A typical AI request run on identical hardware can produce 8.5 times more carbon dioxide in Texas than in Sweden, according to figures released Wednesday by Berget AI, a Swedish AI infrastructure provider. The company is also publishing a free, open-source method for calculating the emissions of every individual request, and it wants other providers to do the same.
The method was developed with the Stockholm Environment Institute, the nonprofit WattTime and Klimatkollen, a Swedish citizen platform that tracks climate data. Berget AI's customers can already follow their own emissions in the company's console as they use the service. Berget AI says that by making the method available to anyone, more providers will report emissions per request, so buyers can compare services on real numbers.
The company's test case involved Kimi K3, an open-weight model. A typical request emits about 0.4 grams of CO₂ in Sweden, 1.6 grams in Ireland, 2.5 grams in Germany and 3.2 grams in Texas. An earlier embargoed version of the release gave different figures for Ireland, Germany and Texas, and the company revised them after a source review of the grid data. The Texas-to-Sweden gap of 8.5 times is the same in the corrected version.
The difference comes mostly from the electricity grids that power the servers. Sweden's grid intensity is about 8 grams of CO₂ equivalent per kilowatt-hour, according to the International Energy Agency's 2024 figures. Texas, based on 2023 data from the U.S. Environmental Protection Agency, is about 333. The method also accounts for other factors. Results depend on the model, the time of day and whether part of a prompt is cached, so the calculator works out each request separately rather than applying a flat average.
Berget AI reports location-based figures, using the average grid mix where its servers run. It also includes emissions from manufacturing the hardware. For a typical request in Sweden, the company says about 89 percent of the emissions come from hardware manufacturing. In Texas, about 90 percent come from electricity.
Few AI buyers get emissions data from their providers. The public figures that do exist are mostly estimates made from the outside, and Berget AI argues that many of the widely cited numbers rely on multiplying an average by a guess about usage volume. Christian Landgren, the company's chief technology officer and co-founder, said that approach cannot be checked.
"Most of the alarming figures about AI's footprint come from multiplying an average by a guess at the volume, and nobody can check them," Landgren said. "The only fair comparison is to calculate each request and add them up. We have published our judgment calls so anyone can check and use them."
Open-weight models, which can run in any data center, are a growing share of AI traffic. They accounted for 29 percent of tokens on Vercel's AI Gateway in June 2026, up from 11 percent in April. For those models, the location where a request runs is a choice the buyer can make, as long as the emissions are visible.
Andreas Lundmark, Berget AI's chief executive, said most European buyers would choose differently if they could see the numbers. "Most companies today have no way of knowing what their use of AI emits," Lundmark said. "With this method our customers can follow their emissions request by request and compare providers on data instead of marketing."
The stakes are rising. The International Energy Agency projects data center electricity consumption will more than double to about 945 terawatt-hours by 2030, with AI as the main driver. Somya Joshi, research director at the Stockholm Environment Institute, said efficiency gains are likely to be outweighed by growing demand. She said methods like this give buyers a transparent basis for decisions and show how grid location and infrastructure affect how sustainable AI can be.
The method follows the Green Software Foundation's SCI for AI specification and is available under the MIT license on GitHub. A calculator at co2.berget.ai applies it to any model and region. The tool builds on open-source estimators, including EcoLogits and the Sustainable AI Group's CLEER dashboard, which estimate emissions from outside the provider. Berget AI says it is the first inference provider to calculate emissions for every request on its own infrastructure and publish the method.
The method has limits. It does not yet cover the water and energy footprint of chip manufacturing, and it excludes training. The partners plan to add those areas over time. Marginal emission rates, which estimate the impact of additional electricity demand on existing and new power plants, will be added to the calculator next month. Berget AI says those figures could significantly change the measured emissions of data centers in Sweden.
Questo articolo è stato scritto con l'assistenza dell'IA.
News Factory APP - notizie agentiche per potenziare il tuo SEO e AEO.