The European Commission's AI Office has opened a targeted consultation on measuring energy consumption and emissions of AI models and systems, signaling regulatory attention to the environmental impact of generative AI deployments [source]. The consultation, published in April 2025, seeks input on methodologies for quantifying the carbon footprint and power draw of foundation models and AI systems covered under the EU AI Act.

The consultation document acknowledges gaps in current measurement standards, noting that energy consumption varies significantly across training, fine-tuning, and inference phases. It proposes frameworks for mandatory reporting by general-purpose AI providers, including disclosure of total energy used per model and emissions per query. The Commission is soliciting feedback on whether reporting should be standardized at the model level, system level, or both.

This follows observations by independent researchers that major providers do not uniformly disclose energy metrics. OpenAI, Anthropic, and Google have published partial data on training costs for select models, but inference-level consumption remains largely unreported. The consultation references studies estimating that a single ChatGPT query consumes approximately 2.9 watt-hours, though providers have not confirmed these figures.

The consultation runs until June 2025 and targets AI developers, cloud infrastructure providers, and civil society organizations. Responses will inform technical annexes to the AI Act's implementing regulations, expected in late 2025. The Commission states that measurement standards are necessary for enforcing Article 52a of the AI Act, which requires transparency about AI systems' environmental impact.

The consultation does not propose specific consumption limits, focusing instead on disclosure requirements and measurement methodologies.