The same AI request emits 8.5 times more CO2 in Texas than in Sweden
Berget AI released an open-source method on 21 October 2026 for calculating the carbon emissions of every individual AI request, developed with researchers from Stockholm Environment Institute, WattTime and Klimatkollen.
A typical request to the open model Kimi K3 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 on identical hardware — 8.5 times more in Texas than in Sweden. The method is free for anyone to use, and Berget AI wants more AI providers to report emissions per request so buyers can compare providers on real figures instead of marketing.
The method follows the Green Software Foundation's SCI for AI specification and is available under the MIT licence on GitHub. It powers a calculator at co2.berget.ai that applies the method to any model and region. Berget AI reports location-based figures using the average grid mix where the servers run and includes the emissions from manufacturing the hardware. Open-weight models, which can run in any data centre, handle a growing share of AI traffic: 29% of tokens on Vercel's AI Gateway in June 2026, up from 11% in April. For those models, where a request runs is the buyer's choice, provided the emissions are visible.
The key facts of the case are these. The method calculates each request separately instead of applying an average, because the result depends on the model, the time of day and whether part of the prompt is cached. Buyers rarely get emission figures from their AI providers, and the public figures that exist are mostly averages estimated from the outside. As far as we know, Berget AI is the first inference provider to calculate emissions for every request on its own infrastructure and publish the method. The partners will keep developing it, for example to cover the water and energy footprint of chip manufacturing, which the current version does not include. Emissions estimates are based on the current average emissions intensity of each regional electricity grid, which has been the norm in greenhouse gas accounting. An emerging approach uses marginal emission rates, which estimate the impact of additional load on both the use of existing power plants and the new generators built to meet increased demand. Marginal estimates will be added to the calculator next month and may significantly change the measured greenhouse gas intensity of data centres in Sweden. Sweden's figure is unchanged.
Sweden's average grid intensity is ~8 g CO₂e/kWh; Texas's ERCOT grid is ~333 g CO₂e/kWh. Berget AI's bet is that visibility will shift demand: "most European buyers would choose differently if they could see the numbers," says CEO Andreas Lundmark.
The measurement system itself introduces a second design tension. The current method uses average grid intensity, which treats every kilowatt-hour in a region as equally carbon-intensive. An emerging alternative, marginal emission rates, asks: what power plant actually ramps up when this data centre draws one more megawatt? The partners will add this next month, and it may significantly change Swedish figures because the marginal generator in a low-carbon grid is not necessarily low-carbon. Marginal rates are more accurate for decision-making but harder to calculate and more volatile; average rates are stable, comparable, and already the norm in greenhouse gas accounting. The choice between them is a trade-off between precision and usability, and the method's credibility depends on making that choice explicit.
A third tension sits in the hardware boundary. The method includes GPU manufacturing at ~1,000 kg CO₂ per unit, with refurbished counted as zero. This is a judgement call: refurbished hardware has already emitted its manufacturing carbon, so the marginal emissions of reusing it are zero. But this creates an incentive to label hardware refurbished even when it is functionally new, or to ignore the upstream supply chain that produces refurbished units. The partners acknowledge that chip manufacturing water and energy are not yet included, so the embodied figure is partial.
The mechanism that produces the outcome, then, is this: a buyer sends a request to an open-weight model running on Berget AI's infrastructure in Sweden. The software measures the request's compute time, multiplies by the measured or estimated power draw of the specific GPU, applies the regional grid intensity, adds the amortised embodied hardware, and returns a gram figure. The same request sent to the same model on the same GPU in Texas gets a different number because the grid intensity multiplier is different. The buyer sees both numbers and can choose. If the buyer uses a closed provider, no such number is available, so no such choice is possible.
The incentive structure rewards Berget AI for publishing because its infrastructure sits on a low-carbon grid. A provider on a high-carbon grid would need to either publish worse numbers, not publish, or invest in offsets or power purchase agreements that the method may or may not credit. The method's use of location-based accounting, not market-based accounting, means that renewable energy certificates bought by a Texas provider do not reduce the reported figure. This is a deliberate choice that follows the Green Software Foundation specification, and it privileges physical infrastructure over financial instruments.
The sharpest unresolved fact is what happens when marginal rates are added next month, and whether the Swedish advantage shrinks or disappears if the marginal generator in Sweden is fossil. The method is designed to evolve, but its authority depends on each version being legible. A buyer comparing providers on per-request grams needs to know which version of the method each provider used, and whether the difference between 0.4 grams and 3.2 grams is a durable physical fact or a temporary accounting boundary.