In short
There is no official energy label for AI models yet, but the building blocks exist. The AI Energy Score rates open models with one to five stars, MLPerf Power benchmarks hardware systems, and the EU requires energy documentation for general-purpose AI models and is building a rating scheme for data centres. The EU has also tendered for work on measuring energy-efficient AI, and a label is under discussion.
Key points
- The AI Energy Score, launched in February 2025, gives models a one-to-five-star rating within each task, based on measured GPU energy.
- MLPerf Power measures the energy efficiency of whole machine-learning systems, from tiny devices to data-centre racks.
- Under the EU AI Act, general-purpose AI providers must document known or estimated energy consumption, but this goes to regulators, not consumers.
- The EU is building an A–G style rating for data centres, and has tendered for work that could lead to an AI energy and emissions label.
- A label can rank like-for-like options but cannot capture everything: hardware, settings and real-world use all change the result.
Fridges, washing machines and light bulbs carry energy labels that let buyers compare them at a glance. People increasingly ask why AI models do not. The short answer is that labelling AI is harder than labelling a fridge, but work towards it is well under way. This article sets out what exists in late 2026 and what a label can and cannot tell you.
Is there an official energy label for AI?
No, not yet. There are voluntary ratings and benchmarks, legal duties to document energy use in the EU, and European Commission work that could lead to a label. None of these is yet a consumer label of the fridge kind.
What is the AI Energy Score?
A voluntary star rating for AI models, based on measured energy use on standard tasks. It was launched in February 2025 by Hugging Face with partners including Salesforce, Cohere and Carnegie Mellon University.
How it works:
- Models are run on a fixed set of common tasks, using the same type of GPU for every model.
- The GPU energy used for a fixed number of queries is measured.
- Each model gets one to five stars, relative to other models in the same task and size class. Ratings are recalibrated as new models are added.
The first release rated 166 models across 10 tasks. A second version, released in late 2025, added reasoning as a benchmarked task; Hugging Face reported that reasoning models used, on average, around 30 times more energy than models without reasoning or with it switched off, largely because they produce far more output tokens.
The score's strength is that it is measured, not estimated, and that it uses the same conditions for every model. Its limit is coverage: closed commercial models can only be included if their providers choose to take part.
What are MLPerf Power and the ML.ENERGY Leaderboard?
Benchmarks that measure energy under controlled conditions, one for hardware systems, one for open models.
- MLPerf Power, from the MLCommons industry consortium, measures the energy efficiency of complete machine-learning systems running standard MLPerf workloads. Its 2025 paper reports 1,841 measurements from 60 systems submitted by more than 20 organisations, ranging from microwatt devices to megawatt data-centre systems. It helps buyers compare hardware, not models.
- The ML.ENERGY Leaderboard, from the University of Michigan, measures the inference energy of widely used open models on known hardware, and publishes its tools so others can repeat the measurements.
High-performance computing has a precedent. The Green500 list ranks supercomputers by performance per watt on a standard benchmark. It is a crude measure, but because everyone runs the same test, it is comparable, and it has pushed efficiency for years. AI benchmarks may end up the same way: imperfect, but useful because they are shared.
What does EU law require?
Documentation for general-purpose AI models, reporting for data centres, and a framework for future standards.
- AI Act, Annex XI. Providers of general-purpose AI models must document the "known or estimated energy consumption of the model". If it is not known, it may be based on information about the computational resources used. This documentation goes to the EU AI Office and national authorities, not to the public.
- AI Act, Article 40. The Commission's standardisation request covers reporting and documentation processes to improve AI systems' resource performance, including the energy-efficient development of general-purpose models. These standards are still being developed.
- Data centres. Under Delegated Regulation (EU) 2024/1364, data centres with at least 500 kW of IT power must report energy and sustainability indicators each year. In 2026 the Commission proposed the next step, a rating scheme with classes for indicators such as power usage effectiveness (PUE) and water usage effectiveness (WUE); it is under scrutiny and is not yet in force.
Is the EU developing an AI energy label?
It is exploring one. The Commission has run a call for tenders under the AI Act to measure and foster energy-efficient and low-emission AI, which includes the possibility of an AI energy and emissions label, and held a targeted consultation in 2026. Whether a label follows, and whether the AI Act provides a legal basis for one, is not settled. Any first framework is likely to be a guide with wide margins of uncertainty rather than a precise score.
Outside the EU, the picture is mostly voluntary. In the UK and the US there is no AI energy label requirement; organisations rely on voluntary schemes such as the AI Energy Score and on standards such as ISO/IEC 21031:2024, the Software Carbon Intensity specification, which defines a carbon rate per unit of work but is not itself a comparative label.
Why is labelling AI harder than labelling a fridge?
Because the same model's energy use changes with the hardware, the settings and the job. A fridge does the same thing every day. An AI model's energy per task depends on:
- Hardware: the chip generation, how many chips, power and boost settings.
- Software: the inference engine, batching, quantisation, caching.
- The task: prompt length, answer length, whether reasoning mode is on.
- The setting: a lab benchmark versus a busy production data centre with spare capacity held for peaks.
- Quality: a model that uses less energy but gets more answers wrong may cost more overall.
A label therefore has to fix the conditions, which makes it comparable but less realistic. That is the same trade-off car fuel-economy tests make.
What should a buyer ask for meanwhile?
Measured energy per task, under stated conditions, with quality alongside. In practice:
- Ask whether the model has an AI Energy Score or other independent measurement.
- Ask what hardware, region and settings the supplier's figures assume.
- Ask for energy and quality on your task, not a generic benchmark.
- Treat ratios such as "35 times more efficient" with care: they say as much about the baseline as the product.
See sustainable AI procurement and measurement is the bottleneck.
What to do next
- Check the AI Energy Score for open models you are considering, and prefer the most efficient that meets your needs.
- Build energy reporting into AI contracts: see sustainable AI procurement.
- Measure your own workloads: see measuring AI energy.
- Avoid unsupported efficiency claims in your own marketing: see making honest green claims.
- Find rated tools and services in the products directory.
Frequently asked questions
Is there an energy label for AI models?
Not an official one. The AI Energy Score is a voluntary star rating for models on standard tasks. The European Commission has commissioned work on measuring energy-efficient AI that could lead to a label, but none has been adopted.
What is the AI Energy Score?
A project launched in February 2025 by Hugging Face with partners including Salesforce, Cohere and Carnegie Mellon University. It measures the GPU energy models use on standard tasks and gives each a rating from one to five stars, relative to others in the same task.
What is MLPerf Power?
A benchmark from the MLCommons consortium that measures the energy efficiency of machine-learning hardware and software systems running standard MLPerf workloads, from microwatt devices to megawatt data-centre systems.
Sources
- Announcing AI Energy Score ratings, Hugging Face (2025)
- AI Energy Score v2: refreshed leaderboard, now with reasoning, Hugging Face (2025)
- AI Energy Score, OECD.AI catalogue of tools (2025)
- MLPerf Power: benchmarking the energy efficiency of machine learning systems from µWatts to MWatts for sustainable AI, Tschand et al., MLCommons, HPCA 2025 (2025)
- The ML.ENERGY benchmark: toward automated inference energy measurement and optimization, Chung et al., NeurIPS Datasets and Benchmarks (2025)
- Regulation (EU) 2024/1689 (AI Act), Annex XI, European Union (2024)
- Regulation (EU) 2024/1689 (AI Act), Article 40, European Union (2024)
- Artificial Intelligence Act: call for tenders to measure and foster energy-efficient and low-emission AI, European Commission (2025)
- Commission Delegated Regulation (EU) 2024/1364 on the first phase of the establishment of a common Union rating scheme for data centres, European Commission (2024)
- European Commission proposes mandatory sustainability label for data centres, Latham & Watkins (2026)
- Software Carbon Intensity (SCI) specification, Green Software Foundation (2024)
- TOP500 and Green500 lists, TOP500 (2026)
Related
Measurement is the bottleneck for green AI
Why AI energy figures are hard to trust or compare: self-reported numbers, estimates not meters, different boundaries, and what good looks like.
Sustainable AI procurement
What to ask AI suppliers about energy, carbon and water, which evidence to request, and contract clauses that make sustainability commitments enforceable.
Choose the right model for the job
How to right-size AI: when to use small or specialised models, when a frontier model earns its cost, how routing works, and when not to use AI.
Honest green claims about AI
How to avoid greenwashing about AI: what you can and can't claim, rebound effects, honest reporting, and the UK, EU and US rules on environmental claims.