Where does an AI query emit the least? The carbon footprint of AI by country

Among the AWS, Azure and Google Cloud regions in the dataset, the lowest is on the Quebec (Canada) grid, where the dataset maps AWS ca-central-1. The same mid-size AI exchange of 1,210 tokens comes to an estimated 0.008 g CO₂e there, against 0.095 g in Bahrain, about 12 times more for identical work. Inside the EU, Sweden is lowest at 0.009 g and Poland highest at 0.070 g, about 7.4 times apart. At the world-average grid the figure is 0.053 g. These are estimates from the SOMA AI (somaai.earth) dataset, which covers the grids of 209 countries and every AWS, Azure and Google Cloud region it maps. A query is served where the data centre is, not where the user sits.

Because electricity is most of the footprint on a fossil-heavy grid and a small part on a clean one. On the world-average grid, electricity is 90% of the estimate for one exchange. The rest, the server hardware, the data-centre building and the amortised training of the model, is about 0.006 g wherever the query runs, which is why even the cleanest grid does not reach zero. Model size multiplies everything, as the class comparison below shows.

Sweden, at 0.009 g per exchange, and the highest in the EU is Poland at 0.070 g, with the EU-27 average at 0.029 g. Spain ranks 12 of 27, at 0.022 g, and the AWS Spain region (eu-south-2) has its own page.

Of the 143 AWS, Azure and Google Cloud regions in the dataset, the cleanest grid is Quebec (Canada), at 0.008 g per exchange, and the most carbon-intensive is Bahrain, at 0.095 g, about 12 times more for identical work. United States regions are matched to their eGRID subregion rather than the national average, because the grids inside the country differ widely.

In this dataset version, Azure canadaeast and Google Cloud northamerica-northeast1, both in Quebec, are mapped to the Canada-wide average rather than the Quebec grid. They will be corrected in the next version.

A frontier model on a clean grid can emit more than a small model on a dirty one, so a company choosing where to cut should look at both the model and the region.

Find out the region each AI provider serves your usage from, and use that region's factor, not your office's country. On cloud platforms the region is in the usage export. Where you can choose, the region is the largest lever after the model class, with no change to the work. Report the location-based figure, and a market-based one beside it only if the provider documents its renewable contracts for that region. The method for the inventory line is in the Scope 3 guide, and the free calculator applies the region for you.

Across the 209 ranked countries the range is wider, from 0.008 g in Nepal to 0.134 g in Turkmenistan. The Central African Republic row has a grid intensity of 0 g per kWh in the 2023 source data, which we treat as a data gap, so it is not ranked and is left out of the range and the median. The median country is Malta, at 0.054 g. The full grid table, with sources per row, is in the dataset on Zenodo under CC BY 4.0, free to reuse with attribution to SOMA AI Emission Factors.

• Grid intensities are annual, location-based averages, for 2023, mostly Ember yearly data on a lifecycle basis, with US subregions from EPA eGRID 2023 scaled to the same lifecycle basis and, for Quebec, Canada Energy Regulator 2023 generation shares on the Ember lifecycle basis.
• A country's average is not the mix of a specific data centre, which may sit on a cleaner or dirtier part of the grid or buy renewable electricity.
• All figures use the class factors for H100 servers, with hardware, construction and amortised training included. Water is not part of this comparison.
• The method paper is under peer review, and corrections go into the next dataset version. The equations are on the methodology page.

Figures per query are one plain exchange, before the Tier 2b reasoning uplift. Electricity term. GPU energy per token from the ML.ENERGY Leaderboard v3, a facility overhead of 1.20, and grid carbon intensity from EPA eGRID 2023 and Ember 2023. Embodied hardware, data-centre construction and amortised training terms are documented on the methodology page. Water combines on-site cooling water (Li et al. 2025) with the water embedded in grid electricity (Reig et al. 2020). The modelled range of the models inside a class, around the class value, is −41% to +48% for class A, −39% to +24% for class B and −67% to +55% for class C.

Dataset: AI Inference Emission and Resource Factors for Corporate GHG Inventories (SOMA AI Emission Factors (SAEF)), Llopis-Garcia, G., SOMA AI, 19 October 2026, CC BY 4.0, concept DOI 10.5281/zenodo.20443585, which resolves to the latest version (this version 10.5281/zenodo.23371447, earlier version 2 values at 10.5281/zenodo.22767475). Method: Accounting for Purchased AI Inference in Corporate GHG Inventories: A Token-Based Method, arXiv:2606.10660, under peer review.

Partners can fetch any row from the AI factors API, and the whole table is a CSV download.

• SOMA AI Emission Factors by cloud region. Per-region pages with carbon and water
• What is the carbon footprint of one ChatGPT prompt? Published per-prompt figures and their basis
• Is AI bad for the environment? Per query against the total