison. Berget also says the current version does not include the water and energy footprint of chip manufacturing. Those omissions limit what the 0.38-gram and 3.2-gram figures represent: they compare estimated emissions associated with serving a request, not the full lifecycle footprint of the model.
The release uses average regional grid emissions, an established accounting approach, while noting that marginal emissions estimates can better capture how additional electricity demand affects generation. Berget said it planned to add marginal estimates to the calculator in November 2026, and warned that those figures could significantly alter Sweden’s result. Until then, its comparison offers a consistent location-based estimate, with the company’s own correction and scope limits attached.