New open-source methodology exhibits the identical AI request emits 8.5 instances extra CO₂ in Texas than in Sweden
Estimates primarily based on the typical emissions depth of every regional electrical energy grid*
STOCKHOLM, 21 October 2026. Swedish AI supplier Berget AI as we speak releases an open-source methodology for calculating the carbon emissions of each particular person AI request, developed with researchers from Stockholm Environment Institute (SEI) and the tech non-profits WattTime and Klimatkollen. Berget AI’s clients can comply with their very own emissions within the firm’s console as they use the service. By making the tactic free for anybody to make use of, Berget AI desires extra AI suppliers to report emissions per request, in order that patrons can examine suppliers on actual figures and select lower-carbon choices.
“Most firms as we speak haven’t any method of understanding what their use of AI emits. With this methodology our clients can comply with their emissions request by request and examine suppliers on information as a substitute of selling. We know most European patrons would select in another way if they may see the numbers,” says Andreas Lundmark, CEO of Berget AI.
A typical request to the open mannequin Kimi K3 emits about 0.4 grams of CO₂ in Sweden, 1.6 grams in Ireland, 2.5 grams in Germany and three.2 grams in Texas, on an identical {hardware} — 8.5 instances extra in Texas than in Sweden. The end result additionally depends upon the mannequin, the time of day and whether or not a part of the immediate is cached, so the tactic calculates every request individually as a substitute of making use of a mean. Berget AI experiences location-based figures, utilizing the typical grid combine the place the servers run, and consists of the emissions from manufacturing the {hardware}.
Buyers hardly ever get emission figures from their AI suppliers, and the general public figures that exist are principally averages estimated from the skin. Open-weight fashions, which might run in any information centre, deal with a rising share of AI visitors: 29% of tokens on Vercel’s AI Gateway in June 2026, up from 11% in April. For these fashions, the place a request runs is the client’s selection, offered the emissions are seen.
“Most of the alarming figures about AI’s footprint come from multiplying a mean by a guess on the quantity, and no person can test them. The solely truthful comparability is to calculate every request and add them up. We have printed our judgement calls so anybody can test and use them,” says Christian Landgren, CTO and co-founder of Berget AI.
“The International Energy Agency tasks data-centre electrical energy consumption to greater than double to about 945 TWh by 2030, with AI because the principal driver. Any promise of effectivity beneficial properties is prone to be outweighed by larger demand, negating any emission reductions from renewables. Methodologies like this provide a clear foundation for knowledgeable choices. Their power is that they clearly show how the situation of the grid and significant utilities matter in the case of how sustainable AI might be,” says Somya Joshi, Research Director at SEI.
About the tactic
The companions mapped and measured the environmental affect of Berget AI’s utilities, together with vitality, water and {hardware}, in addition to the carbon footprint of Klimatkollen’s Garbo AI, which runs on it. They mixed the information with present analysis on translating it into carbon emissions and broader environmental affect.
The methodology follows the Green Software Foundation’s SCI for AI specification and is obtainable below the MIT licence on GitHub. The calculator at co2.berget.ai applies it to any mannequin and area. It builds on open-source estimators comparable to EcoLogits and the Sustainable AI Group’s CLEER dashboard, which estimate emissions from outdoors the supplier. As far as we all know, Berget AI is the primary inference supplier to calculate emissions for each request by itself utilities and publish the tactic. The companions will hold creating it, for instance to cowl the water and vitality footprint of chip manufacturing, which the present model doesn’t embody.
* Emissions estimates are primarily based on the present common emissions depth of every regional electrical energy grid, which has been the norm in greenhouse fuel accounting. An rising method makes use of marginal emission charges, which estimate the affect of further load on each using current energy crops and the brand new mills constructed to fulfill elevated demand. Marginal estimates will likely be added to the calculator subsequent month and should considerably change the measured greenhouse fuel depth of knowledge centres in Sweden.
Correction, 21 October 2026: an earlier model of this launch, distributed below embargo, acknowledged 1.2 g (Ireland), 2.1 g (Germany) and three.9 g (Texas) per request and a ten-fold distinction between Texas and Sweden. A supply evaluation of the grid intensities printed with this launch corrected them to Ireland ~224, Germany ~337 and Texas ~333 g CO₂e/okayWh; the corrected per-request figures seem above and the Texas–Sweden distinction is 8.5×. Sweden’s determine is unchanged.
Facts
| Grid depth | Sweden ~8 g CO₂e/okayWh (IEA 2024) · Ireland ~224 (SEAI 2024) · Germany ~337 (Ember 2024) · Texas (ERCOT) ~333 (EPA eGRID 2023) · world common ~470 (Ember 2024) |
| Embodied {hardware} | ~1,000 kg CO₂ per information centre GPU; refurbished counted as zero |
| Idle energy | An idle B300 nonetheless attracts 236 W (measured over seven days in manufacturing) |
| Cooling | Nordic free air: PUE ~1.15, no water. Hot climates: towards 1.8, as much as 2 L water/okayWh |
| Scope | Training excluded: self-reported figures differ ±50% |
| Open-model visitors | 29% of tokens on Vercel AI Gateway, June 2026 (about 11% in April); ~69% of named-model tokens on OpenRouter |
| Kimi K3, one typical request | Sweden 0.38 g CO₂e (89% from {hardware} manufacturing). Texas 3.2 g (90% from electrical energy), plus about 6 mL of cooling water. Energy about 4 Wh in each. Calculated with @bergetai/co2-calculator 1.1.1: B300 node, shared deployment, production-typical request of 5,000 enter and 387 output tokens, weighted by a typical day’s visitors. |
About Berget AI
Berget AI champions AI sovereignty to unlock innovation and progress in Europe. Founded in Sweden in 2024, Berget AI affords inference and AI utilities for personal and public sector organisations. Since 2025 it has reported the CO₂ emissions of each API request.
About SEI, WattTime and Klimatkollen
Partners: Stockholm Environment Institute (data associate), WattTime (scientific associate), Klimatkollen (associate).
Stockholm Environment Institute is an global non-profit analysis institute that tackles local weather, surroundings and sustainable improvement challenges. To study extra, go to www.sei.org.
Klimatkollen.se is a Swedish citizen platform that presents local weather information about municipalities and firms. The goal is to contribute to extra engaged residents, a extra fact-based local weather debate, and measures that scale back emissions in keeping with the Paris Agreement. Klimatkollen is run by the non-profit affiliation Klimatbyrån.
WattTime is an environmental tech nonprofit that empowers all individuals, firms, policymakers, and international locations to slash emissions and select cleaner vitality. Founded by UC Berkeley researchers, we analysis, develop, and share data-driven instruments and insurance policies that assist anybody who’s to scale back extra emissions, sooner. www.watttime.org
