Eight mentions, one obligation

Energy appears eight times in the EU AI Act. Most of those are aspiration, sandbox eligibility, or a job for the Commission. Exactly one is a binding obligation on a provider to record a number, and it is narrower than almost any summary of it suggests.

This piece works from the instrument. All clause references were read in the consolidated text of Regulation (EU) 2024/1689 as amended to 27 July 2026, with recitals checked against the original Official Journal text, because consolidated versions omit preambles.

The one binding line

Annex XI, Section 1, point 2(e) requires providers of general-purpose AI models to include, in the technical documentation, the "known or estimated energy consumption of the model". The neighbouring point 2(d) requires the computational resources used for training. The obligation is switched on by Article 53(1)(a), and it has applied since 2 August 2025, when the general-purpose AI chapter came into application.

That is the whole of it. There is no equivalent for deployers, none for high-risk systems, and none anywhere in the Act for the energy consumed when a model is actually used.

Four features that limit what it delivers

1. It is written to be read by regulators, not published. Article 53(1)(a) requires the documentation to be drawn up and kept up to date "for the purpose of providing it, upon request, to the AI Office and the national competent authorities". It is a file you produce when asked. Article 78 attaches confidentiality obligations to what authorities receive.

2. It does not travel downstream. Annex XII lists what a general-purpose model provider must give to the providers who build on the model. It contains no energy item and no compute item. So an organisation integrating a foundation model has no entitlement under the Act to the figure its supplier has recorded.

3. There is no method. The Act specifies no unit, no measurement boundary and no time period. Article 53(5) empowers the Commission to adopt delegated acts detailing "measurement and calculation methodologies with a view to allowing for comparable and verifiable documentation", naming points 2(d) and (e) specifically. A power to make figures comparable is an admission that they are not comparable yet. The Act's only fallback is that where energy consumption is unknown, it may be estimated from information about computational resources used.

4. Open-source models are largely outside it. Article 53(2) exempts providers of models released under a free and open-source licence, with weights, architecture and usage information published, from Article 53(1) points (a) and (b) altogether, unless the model is a general-purpose AI model with systemic risk.

Systemic-risk models: one more line, still training-only

For general-purpose AI models with systemic risk, Annex XIII(c) adds the "estimated energy consumption for the training". The wording repays attention. It is training energy, estimated, for the subset of models that cross the systemic-risk threshold. Inference, the energy consumed every time anyone actually uses the model, is not addressed there or anywhere else in the Act.

Whether Annex XI point 2(e) itself reaches inference is genuinely unsettled. It says "energy consumption of the model", which is broader than point 2(d)'s training-specific wording, but it sits under a chapeau concerned with the development process. Anyone telling you the Act clearly does or clearly does not cover inference energy is reading past the ambiguity.

High-risk systems: nothing, and a duty on the Commission

Articles 8 to 15, which carry the substantive requirements for high-risk AI systems, impose no resource-use or environmental requirement. The closest thing is Annex IV point 2(c), which asks for the computational resources used to develop, train, test and validate the system. That is compute, not energy, and following the Digital Omnibus it does not apply until 2 December 2027 for standalone Annex III systems and 2 August 2028 for Annex I embedded ones.

The only provision in the Act tying energy to a high-risk system is Article 40(2), and it points the other way:

"The standardisation request shall also ask for deliverables on reporting and documentation processes to improve AI systems' resource performance, such as reducing the high-risk AI system's consumption of energy and of other resources during its lifecycle, and on the energy-efficient development of general-purpose AI models."

That is an instruction to the Commission to ask European standardisation bodies for deliverables. It creates no provider obligation, and a harmonised standard would not create one either: under Article 40(1) such standards confer a presumption of conformity with requirements that exist, and Chapter III Section 2 contains no energy requirement for a standard to attach to. Article 40 sits in Chapter III Section 5, which Article 113 does not carve out, so it has applied since 2 August 2026.

As far as I can establish, it has not yet been executed for energy. The standing standardisation request, Commission Implementing Decision C(2025) 3871 final of 23 June 2025, contains no occurrence of the string "energ" in its enacting text, its recital 6 fixes the scope to the earlier 2023 annex, and the Commission's own explainer on AI Act standardisation lists ten areas for harmonised standards with no energy item among them. That evidence is strong but not complete: Annexes I and II to that decision, which carry the list of requested deliverables, could not be retrieved.

What the Code of Practice adds, and who it still excludes

The Commission-facilitated Model Documentation Form for general-purpose AI providers, published 10 July 2025, does what the Act declines to do. It asks for the figure in megawatt-hours to at least two significant figures. It also shows the limits plainly: it permits "N/A" where a compute or hardware provider will not supply the information, it marks the energy field for the AI Office and national competent authorities rather than for downstream providers, and it asks for inference in floating point operations rather than in energy.

So the one hard number the Act generates about AI energy is a training-only figure, in a form the public does not see, which a provider may decline to produce by pointing at its cloud vendor.

Why this matters more than it looks

Enforcement is now real. Article 101 lets the Commission fine general-purpose model providers up to 3% of worldwide annual turnover or 15 million euro, whichever is higher, and it has applied since 2 August 2026. A documentation obligation with a penalty attached is a different thing from a voluntary disclosure, even when nobody outside a regulator reads the document.

The practical consequences for an organisation buying rather than building AI:

  • Do not expect the Act to source your emissions data. Annex XII does not give you the number. If you need your supplier's energy figures for your own reporting, that is a procurement and contract question, not a compliance entitlement.
  • Do not treat a supplier's figure as comparable to anyone else's. Until a methodology exists under Article 53(5), two providers can produce honest numbers on incompatible bases.
  • Training energy is not your operating footprint. Both Act-mandated figures concern training. Your own consumption is dominated by inference, which the Act does not reach.
  • Watch 2 August 2028. Article 112(6) requires the Commission to report on progress on standardisation deliverables for energy-efficient development, and to assess "the need for further measures or actions, including binding measures or actions". The Act is telling you it expects this area to be revisited.

What I could not establish

Two gaps, stated rather than papered over. First, I could not confirm whether any delegated act under Article 53(5) has been adopted; the Commission's own Model Documentation Form of July 2025 is drafted around its absence, but that is evidence of the position then, not now. Second, the annexes to the standardisation request could not be retrieved, so the conclusion that it contains no energy deliverable rests on the decision body, its recital 6 and the Commission's explainer rather than on the annex itself.

Both would change the detail here, not the shape: the Act still contains one binding provider-facing energy line, still confines it to training, and still routes it to regulators rather than to the market.

This article describes the text of Regulation (EU) 2024/1689 as consolidated to 27 July 2026. It is general information about the state of the instrument, not legal advice, and the position on delegated acts and standardisation requests can change without the Act itself changing. Verify against the consolidated text and your own advisers before relying on it.

Related reading

Sources: Regulation (EU) 2024/1689, consolidated text to 27 July 2026 (EUR-Lex) · General-Purpose AI Code of Practice and Model Documentation Form, 10 July 2025 (European Commission) · Understanding the standardisation of the AI Act (European Commission)