AI carbon footprint in Global average: emissions per token and per query
Factor version soma-ai-ef-2026.10-dataset-v3.0.0. Last updated . Location-based grid averages, not renewable contracts. Estimates from published benchmarks.
Part of the SOMA AI Emission Factors cluster.
How much CO₂ does an AI query emit when served from Global average?
A 1210-token query to a class B model (gpt-4o) served from Global average emits about 0.053 g CO₂e, and a million tokens emit 0.0388 kg CO₂e.
It sits between EU North - Sweden (0.00714 kg per million tokens for class B) at the low end and Asia Pacific - Singapore (0.0401 kg) at the high end. The factor applies when the serving region is unknown.
Which grid supplies the electricity in Global average?
This page uses the World row of the dataset's grid table, 483.2 g CO₂e per kWh in 2023, an aggregate and not a single grid. The source is Ember yearly electricity data, CO2 intensity 2023, 483.19 g/kWh (lifecycle, IPCC AR5). For comparison, the EU-27 aggregate is 236 g per kWh, 2.0 times lower than the global average.
No provider region in the dataset's cloud mapping table resolves to the World grid. The row is the fallback the dataset applies when no region is known or a model is served from the EU without a country, so it has no AWS, Azure or Google Cloud region id of its own.
Which grids have an intensity close to World?
| Difference g | Grid | Code | g CO₂e per kWh | Ratio |
|---|---|---|---|---|
| 0.6 | United Arab Emirates | ARE | 483.8 | 1.001 |
| 1.7 | Macao | MAC | 484.9 | 1.004 |
| -3.2 | Cabo Verde | CPV | 480 | 0.993 |
| 5.8 | Mexico | MEX | 489 | 1.012 |
| 8.0 | Cambodia | KHM | 491.2 | 1.017 |
| -8.5 | Viet Nam | VNM | 474.7 | 0.982 |
| 9.4 | Japan | JPN | 492.6 | 1.019 |
| 10.4 | Malta | MLT | 493.6 | 1.022 |
| -11.1 | MRO West | MROW | 472.1 | 0.977 |
| 11.7 | Turkey | TUR | 494.9 | 1.024 |
| Region | Class B kg | Difference kg | Ratio |
|---|---|---|---|
| EU North - Sweden | 0.00714 | -0.0317 | 0.18 |
| EU West - France | 0.0082 | -0.0306 | 0.21 |
| EU North - Finland | 0.0102 | -0.0286 | 0.26 |
| EU West - Belgium | 0.014 | -0.0248 | 0.36 |
| EU South - Spain | 0.0165 | -0.0223 | 0.42 |
| EU West - UK (London) | 0.0212 | -0.0177 | 0.55 |
| Netherlands | 0.0235 | -0.0153 | 0.61 |
| EU West - Ireland | 0.0245 | -0.0143 | 0.63 |
| US East - N. Virginia | 0.0261 | -0.0128 | 0.67 |
| US West - Oregon | 0.0275 | -0.0114 | 0.71 |
| EU Central - Germany | 0.0303 | -0.0085 | 0.78 |
| US South Central - Texas | 0.0312 | -0.0077 | 0.80 |
| US South - Southeast | 0.0351 | -0.0037 | 0.90 |
| US Central - Iowa | 0.0381 | -0.0008 | 0.98 |
| Asia Pacific - Japan | 0.0395 | 0.0007 | 1.02 |
| Asia Pacific - Singapore | 0.0401 | 0.0013 | 1.03 |
Emission factors for Global average by model class
| Model class | Electricity kg | Hardware kg | Construction kg | Training kg | Total kg CO₂e | g per query | Water L | mL per query |
|---|---|---|---|---|---|---|---|---|
| Class A (small) | 0.01052 | 0.00027 | 0.00014 | 0.00060 | 0.012 | 0.016 | not available | not available |
| Class B (mid-size) | 0.03445 | 0.00101 | 0.00046 | 0.00293 | 0.039 | 0.053 | not available | not available |
| Class C (frontier) | 0.12827 | 0.00388 | 0.00169 | 0.01284 | 0.147 | 0.2 | not available | not available |
| Model class | Central | Class range low | Class range high | High case | B200 | Development |
|---|---|---|---|---|---|---|
| Class A | 0.01153 | 0.007116 | 0.01671 | 0.01851 | 0.0101 | 0.01434 |
| Class B | 0.03885 | 0.02502 | 0.04717 | 0.06173 | 0.02718 | 0.04697 |
| Class C | 0.1467 | 0.05751 | 0.2197 | 0.2319 | 0.1509 | 0.1749 |
How much does siting matter for a model served from Global average?
Electricity is 88.7% of the class B total here (class A 91.2%, class C 87.4%), with hardware at 2.6%, construction 1.2% and training 7.6%. Electricity carries most of the 0.0388 kg, so the grid is the biggest lever.
The example company (2,323 million class B tokens a year) emits 90 kg CO₂e from Global average. Moving to US Central - Iowa would save 1.8 kg (2%) or US South - Southeast would save 8.6 kg (9.6%).
| Region | Grid g per kWh | Class A | Class B | Class C |
|---|---|---|---|---|
| US Central - Iowa | 472.1 | 0.562 | 1.84 | 6.85 |
| US South - Southeast | 431 | 2.64 | 8.64 | 32.2 |
How much water does AI usage consume in Global average?
Water: not available for Global average. The dataset publishes no off-site water factor for this region, so SOMA prints no water figure rather than a partial one.
Where do these numbers come from and how do I cite them?
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.
Cite as "0.0388 kg CO₂e per million tokens, class B, Global average, SAEF, 19 October 2026" and link somaai.earth/ai-footprint/factors/global-avg.