En­er­gy and re­source ef­fi­cien­cy

• Requirements for AI models and high-risk AI systems
• Codes of conduct
• Further information

Today increasingly large amounts of electricity, water and raw materials are already needed for manufacturing computer chips, building and operating data centres, and training and storing AI models and systems and using them on a daily basis. The International Energy Agency (IEA) estimates that in 2024 data centres accounted for up to 1.5% of global electricity demand and forecasts that demand from data centres will double by 2030.

Requirements for AI models and high-risk AI systems

Article 53 of the AI Act requires providers of general-purpose AI models to draw up and keep up-to-date technical documentation for their models. This technical documentation includes information about the energy consumption of the models. If the energy consumption of a model is unknown, it can be based on information about the computational resources used.

The Transparency Chapter of the General-Purpose AI Code of Practice published by the European Artificial Intelligence Board (AI Board) includes a Model Documentation Form that providers can use to document the energy consumption of their models. More information about the Code of Practice can be found here.

The European Commission issues standardisation requests under Article 40(2) of the AI Act relating to the resource and energy efficiency of AI models and systems. One such request addresses reducing the consumption of energy and other resources during the lifecycle of high-risk AI systems and the energy-efficient development of general-purpose AI models.

The standards to be developed in response to the above-mentioned standardisation request have not yet been delivered. The detailed requirements for documenting the resource and energy efficiency of AI have therefore not yet been defined.

Codes of conduct

The European AI Office and the Member States encourage and facilitate the drawing up of codes of conduct for the voluntary application of specific requirements in accordance with Article 95 (2) point (b) of the AI Act. One of the elements to be covered is assessing and minimising the impact of AI systems on environmental sustainability, including energy-efficient programming and techniques for the efficient design, training and use of AI. The codes of conduct should have clear objectives and key indicators to measure the achievement of those objectives. Codes of conduct can be drawn up by individual providers or deployers of AI systems or by organisations representing them.

Further information

Green AI Guidelines: The guidelines set out specific measures and tools that are aimed at making the design and use of AI systems in the fields of management, procurement and operation as environmentally sustainable as possible. They include guidance relating to green and efficient software development and to the choice of suitable hardware and data centres. The guidelines were developed by the Green-AI Hub Mittelstand of Germany’s Federal Ministry for the Environment, Climate Action, Nature Conservation and Nuclear Safety (BMUKN) together with the German AI Association and other players from the environmental, green IT and AI sectors.

SustAIn: The project consortium (AlgorithmWatch, Institute for Ecological Economy Research (IÖW), Distributed Artificial Intelligence Laboratory of TU Berlin), which is supported by the BMUKN, has drawn up checklists with criteria for the socially, environmentally and economically sustainable design of AI systems during their lifecycle.

SustainML: The consortium from industry and academia (IBM, German Research Center for Artificial Intelligence (DFKI), Inria, University of Copenhagen (UCPH), UPMEM and RPTU University Kaiserslautern-Landau) is dedicated to creating a sustainable, interactive framework for machine learning. This framework includes resource-aware AI optimisation methods, design assistants that make carbon footprints transparent for AI developers, and catalogues and libraries of energy-optimised parameterised machine learning models.