Last reviewed:
What is the AI carbon footprint? Definition and business implications
The AI carbon footprint is the set of greenhouse gas emissions related to its life cycle: model training, component manufacturing, data centre operation, and inference at usage. It is becoming a significant consumption item on a global scale.
The AI carbon footprint breaks down into four main items. Hardware manufacturing: production of GPUs, servers, coolers, whose upstream footprint is rarely accounted for. Model training: a high-energy-cost operation, one-off but repeated at each new generation. Training GPT-3 consumed about 1,287 MWh and emitted 552 tons of CO2 equivalent (Patterson et al., 2021). Data centre operation: a continuous item, dominant cumulatively. According to the IEA, global data centres consumed 415 TWh in 2024 (1.5% of global electricity), with a projection of around 950 TWh in 2030. Inference at usage: a small unit operation multiplied by billions of daily calls. According to the IEA, electricity consumption from AI-focused data centres grew by 50% in 2025.
Concrete example
Take a European company deploying an AI assistant for 500 employees, at 20 requests per day over 220 working days: 2.2 million requests per year. Google measures 0.24 Wh for a median text prompt on Gemini, data centre overhead included. Direct consumption would therefore be around 530 kWh per year. With the average carbon intensity of electricity generated in the EU in 2024 (about 183 g of CO2e per kWh according to the European Environment Agency), that is on the order of 100 kg of CO2e per year. The result varies widely from one country to another depending on the electricity mix. That is modest compared with the carbon balance of an SME. But the ratio changes for intensive uses: video generation, multi-step reasoning models, autonomous agents. The IEA estimates that these tasks can consume hundreds or even thousands of times more energy per query than simple text generation. The choice between an optimised lightweight model and a flagship model has a direct impact on the footprint.
FAQ
What is the AI carbon footprint?
It is the set of greenhouse gas emissions related to its life cycle: component manufacturing, model training, data centre operation, and inference at usage. It is becoming a significant consumption item on a global scale.
What are the main emission items?
Four. Hardware manufacturing (GPUs, servers, coolers), whose upstream footprint is rarely accounted for. Model training: training GPT-3 consumed about 1,287 MWh and emitted 552 tons of CO2e. Data centre operation, a continuous and cumulatively dominant item: according to the IEA, global data centres consumed 415 TWh in 2024. Inference at usage, small per unit but multiplied by billions of daily calls.
What is the order of magnitude for a company?
For an assistant used by 500 employees at 20 requests a day over 220 days, i.e. 2.2 million requests a year, direct consumption would be around 530 kWh, on the order of 100 kg of CO2e per year with the EU average electricity mix in 2024. That is modest, but the ratio changes sharply for intensive uses (video, multi-step reasoning, agents), which can consume hundreds or even thousands of times more energy per query than simple text generation.
How can this footprint be reduced?
Through usage sobriety, a double dividend: it cuts the bill (inference cost proportional to tokens) and the footprint (energy consumption proportional). Model choice must include an explicit energy criterion, since for the same task a small model consumes far less than a flagship one. Ask providers for their per-request consumption measurements.
Is there a regulatory obligation?
Yes. Since 2 August 2025, the European AI Act has required providers of general-purpose models to document the known or estimated energy consumption of their models (Annex XI). The Commission can enforce these obligations with fines since 2 August 2026. This first baseline foreshadows more demanding reporting.
See also
Further reading
Sources
- Energy and AI, International Energy Agency, executive summary, 2025. https://www.iea.org/reports/energy-and-ai/executive-summary
- Key Questions on Energy and AI, International Energy Agency, executive summary, 2026. https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary
- Carbon Emissions and Large Neural Network Training, Patterson et al., arXiv:2104.10350, 2021. https://arxiv.org/abs/2104.10350
- Measuring the environmental impact of delivering AI at Google Scale, Google, arXiv:2508.15734, 2025. https://arxiv.org/abs/2508.15734
- Greenhouse gas emission intensity of electricity generation in Europe, European Environment Agency, 2026. https://www.eea.europa.eu/en/analysis/indicators/greenhouse-gas-emission-intensity-of-1
- Regulation (EU) 2024/1689 (AI Act), Article 53, Annex XI and Article 113. https://eur-lex.europa.eu/eli/reg/2024/1689/oj