- The EU AI Act requires General-Purpose AI model providers to report estimated energy consumption starting August 2025, with the EU AI Office able to demand technical documentation without prior notice.
- No standardised methodology currently exists for measuring AI energy use across training, inference and hardware manufacturing, making provider-to-provider comparison impossible.
- The European Commission’s May 2026 consultation on standardised measurement and labelling means GPAI providers operating in the EU will need compliance tracking built into development processes before the final format is settled.
The EU is moving to put mandatory energy labels on AI models, similar to the efficiency ratings on fridges and washing machines. The EU AI Act’s requirement for General-Purpose AI providers to report energy consumption, combined with the European Commission‘s May 2026 consultation on standardised measurement, is the most detailed government-led attempt yet to put numbers on AI’s environmental costs. Whether it produces comparable, enforceable data is still an open question.
Europe’s mandate for AI energy transparency
The EU AI Act entered into force on August 1, 2024. Among its provisions is a requirement for providers of General-Purpose AI (GPAI) models to document and report their models’ known or estimated energy consumption, with that obligation applying from August 2025. The EU AI Office, established under the Act, can demand technical documentation on energy consumption from GPAI providers without prior notice. The framing is deliberate: the goal is to move beyond the current landscape of inconsistent, incomparable self-reporting and replace it with a government-enforced disclosure framework. The Commission is also developing standardised energy and emissions labels for AI models, drawing directly on the appliance-labelling model that already governs consumer hardware across Europe.
The scale of AI’s energy costs
Pinning down what AI actually costs in energy terms is harder than it sounds. Consumption varies substantially depending on whether you are counting training runs, inference at scale, hardware manufacturing or some combination of all three. Current industry practice offers no consistent answer on which of those to include, which means published figures from different providers are rarely measuring the same thing.
EU data centre electricity consumption is projected to rise significantly, driven by AI computation and broader digitalisation, though the precise trajectory depends on which cost components end up inside the measurement boundary. The Commission’s May 2026 consultation explicitly sought input on what data is actually accessible across the full lifecycle, a signal that the Commission regards the measurement problem as unsolved, not merely unstandardised.
Why self-regulation falls short
Voluntary sustainability reporting has a structural problem: the companies producing it also set the metrics, select the data and publish the results with no external audit requirement. There are no standardised benchmarks for AI energy disclosure, which makes comparison across providers essentially impossible. The competitive dynamics of the AI sector give developers good reason to prioritise deployment speed, and limited reason to invest in environmental accounting that could be used against them by regulators or rivals.
Investor patience with AI infrastructure spending is also showing signs of strain. Scrutiny over whether large cloud provider investments are generating adequate returns has intensified, according to recent financial commentary, raising questions about fiscal discipline alongside environmental ones. That external pressure adds some incentive for transparency, but it is not a substitute for enforceable standards.
Emerging external oversight bodies
The EU AI Office is the most direct oversight mechanism in play, with authority to demand technical documentation from GPAI providers on short notice. Outside Europe, the picture is more fragmented.
The U.S. Government Accountability Office published a report in May 2026 identifying weak oversight of federal IT spending, including AI-related expenditure, and flagged risks of waste and duplication. Specific problems cited included agencies repurchasing software licences for cloud use and incurring unexpected fees for AI services. The mandate is narrower than the EU’s, focused on federal procurement rather than industry-wide environmental accountability.
The OECD‘s Expert Group on AI Compute and Climate has been working since late 2020 on frameworks for measuring and benchmarking domestic AI computing supply. An October 2025 working paper presented a methodology for tracking the global distribution of public cloud compute for AI, aimed at giving governments better data for national AI strategy decisions. The OECD’s output is advisory rather than binding.
What accountability in practice requires
For GPAI providers operating in the EU, voluntary sustainability reporting is no longer a sufficient response. Compliance will require tracking systems built into development processes from an early stage, capable of producing documentation that meets whatever standardised format the Commission finalises. How demanding that format will be is still being worked out, and enforcement without a settled measurement standard will be difficult. The gap between the Act’s current requirements and a functional labelling system remains substantial. For more on AI agents and automation tools, visit our AI Agents section.
