There is a lot of public debate lately about controlling artificial intelligence, and it mostly focuses on safety, misinformation and jobs.
But there is a question that sustainability professionals keep thinking about, and it receives less attention: how much energy, water and land does AI actually use, and who is responsible for measuring it?
That question is now moving onto the regulatory agenda. In Europe, providers of general-purpose AI models are already required to document their models’ energy use, and a new rating scheme for data centres is close to becoming law. In the United States, the debate has turned to who pays for the power that AI needs.
It is important to be precise about where things stand. There are no binding limits on AI energy consumption today. What is emerging is a transparency regime: rules that require energy use to be measured, documented and compared. Standards, labels and possible performance requirements are still being built. But the direction is clear, and it brings AI into the same territory as any other material sustainability issue.
The scale of AI energy consumption is growing quickly. A June 2026 report from the United Nations University Institute for Water, Environment and Health (UNU-INWEH) estimated that global data centres consumed around 448 TWh of electricity in 2025. That is more than most countries, meaning that if data centres were a single country, they would rank just behind France.
The report’s projections for 2030 are significant:
One finding is particularly relevant for organisations that use AI rather than build it. According to the same report, inference — the everyday use of AI models after they are trained — accounts for 80 to 90 per cent of total AI energy use. In other words, AI’s footprint is driven less by a handful of large training runs and more by the millions of queries made every day by companies and individuals.
In Europe, the International Energy Agency expects data centre electricity use to rise from around 68 TWh in 2024 to 114 TWh by 2030, more than 3% of total EU electricity demand. Water is also becoming a concern: Bloomberg has reported that AI data centre water use is increasing in drought-stricken areas.
The AI Act: energy documentation for general-purpose AI models
The EU AI Act is best known for its risk-based rules on high-risk AI systems, many of which have now been postponed. The Digital Omnibus on AI (Regulation (EU) 2026/1744), which entered into force in July 2026, moved most high-risk obligations to December 2027 or later.
The rules for general-purpose AI models were not delayed. Since August 2025, providers of these models have had to prepare technical documentation that includes the known or estimated AI energy consumption of their models. Where actual consumption is unknown, providers may estimate it from the computing resources used. Models already on the market before August 2025 have until August 2027 to comply.
Since 2 August 2026, the EU AI Office has had the power to enforce these obligations. Non-compliance with obligations for general-purpose AI models can lead to fines of up to €15 million or 3% of worldwide turnover, whichever is higher.
The AI Act also looks further ahead. It asks the Commission to work with standards bodies on AI standards for energy efficiency and resource use, encourages voluntary codes of conduct on energy-efficient AI, and requires the Commission to report on progress towards energy-efficient general-purpose AI models from August 2028. Those reports may recommend binding measures.
Measuring AI energy consumption: the next step
It’s important to be able to compare documents. This spring, the Commission consulted developers and deployers of AI on how to measure the AI energy consumption and emissions of models and systems, covering both training and inference. The work is part of a Commission study that will feed into a measurement framework for the AI Act and support the design of a possible AI energy and emissions label.
A label is not yet a legal requirement, but it signals the direction of the policy discussion: from ‘disclose what you use’ to ‘make it comparable for buyers’.
Data centres: from reporting to ratings
The second building block concerns the infrastructure behind AI. On 21 September 2026, the European Commission proposed a common EU rating scheme for data centres with at least 500 kW of installed IT power. Ratings will cover energy and water efficiency, local water stress, waste heat reuse and how facilities interact with the electricity grid.
The proposal is now in a two-month period during which the European Parliament and the Council can object. The first labels for individual data centres are expected in 2027. At this time, the Commission is consulting on minimum performance standards for data centres until 14 December 2026, with a legislative proposal planned for 2027.
This is the first point at which binding performance requirements, not only disclosure, could apply to the infrastructure that powers AI.
Outside Europe, the debate has taken a different form. In the United States, the focus is on electricity prices and grid costs. A draft Senate bill would require large data centres to pay their share of existing transmission costs and the new grid costs they trigger. Local opposition has reportedly led to the cancellation of around $260 billion of planned US data centre projects this year, and Amazon has announced that it will publish the energy and water use of its data centres annually.
At the same time, a bipartisan permitting bill designed to accelerate energy projects for the AI boom has divided opinion, winning support from both fossil fuel and clean power companies, while many environmental groups oppose its narrower environmental reviews.
The common thread is accountability. Whether through regulation, grid pricing or community pressure, AI’s energy consumption is becoming something companies are expected to measure and explain.
The current framework has clear gaps, for example there aren’t any binding limits because neither the AI Act nor the data centre rules have concluded the AI energy consumption or water used by these models. Measuring AI energy consumption is a method still under development. Also, what is documented regarding AI Act’s energy cosumption is only limited to providers of general-purpose AI models, not to the many companies that use AI in their operations. The data collected is still unclear, especially regarding information on water use, hardware and supply chains. What’s next concerning policy is likely the focus on more standardised footprint reporting where carbon, water and land footprints will be treated by investors as material risks.
For most organisations, the immediate impact is not a new AI-specific obligation. It is the connection between AI and the sustainability frameworks they already work with.
The control of AI is often discussed as a question of safety and ethics. It is also becoming a question of resources. Europe has taken the first steps by requiring energy documentation for general-purpose AI models and by building a rating scheme for data centres. Binding limits do not exist yet, but transparency is becoming the default, and transparency tends to be the first step towards accountability.
For sustainability, ESG and reporting professionals, this means AI can no longer be treated as a purely technical topic. Understanding how AI affects energy use, emissions and environmental claims is becoming part of the core skill set. The online course “Sustainability Intelligence: An Introduction to AI in Sustainability/ESG Practices” helps professionals build that understanding, enabling them to bring AI’s footprint into reporting and strategy with confidence.
