The GPAI Code of Practice: what signing it means for energy reporting
What the code is
The AI Act's obligations for general-purpose models are written in general terms; the Code of Practice, drafted through 2025 by independent chairs with input from industry, civil society and academia and finalised in July, translates them into concrete commitments. Providers that sign gain a presumption of compliance with the relevant articles. Those that do not must show compliance by other means and can expect closer scrutiny from the AI Office.
The energy field
The code's transparency chapter includes a model documentation form, the record a provider keeps and shares with the AI Office and with companies that build on the model. Among its fields is energy consumption for training, known or estimated, consistent with Annex XI. The form is a document, not a publication; downstream providers see it, the public does not.
What it does not ask
Energy per prompt. Water. Emissions, or the grid intensity that would turn energy into emissions. Hardware. The index's nine-item disclosure checklist contains eight items the form does not touch, and the ninth, training emissions in model cards, only in the sense that a provider could choose to publish what it documented.
Who signed
Most large providers indicated they would sign in the months after publication; Meta was reported to have declined, objecting to the code's scope. Signing does not change what a company publishes about its footprint. Google signed and publishes the most in the index; OpenAI signed and publishes one sentence. Mistral, which helped shape the European debate, published a full life-cycle assessment the same month, which no code required.
The index's reading
The code is a compliance instrument, and a reasonable one for its purpose. It is not an environmental disclosure regime, and it should not be described as one. Where the index credits a company for transparency, it is for what the company put in public, not for a form it filed in Brussels.
Companies in this post
Mistral AIThe only frontier lab with a peer-reviewed life-cycle assessment.75
OpenAI (ChatGPT)A single blog-post number, and gigawatts of gas-backed capacity on the way.40
Google (Gemini)The most detailed per-prompt disclosure — attached to the fastest-growing footprint.71
Meta (Llama / Meta AI)Model cards disclose training emissions; Hyperion will be powered by new gas plants.56
Sources
- 1.General-Purpose AI Code of PracticeEuropean Commission (via artificialintelligenceact.eu) · Jul 10, 2025
- 2.EU AI Act — Annex XI: technical documentation for general-purpose AI modelsartificialintelligenceact.eu · Jul 12, 2024
- 3.SCI for AI and the EU AI ActGreen Software Foundation · Mar 1, 2025
Figures labelled as estimates on the index carry roughly ±3× uncertainty; disclosed figures are the company’s own. Corrections: team@ecogpt.com.



