Green AI

Green AI is a design philosophy and research field concerned with reducing the energy consumption and environmental impact of artificial intelligence while maintaining useful performance. It emerged in response to the rapidly growing computational demands of modern machine learning, particularly the training and serving of large models, whose electricity use, carbon emissions, and water consumption have become significant. Green AI is often contrasted with "Red AI", a term describing research that pursues state-of-the-art accuracy by deploying ever-larger amounts of compute without regard to efficiency or cost.

The Environmental Cost of AI

The scale of AI's resource use is substantial. Frequently cited analyses estimate that training a single large transformer model can emit hundreds of metric tons of carbon dioxide — comparable to the lifetime emissions of several cars — with figures varying widely depending on model size, hardware, and the carbon intensity of the electricity grid used. Training GPT-3, for example, has been estimated to have released on the order of 500 metric tons of carbon dioxide, whereas much smaller models emit only a fraction of that.

The footprint is not limited to training. Inference — serving a model to users — can dominate total lifetime energy use for popular deployed systems, since each query consumes power and a successful model may serve billions of requests. Data centres housing AI hardware also consume water for cooling, raising local resource concerns in water-stressed regions.

Techniques for Reducing AI's Footprint

Green AI draws on a portfolio of methods spanning the model lifecycle. Model compression techniques — including pruning (removing redundant weights), knowledge distillation (training a small model to mimic a larger one), and quantisation (representing weights with fewer bits) — shrink models so they require less energy for both training and inference, often with minimal accuracy loss. Efficient architectures and approaches such as mixture-of-experts activate only part of a model per query, reducing computation. Mixed-precision training and improved optimisation reduce the energy of training runs, while neural architecture search can be directed to find energy-efficient designs.

Beyond the model itself, operational choices matter greatly. Running workloads in data centres powered by renewable energy, scheduling training when and where the grid is cleanest (carbon-aware computing), improving data centre power usage effectiveness, and adopting liquid cooling all reduce the real-world footprint of a given amount of computation. Edge AI — running models on local devices — can also lower energy by avoiding constant data transmission to the cloud.

Measurement and Reporting

A foundational principle of Green AI is making efficiency a reported, comparable metric rather than an afterthought. Researchers advocate publishing the computational cost, energy use, and estimated emissions of experiments alongside accuracy, so that the community can weigh performance gains against their resource cost. Tools exist to estimate the carbon emissions of training runs based on hardware, runtime, and grid carbon intensity. Treating efficiency as a first-class metric encourages models that deliver strong performance per unit of energy rather than maximum performance at any cost.

Significance

As AI deployment scales globally, the aggregate energy demand of data centres has become a material factor in national energy planning and corporate sustainability commitments. Green AI connects technical research with these broader concerns, offering a path to continue advancing capability while limiting environmental harm. It also has an equity dimension: efficient models are cheaper to train and run, lowering barriers for smaller organisations and countries that lack access to vast compute resources.

References

1. Schwartz, R., Dodge, J., Smith, N. A., Etzioni, O. (2020). Green AI. Communications of the ACM.
2. Strubell, E., Ganesh, A., McCallum, A. (2019). Energy and Policy Considerations for Deep Learning in NLP. ACL.
3. Springer Machine Learning. (2025). An efficient model training framework for green AI. link.springer.com.
4. ScienceDirect. (2026). Eco-conscious green AI: Reviewing sustainability approaches to minimize environmental impact. sciencedirect.com.