EUはすでにAIモデル提供者にエネルギー消費量の文書化を求めている。ただし公開される数字ではない
EU AI法の附属書XIは、汎用AIモデルの提供者が作成・保持する技術文書に「既知または推計のエネルギー消費量」を含めることを求めている。義務は2025年8月2日から適用されている。情報は下流提供者とAI Officeの求めに応じて提供されるもので、一般公開されるものではない。すでに市場に出ているモデルの期限は2027年8月2日である。
Almost every number in the public argument about AI energy use is an outside estimate. A researcher counts GPUs, assumes a utilisation rate, picks a grid emissions factor and publishes a range. The company that actually metered the training run says nothing.
That asymmetry has a legal answer in Europe, and it has been in force for more than a year. Since 2 August 2025 the providers of general-purpose AI models have had to write their own energy figure down.
They do not have to publish it, and that distinction decides what the obligation is actually worth to anyone outside the company.
The obligation is one line in an annex
The AI Act, Regulation (EU) 2024/1689, sets out in Annex XI what a provider of a general-purpose AI model must put in its technical documentation. Section 1 lists the items, and says they apply “as appropriate to the size and risk profile of the model”.
Two of them sit next to each other and are worth reading in full.
Point 2(d) asks for “the computational resources used to train the model (e.g. number of floating point operations)” together with “training time, and other relevant details related to the training”.
Point 2(e) asks for the “known or estimated energy consumption of the model”.
A note immediately after it says that “where the energy consumption of the model is unknown”, the figure “may be based on information about computational resources used”.
So the law asks for compute and it asks for energy, and it accepts a conversion from the first to the second when the second was never measured.
Compute is the fallback, and the Annex names no methodology
That fallback is generous and it is honest about why it exists. Many providers rent capacity. A company training on someone else’s cluster may genuinely not hold a meter reading, and the Annex does not pretend otherwise.
It also does not prescribe a method. Annex XI names no conversion factor, no boundary for what counts as the model’s energy, and no rule on whether inference belongs in the number at all. Point 2(e) says “of the model”, not “of the training run”.
Two providers can both comply and produce figures that are not comparable. One can report metered training electricity. The other can report a compute-derived estimate with a different assumed efficiency. Nothing in the text forces them onto the same basis.
For anyone building an energy argument on these disclosures, that is the first thing to check and the last thing to assume.
Documentation is not publication
The second trap is reading an Annex XI item as a public disclosure. It is not.
The Code of Practice for general-purpose AI models, published by the European Commission on 10 July 2025, sets out how signatories meet Article 53(1). Its Transparency chapter contains three measures and a Model Documentation Form that gathers the Annex XI items in one place.
Measure 1.2 governs who sees what. Signatories publish contact details so that the AI Office and downstream providers can request access. Information intended for the AI Office or national competent authorities is made available “only to be made available following a request from the AI Office”, and such requests cover only items “strictly necessary” for the task at hand. Downstream providers get the information in the Model Documentation that is intended for them, and additional information on request where they need it to understand the model and meet their own obligations under the Act.
Article 78 of the Act then binds every recipient to confidentiality, “in particular intellectual property rights and confidential business information or trade secrets”.
The energy figure exists, is kept up to date, and is reachable by a defined set of people for defined reasons, none of which include general publication.
Free and open-source models sit outside it
Article 53(2) exempts providers of general-purpose AI models released under a free and open-source licence meeting the conditions in that provision, unless the model is a general-purpose AI model with systemic risk. The Transparency chapter repeats the carve-out in the same terms.
Open weights therefore do not imply open documentation, and the energy figure is the item most likely to be missing where the weights are most freely available.
The dates that are already behind us
Article 113 of the AI Act stages the application. Chapter V, which carries the general-purpose AI model obligations, applies from 2 August 2025. The Regulation as a whole applies from 2 August 2026.
Article 111(3) is the one a diligence process should read. Providers of general-purpose AI models “that have been placed on the market before 2 August 2025” must “take the necessary steps in order to comply with the obligations laid down in this Regulation by 2 August 2027”.
So the model families that dominate production use today are inside a transition that ends in under two years. A model shipped after 2 August 2025 should already have the documentation. A model shipped before it may not, lawfully, until 2027.
The Code of Practice adds a retention rule on top. Signatories keep previous versions of the Model Documentation “for a period ending 10 years after the model has been placed on the market”. The record is meant to outlive the model.
A downstream buyer has fourteen days of leverage
The practical consequence runs through the supply chain rather than the press release.
If you integrate a general-purpose AI model into your own system, you are a downstream provider, and the Code of Practice commits signatories to answer your request “within a reasonable timeframe, and no later than 14 days of receiving the request save for exceptional circumstances”.
That is a usable procurement right. A company with a climate target that now runs a large share of its workload through someone else’s model has a route to the provider’s own energy and compute figures, under confidentiality, on a fixed clock.
Most buyers are not using it. The request costs an email and the answer, however imperfect its methodology, beats the public estimate it would otherwise have to cite.
The EU measures the building and the model separately
Europe now asks two different questions of the same electricity.
Data centres above a size threshold report energy, water and waste heat into an EU database, which we covered in Europe Measures Its Data Centres at 500 kW. The Form Asks About Water., and that regime measures the facility.
Annex XI measures the model. A model trained in Virginia and served from Dublin touches both regimes at different points, and nothing in either text reconciles one to the other. For an investor looking at AI infrastructure, that gap is where the exposure sits, because a customer asking what its inference cost in electricity is asking a question no single European filing currently answers.
What to do with this
If you provide a model, the compliance question is not whether you can produce a number. It is whether you can defend the basis of the one you produce, because the Annex lets you choose it and the AI Office can ask you to explain it.
If you buy a model, send the request. Ask what the figure covers, whether it is metered or derived from compute, and whether inference is in it. The answers to those three questions tell you more than the number.
If you invest in either, note the asymmetry in the timetable. The documentation obligation is already live for anything new, the transition for incumbents closes on 2 August 2027, and the records survive a decade.
Tech for Impact Summit 2027 is co-hosted by the United Nations University in Tokyo on 18 and 19 May 2027, an invitation-only gathering of leaders across business, policy and culture working on technology against the hardest problems. The energy behind AI, and who has to account for it, is on the agenda.