- EN English
Decision on how the European Commission assessed energy consumption provisions of the Code of Practice related to the AI Act (case 3162/2025/MIK)
Decision
Case 3162/2025/MIK - Opened on Thursday | 19 February 2026 - Decision on Tuesday | 06 October 2026 - Institution concerned European Commission ( No further inquiries justified ) - Country Belgium
Complaint submitted
29/10/2025Analysis of the complaint
29/10/2025Inquiry ongoing
19/02/2026Inquiry outcome
06/10/2026
The complainant, a Member of the European Parliament, raised concerns about what he believed was a discrepancy between the AI Act and the Code of Practice for general purpose AI (GPAI). The European Commission recently confirmed that the Code, developed by stakeholders, is an ‘adequate’ voluntary tool for GPAI providers to demonstrate compliance with the AI Act. However, the complainant argued that the Code exempts GPAI providers from their duty under the AI Act to document or at least estimate the energy consumption of their models during their training and testing.
The Ombudsman inquired into this issue and found the Commission reasonably explained how its AI Office intends to assess the energy consumption of GPAI where the providers lack certain information necessary to make the estimation. Thus, the Commission provided the necessary assurances that it will fulfil the underlying objective of the AI Act in this regard. In this context, the Ombudsman noted that the Code is intended as a temporary solution, which may be further improved and is likely to be replaced by targeted instruments, such as ‘harmonised standards’, developed by the European Standardisation Organisations.
On this basis, the Ombudsman found that, at this stage, no further inquiries into this complaint are justified.
Background to the complaint
The requirement to document or estimate the energy consumption of GPAI under the AI Act
1. General-purpose AI (GPAI) systems are designed to perform many different tasks (such as for instance online AI chats). GPAI providers play a key role because many AI systems designed for more specific tasks (such as, for instance, speech recognition) are built using GPAI models. Due to this important role of GPAI, regulators and companies involved in AI need enough information about GPAI models to understand their capabilities, risks, and impacts, including environmental impacts.
2. This is why, under the AI Act[1], GPAI providers are required to keep up-to-date technical documentation about their models, including information about the energy consumption of these models, and share this information with the competent authorities upon request. This documentation must also cover the “training and testing process” of the models,[2] that is, the development processes during which the model is being shown extremely large amounts of data so it can learn patterns to make plausible predictions and generate answers to users’ prompts in the future. This documentation must contain “known or estimated energy consumption of the model”, including the energy consumption during the process for developing the model.[3]
3. At the same time, the AI Act stipulates that “where the energy consumption of the model is unknown, the energy consumption may be based on information about computational resources used”.[4] In other words, if a provider does not know exactly how much energy the model uses or used during training, it is allowed to estimate it by looking at the computing power and hardware used to create or run the model, that is: the number of servers, chips, training time, and computing hours. The AI Act assumes that it is possible for providers to work out a reasonable estimate of energy consumption based on these factors.
The role of the Code of Practice for GPAI
4. The AI Act, as many pieces of EU legislation concerning product safety, lays down general legal requirements, which need further specification of technical solutions to meet these requirements. Therefore, the Commission’s AI Office[5] must encourage and facilitate the drawing up of ‘codes of practice’ fostering the proper application of the AI Act. It may invite GPAI providers, national competent authorities, civil society, academia, and other relevant stakeholders to contribute to drawing up such codes.[6] According to the AI Act, “the codes of practice should represent a central tool for the proper compliance with the obligations provided for under [the AI Act] for providers of general-purpose AI models. Providers should be able to rely on codes of practice to demonstrate compliance with the obligations.”[7]
5. At the same time, the codes of practice are intended as temporary solutions, until ‘harmonised standards’ are developed by the European Standardisation Organisations and approved by the Commission. As the AI Act says, “once a harmonised standard is published and assessed as suitable to cover the relevant obligations by the [Commission’s] AI Office, compliance with a European harmonised standard should grant providers the presumption of conformity” with the legal obligations imposed by the AI Act.[8] In addition, the Commission may, under some conditions, establish ‘common specifications’ through implementing acts, where harmonised standards turn out to be ultimately unavailable.[9]
6. Pursuant to the AI Act, the Commission should assess the adequacy of the codes of practice intended to facilitate the application of the AI Act and publish its assessment.[10] In that regard, a Code of Practice for GPAI has recently been drawn up by stakeholders, and the Commission (as well as the AI Board) assessed it as adequate.[11]
7. According to the Model Documentation Form proposed by this Code, “if the amount of energy used for training cannot be estimated due to the lack of critical information from a compute or hardware provider”, the users of the Code may explain the type of information they lack.[12]
The complaint
8. The complainant, a Member of the European Parliament, considered that the above solution proposed by the Code is not in line with the relevant provisions of the AI Act.
9. The complainant raised his concerns with the Commission. The Commission Vice-President responsible for Tech Sovereignty, Security and Democracy replied to the complainant’s concerns in September 2025. Dissatisfied with this reply, the complainant turned to the Ombudsman in October 2025.
The inquiry
10. The Ombudsman opened an inquiry to clarify how the European Commission assessed energy consumption provisions of the Code.
11. In the context of the inquiry, the Ombudsman inquiry team met with the representatives of the Commission to discuss the issue raised by this complaint. Also, the Ombudsman obtained the complainant’s comments on the meeting report.
Arguments presented to the Ombudsman
12. The complainant argued that, on the one hand, Article 53(1) and Annex XI of the AI Act require the providers of GPAI to draw up, keep up-to-date, and deliver to the AI Office and national competent authorities, upon request, information about the known or estimated energy consumption of their models used in the GPAI systems, including the energy consumption during their development, that is, training and testing. On the other hand, the Code appears to exempt providers from delivering this information “if the amount of energy used for training cannot be estimated due to the lack of critical information from a compute or hardware provider”. In such a case, providers may merely explain which type of information they lack. In this context, the complainant argued that there is a discrepancy between the AI Act and the Code.
13. The complainant added that the AI Act has a unique role in mitigating the environmental risks of AI. According to the complainant, projections suggest that AI-related electricity consumption could more than double by 2030, potentially comprising 3% of the EU’s total energy use, while in Ireland this figure could exceed 20%. The complainant was further concerned that individual data centres already consumed as much electricity as 100,000 households annually.
14. In its reply to the complainant, the Commission argued that, if providers lack critical information from a compute or hardware provider necessary to assess the energy consumption of the model during training, the AI Office will nonetheless estimate the energy consumption.
15. During the meeting with the Ombudsman inquiry team, the Commission representatives pointed out that Annex XI of the AI Act acknowledges that the exact energy consumption data may be unknown and can alternatively be estimated based on the computational resources used. However, there may be instances in which model providers cannot make an estimate because they lack critical data from the providers of computational resources hosting their training infrastructure. Therefore, according to the Code, model providers can indicate “N/A” [Not Available] and clearly indicate which critical information they lack. The AI Office will then assess whether the provided indication of missing critical information is reasonable. In such cases, the AI Office will make its own assessment of the energy consumption based on the information provided and other information at its disposal, as well as the findings of the ongoing Commission study on energy-efficient and low-emission AI[13].
16. In addition, the Commission representatives clarified that, when the Commission assesses a code of practice as adequate, GPAI providers have a legitimate expectation that the Commission will consider that adherence to an endorsed code of practice is sufficient to demonstrate compliance with their respective obligations under the AI Act.
17. In response to the Commission’s explanation presented during the meeting with the Ombudsman inquiry team, the complainant argued that the approach proposed by the Commission is neither necessary nor acceptable under the AI Act. According to the complainant, the provider has more information than the Commission about the training of its model, so there is no reason why making the estimate should be feasible for the Commission but not for the provider. Also, the complainant contended that if the Commission has information that may be helpful in making the estimates of energy consumption, it should publish that information, thereby sharing it with GPAI providers who struggle with making their estimations. In addition, the complainant pointed out that the Commission’s approach is unlawful because there is no legal basis in the AI Act for the Commission to make an estimate of energy consumption. For the complainant, the Commission should initiate a rapid update of the Code.
The Ombudsman's assessment
18. The Ombudsman notes that the Commission assessed the Code to be an adequate tool for providers to demonstrate compliance with the AI Act, including in relation to maintaining the up-to-date documentation concerning the energy consumption of GPAI. The Ombudsman understands that, consequently, the Commission will not, in principle, depart from solutions prescribed by the Code in assessing individual cases. Therefore, this Code may play a pivotal role in guiding its signatories in complying with the AI Act, at least at the early stage of its implementation, when harmonised standards or the Commission’s specifications are not yet available.
19. The Ombudsman takes note of the complainant’s concern that the Code appears to exempt the providers, in instances in which they lack ‘critical information’, from their duty to estimate the energy consumption of their models during training and testing instead of requiring them to take extra steps to obtain this information.
20. In such situations, the Commission appears to argue that providers will discharge their duty under the AI Act by explaining which information they lack. Based on this and other information at its disposal, according to the Commission, its AI Office will be in a position to estimate the relevant energy consumption in a sufficiently reliable manner, thus fulfilling the underlying objective of the AI Act. The Commission accepts this solution because it is concerned that, in some cases, it may be technically impossible for providers of GPAI to obtain technical information about the computational resources used for their models.
21. However, the Ombudsman also notes the complainant’s concern that the AI Office is potentially assuming a responsibility that the AI Act places upon the GPAI providers.
22. In general, the Ombudsman observes that the process of implementing the AI Act is particularly complex, as it involves an assessment of entirely new technological opportunities, challenges, and solutions, and implies trial phases and ongoing improvements. The Ombudsman recognises the Commission’s expertise and discretion in determining how to best proceed in attaining the AI Act’s objectives, especially given the rapid technological evolution of GPAI.
23. The Ombudsman further notes that the Code is intended to be a temporary instrument fostering compliance with the AI Act. Consequently, the solutions it proposes may be further improved. In particular, under the AI Act, the Commission must regularly monitor and evaluate the achievement of the objectives of the codes of practice.[14] Moreover, the Commission AI Office is obliged to encourage and facilitate the review and adaptation of the codes of practice, in particular in light of emerging standards.[15] In any case, codes of practice should be ultimately replaced by harmonised standards or, if the adoption of those fails, the Commission’s own common specifications.[16]
24. In addition, the Ombudsman observes that, under Article 40(2) of the AI Act, the Commission is bound to 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 GPAI. The Ombudsman understands that these deliverables may further improve the tracing and help reduce the energy consumption of AI in the context of the gradual process of the AI Act’s implementation.
25. In this context, the Ombudsman finds the Commission’s position reasonable and that, at this stage, no further inquiries into this case are justified.
26. At the same time, the Ombudsman trusts that, in the context of the gradual and challenging process of implementing the AI Act, the Commission will strive to improve technical solutions to enable GPAI providers to meet the requirements of this Act, including in relation to documenting the energy consumption of GPAI.
Conclusion
Based on the inquiry, the Ombudsman closes this case with the following finding[17]:
No further inquiries are justified given the Commission’s commitment to fulfil the AI Act’s objectives in terms of tracing the energy consumption of GPAI even in cases in which their providers lack critical information about computational resources used. The Ombudsman trusts that the Commission will strive to improve technical solutions to enable GPAI providers to meet the requirements of the AI Act in the gradual process of its implementation.
The complainant and the Commission will be informed of this decision.
Teresa Anjinho
European Ombudsman
Strasbourg, 06/10/2026
[1] Regulation 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence (...) (the ‘AI Act’), OJ L 2024/1689, https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX%3A32024R1689.
[2] Article 53(1)(a) of the AI Act.
[3] Annex XI, Section 1, point 2(e) of the AI Act.
[4] Ibid., the last sentence.
[5] According to Article 3(47) of the AI Act, “‘AI Office’ means the Commission’s function of contributing to the implementation, monitoring and supervision of AI systems and general-purpose AI models, and AI governance, provided for in Commission Decision of 24 January 2024 (...)”. See https://digital-strategy.ec.europa.eu/en/policies/ai-office.
[6] Article 56 of the AI Act.
[7] Recital 117 of the AI Act.
[8] Ibid.
[9] Article 41(1) of the AI Act.
[10] Article 56(6) of the AI Act.
[11] https://digital-strategy.ec.europa.eu/en/policies/contents-code-gpai
[12] Ibid. See ‘Model Documentation Form’ attached to the ‘Transparency Chapter’.
[13] https://digital-strategy.ec.europa.eu/en/consultations/targeted-consultation-measuring-energy-consumption-and-emissions-ai-models-and-systems
[14] Article 56(6) of the AI Act.
[15] Article 56(8) of the AI Act.
[16] Articles 41 and 56 of the AI Act, as well as Recital 117 of the AI Act.
[17] This complaint has been dealt with under delegated case handling, in accordance with the Decision of the European Ombudsman adopting Implementing Provisions