A typical request to the open-weight model Kimi K3 emits about 0.4 grams of CO2 when it runs on servers in Sweden. The identical request, on identical hardware, emits 3.2 grams in Texas — 8.5 times more. That gap is the centerpiece of a new open-source method for calculating the carbon emissions of individual AI requests, released on October 21 by the Swedish AI provider Berget AI together with researchers from the Stockholm Environment Institute (SEI) and the nonprofits WattTime and Klimatkollen.
Why Per-Request Carbon Accounting Matters For more context on this story, see our ongoing AI news.
Most companies that use AI have no way of knowing what their usage actually emits, according to Berget AI. Public figures that do exist are usually broad averages estimated from the outside, computed by multiplying an assumed average footprint by a guess at total volume — numbers nobody can independently check.
The new method takes the opposite approach: it calculates emissions request by request, based on the average emissions intensity of the regional electricity grid where the servers actually run, and folds in the embedded carbon of manufacturing the hardware. The result depends on the model used, the time of day, and whether part of the prompt is served from cache, so each request is computed individually rather than via a flat average.
"Most companies today have no way of knowing what their use of AI emits," said Andreas Lundmark, CEO of Berget AI, in the announcement. "With this method our customers can follow their emissions request by request and compare providers on data instead of marketing. We know most European buyers would choose differently if they could see the numbers."
The Grid Is the Variable That Dominates
The headline comparison — 0.4 grams of CO2 per Kimi K3 request in Sweden versus 3.2 grams in Texas — comes from differences in grid mix alone. The same request emits 1.6 grams in Ireland and 2.5 grams in Germany, according to the press release. Because the underlying hardware is identical in each case, the delta is almost entirely a function of where the data center plugs in.
That variable is becoming a procurement decision rather than a fixed constraint. Open-weight models, which can legally run in any data center, now handle a growing share of global AI traffic. Berget AI cites figures from Vercel's AI Gateway showing open-weight models served 29% of tokens in June 2026, up from 11% in April. For those workloads, the buyer can choose the geography — provided emissions are actually visible.
"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," said Christian Landgren, CTO and co-founder of Berget AI. "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."
Built on Existing Standards, Released Under MIT
The method follows the Green Software Foundation's SCI for AI specification and is available under the MIT license on GitHub. Berget AI has also launched a public calculator at co2.berget.ai that applies the method to any model and region, building on prior open-source estimators such as EcoLogits and the Sustainable AI Group's CLEER dashboard. Customers of Berget's own inference service can watch their emissions accumulate in the company console as they query models.
The partners mapped and measured the environmental impact of Berget's infrastructure directly — energy, water, and hardware — along with the carbon footprint of Klimatkollen's Garbo AI, which runs on that infrastructure, then combined the measurements with current research on translating resource use into carbon and broader environmental impact.
Demand Growth Still Outpaces Efficiency
The release arrives amid sharply rising projections for data center power consumption. Somya Joshi, research director at SEI, pointed to International Energy Agency figures projecting data-center electricity consumption will more than double to about 945 TWh by 2030, with AI as the principal driver.
"Any promise of efficiency gains is likely to be outweighed by higher demand, negating any emission reductions from renewables," Joshi said in the announcement. "Methodologies like this offer a transparent basis for informed decisions. Their strength is that they clearly demonstrate how the location of the grid and critical infrastructure matter when it comes to how sustainable AI can be."
What It Means for AI Buyers
For companies running their own open-weight deployments, the practical implication is straightforward: the same model can carry a carbon price that varies by a factor of eight depending on region, and tools now exist to measure that difference before signing a contract. Berget AI says it wants other inference providers to adopt per-request reporting so buyers can compare providers on measured figures rather than marketing claims.
The method's authors acknowledge their figures are location-based estimates using average grid mix rather than hourly marginal emissions, a simplification common to this class of tools. Even so, the transparency of the assumptions — all published for review — is what distinguishes the effort from the opaque averages that have dominated AI energy discourse so far.
Whether other providers follow suit may depend less on the code than on procurement pressure: if enterprise buyers start asking for per-request emissions in RFPs, average-based reporting will become harder to defend.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
