In short
If you provide a general-purpose AI model in the EU, the AI Act's Annex XI requires your technical documentation to include the model's known or estimated energy consumption, along with the computational resources used for training. Track training compute and energy from the start, record the method, and estimate where you cannot measure. If you only use AI, the duty falls on your supplier, but you can ask for the figures.
Steps
- Check whether you are a provider of a general-purpose AI model or a deployer of someone else's.
- If you are a provider, log training compute, time and hardware for every significant run.
- Measure energy where you can and estimate the rest with a written method.
- Keep the figures and method in the model's technical documentation.
- If you are a deployer, ask your model provider for its documented energy figures.
The duty is to document known or estimated energy use, so the practical task is to record it as you go. Annex XI of the EU AI Act lists what providers of general-purpose AI models must include in their technical documentation, and that includes the model's known or estimated energy consumption and the computational resources used to train it. Where energy is not known, it may be based on information about the computational resources used.
This is a summary, not legal advice. Check the current text and guidance, which the European Commission updates.
Does it apply to me?
- Providers that place a general-purpose AI model on the EU market: yes.
- Organisations that only use AI tools or build on someone else's model: generally not this duty, but the documentation your provider holds is useful in procurement.
- Outside the EU: it applies if you supply such a model into the EU. Elsewhere, it is a sensible standard to follow voluntarily.
How do I prepare the figures?
- Wrap training runs in an energy tracker such as CodeCarbon, and keep hardware, hours and location.
- Include tuning and failed runs, not only the final one, and say so.
- State the boundary (accelerators only, or the whole server and cooling).
- Publish the same data in your model card; Hugging Face's
co2_eq_emissionsmetadata has fields for emissions, training type, location and hardware.
Common mistakes
- Reconstructing figures at the end from memory.
- Reporting a number without the method.
- Assuming a deployer has nothing to do: asking providers for these figures is good procurement.
See measure the footprint of training and energy labels for AI.
Sources
- AI Act, Annex XI: technical documentation for providers of general-purpose AI models, European Commission AI Act Service Desk (2024)
- CodeCarbon, CodeCarbon (2026)
- Displaying carbon emissions for your model, Hugging Face Hub documentation (2026)
Related
How to write a green AI policy
A step-by-step guide to writing an organisation's green AI policy, with a template outline covering use, models, measurement, buying, claims and review.
How do I measure the carbon footprint of training a model?
Track the energy of a training run with a tool such as CodeCarbon, multiply by local grid carbon intensity, and report hardware, time and location.
How do I put sustainability into AI procurement?
What to ask AI suppliers about energy, carbon, water and location, how to score the answers in a tender, and what to write into the contract.
Energy labels and benchmarks for AI
Where AI energy ratings stand in 2026: AI Energy Score, MLPerf Power, EU AI Act documentation and data-centre rating, and what labels can show.