EU AI Act Annex XI: energy must be documented, not disclosed

ecogpt.coPublished Oct 19, 2026 · 2 min read

From August 2025, providers of general-purpose AI models in Europe must keep technical documentation that includes the model's energy consumption, known or estimated. They do not have to publish it. Here is what the rule requires, what it allows, and what it leaves out.

Annex XI lists energy consumption among the items in the technical documentation for general-purpose AI models.
Where consumption is unknown, an estimate based on compute is acceptable.
The documentation goes to the AI Office and national authorities on request, not to the public.
Water, embodied carbon and inference are not covered; the obligation is about training.

The obligation

The EU AI Act's rules for general-purpose AI models began to apply on 2 August 2025. Annex XI sets out the technical documentation a provider must prepare and keep up to date, and among the items it lists is the model's energy consumption: the known or estimated energy used in training. The provider must be able to supply it to the European AI Office and to national authorities on request.

What "known or estimated" means

The text allows an estimate where measurement is not available, and the drafting anticipates that estimates will be based on compute: the number of operations the training run required, converted to energy by assumptions about hardware and efficiency. That is roughly how the index estimates training footprints for companies that publish nothing. The difference is that the index publishes its working. The Act does not require the provider to publish anything.

What it does not cover

Inference, which is where most of the electricity of a deployed model goes; water, in any form; the embodied carbon of hardware; the grid on which the model was trained, so an estimate in kilowatt-hours carries no emissions figure. The Green Software Foundation and others have argued for a carbon-intensity metric to sit alongside energy. The Act does not provide one.

The code of practice

The General-Purpose AI Code of Practice, published in July 2025 as a voluntary route to compliance, includes a model documentation form with an energy field. Signing it is optional; the documentation remains confidential to regulators. Mistral's life-cycle assessment, published the same month, shows what a public version would look like: absolute training impact, marginal inference impact, and the ratio between them, across carbon, water and minerals.

Why the index exists in the gap

A regulation that requires documentation but not disclosure produces compliance without information. The index's disclosure checklist asks for the things the Act does not: per-prompt energy and water, training emissions in model cards, a public annual report, location-based Scope 2, fleet WUE, third-party review and hardware circularity. Only Google and Mistral come close. The Act, as written, would not have changed a single score.