How much does grid carbon intensity change the emissions of an AI prompt in Germany versus France versus the US?
For analysts comparing where an AI prompt is served. Every figure is an estimate computed from the open SOMA AI Emission Factors. Factor version soma-ai-ef-2026.11-dataset-v3.1.0. Factors last updated . Location-based grid averages, not renewable contracts. Estimates from published benchmarks.
Page updated
How much does grid carbon intensity change the emissions of an AI prompt in Germany versus France versus the US?
Enough to change the electricity part of the footprint by a factor of about 6.8 between Germany and France, and 7.4 between the US average and France. For one mid-size model exchange of 928 input and 282 output tokens, the SOMA AI estimate is 0.011 g CO₂e on the French grid, 0.041 g on the German grid and 0.044 g on the US average. The total moves less than the electricity, a factor of 3.8 for Germany and 4.1 for the US, because hardware and training emissions do not depend on the grid.
What do the numbers look like side by side?
| Place | Grid basis | Grid g CO₂e per kWh | Electricity part, g | Total, g | Ratio to France |
|---|---|---|---|---|---|
| Germany | Country grid | 364 | 0.036 | 0.041 | 3.8 |
| France | Country grid | 53.3 | 0.0053 | 0.011 | 1.0 |
| United States, average | Country grid | 393 | 0.039 | 0.044 | 4.1 |
Grid intensities are annual location-based averages, mostly from Ember yearly data.
What changes if the prompt runs in a cloud region instead of a country?
| Region | Grid basis | Grid g CO₂e per kWh | Electricity part, g | Total, g | Ratio to France |
|---|---|---|---|---|---|
| EU Central - Germany | Cloud region | 364 | 0.036 | 0.041 | 3.8 |
| EU West - France | Cloud region | 53.3 | 0.0053 | 0.011 | 1.0 |
| US East - N. Virginia | Cloud region on a US subregion grid | 304 | 0.03 | 0.036 | 3.3 |
| EU North - Sweden | Cloud region | 38.4 | 0.0038 | 0.0094 | 0.9 |
The country that hosts the data centre decides the grid, not the country of the user. An employee in Madrid whose prompt is served from Virginia carries the Virginia figure. Providers rarely say where a consumer chat is served, so state the region you assumed. A US cloud region sits on a sub-national grid, which can differ from the national average in the first table.
Why is the total ratio smaller than the grid ratio?
The estimate has four terms. Electricity scales with the grid. Embodied hardware, data-centre construction and amortised training do not. On the French grid those fixed terms are about 0.0056 g of the 0.011 g total, so a low-carbon grid cannot drive the total towards zero. On a high-carbon grid, electricity dominates and the grid choice matters most.
Is this location-based or market-based?
Location-based. SAEF uses the average carbon intensity of the grid where the electricity is generated and consumed. A market-based figure reflects the supplier's contracts, such as renewable certificates or power purchase agreements, and can be much lower than the local grid. Google's published 0.03 g CO₂e for a median Gemini prompt is market-based, so it is not comparable with a location-based number. Which basis to report depends on the standard you follow, and the GHG Protocol asks for both in Scope 2. For the estimate method, see the methodology page.
Keep reading
- Where does an AI query emit the least? The carbon footprint of AI by country. The same exchange on 210 national grids
- Why is Google's 0.24 Wh per Gemini prompt figure so different from academic estimates? Location-based against market-based, in Google's figure
- What is the carbon footprint of one ChatGPT prompt? Grams per prompt and why the region decides