In short
Researchers can make AI greener in two ways: by developing methods that need less computation, and by measuring and reporting the energy their own work uses. Pruning, quantisation, distillation, better numerical methods and hybrid physics-ML models cut the cost of AI, while tools such as CodeCarbon and the Green Algorithms calculator make the cost visible in papers.
Key points
- Patterson and colleagues found that the choice of model, data centre and processor can reduce the carbon footprint of training by up to around 100 to 1,000 times.
- Inference, not only training, deserves research attention: it runs every time a model is used.
- Report energy, hardware, run time and location in every paper, using a free tool such as CodeCarbon or the Green Algorithms calculator.
- Mathematics contributes beyond algorithms: to metrics, uncertainty, benchmarks and modelling how AI use spreads.
- Set a compute budget for a project, including failed runs and hyperparameter searches, not just the final model.
Why does green AI matter for research?
Because research sets the methods everyone else copies, and because research itself uses a lot of computing. A technique that halves the cost of inference, once adopted, saves energy every time any model using it runs.
The case was made early. Strubell, Ganesh and McCallum's 2019 paper quantified the financial and environmental cost of training several language models and proposed ways to cut it. Patterson and colleagues later found that the choice of model architecture, data centre and processor can reduce the carbon footprint of training by up to around 100 to 1,000 times. Luccioni, Jernite and Strubell showed that running multi-purpose generative models can be orders of magnitude more expensive than task-specific models for the same task.
A 2026 UK green AI sandpit of mathematicians, computer scientists and industry framed the research agenda in three stages, each open to different techniques: training from scratch, fine-tuning or adaptation, and deployment, whether in the cloud or on a low-energy device. It also noted that much existing work sits in project deliverables and little-known code repositories, so a first step for any project is to find out what has already been done.
Which methods make models more efficient?
The main families are removing what the model does not need, using fewer bits, learning from less data, and building known structure into the model. Most can be combined.
Pruning and sparsity
Pruning removes weights that contribute little to the output. Sparsity means most values in a matrix or activation are zero, so computation can skip them. Large networks often have many more parameters than the final task needs, because over-parameterisation helps optimisation, so much can be removed after training.
Open questions include: how to set sparsity targets before training rather than by trial and error, how to keep dynamic sparsity (where the zeros depend on the input) efficient on real hardware, and why sparsity is still under-used given architectural constraints in mainstream chips. Theoretical gains often exceed measured speed-ups, so empirical results on real hardware matter.
Quantisation and mixed precision
Quantisation stores and computes with fewer bits, for example 8-bit integers instead of 32-bit floating point. Mixed precision uses lower precision where it is safe and higher precision where it is needed. UNESCO and UCL report that compression such as quantisation can save up to 44% of energy while keeping accuracy. Numerical analysts are well placed to say where reduced precision is safe and how errors propagate.
Distillation and smaller models
Distillation trains a small "student" model to mimic a large "teacher". The student is cheaper to run on every request. Related work on parameter-efficient fine-tuning, which adapts a small number of extra parameters rather than the whole model, and on continual learning avoids retraining from scratch.
Reusing models
Instead of training new models, combine existing ones. Wortsman and colleagues showed that averaging the weights of several fine-tuned models ("model soups") can improve accuracy without any extra inference cost. Merging and reusing models is a promising area for efficiency research.
Data efficiency
Training on less, better data cuts training energy directly. Techniques include removing duplicate and low-value examples, active learning (choosing the most informative examples to label), and architectures that generalise from few examples.
Numerical methods and data movement
Much of AI is linear algebra, and decades of numerical analysis apply. Randomised numerical linear algebra and sketching (compressing a matrix while keeping its important properties) can reduce both arithmetic and data movement, which on many devices costs more energy than the arithmetic itself. Approximate attention, which computes only the parts of the attention matrix that matter, is an active example in large language models. Automatic differentiation, the general technique behind backpropagation, links AI to existing scientific codes.
Hybrid physics-ML
Rather than replacing a whole simulation with a neural network, replace only the part that suits it. A common pattern is a learned surrogate for the most energy-intensive component of a climate or engineering model, keeping the rest of the physics. Physics-informed models build known laws into the loss function or architecture, which can need far fewer parameters and less data. Other structural tools include convexity, unrolled algorithms (turning the steps of a classical solver into network layers) and symmetry, as in geometric deep learning.
The sandpit also raised less explored ideas:
- adding an energy or "carbon" term to the training objective, for example penalising active neurons or expensive experts in a mixture-of-experts model;
- error correction during long computations, checking at intervals so a run does not have to be repeated; and
- detecting when multi-agent AI systems are stuck in loops and wasting tokens.
See efficient training and inference and maths for efficient AI.
How should I measure and report energy in papers?
Measure as you go, report a standard set of facts, and include the runs that did not make the paper. Without this, efficiency claims cannot be checked or compared.
Free tools:
- CodeCarbon, an open-source Python package, estimates emissions from the power drawn by CPU, GPU and memory and the carbon intensity of the local grid. Add it to training and evaluation scripts.
- The Green Algorithms calculator, from Lannelongue, Grealey and Inouye, estimates the footprint of a computation from hardware, run time, memory and location. It suits HPC jobs and work done before you started logging.
- Hardware counters, such as RAPL on many CPUs and the GPU vendors' power interfaces, give direct readings that these tools build on.
What to report:
| Item | Example |
|---|---|
| Hardware | GPU model and count, CPU, memory |
| Run time | GPU-hours for final runs and for all tuning and failed runs |
| Energy | kWh, measured or estimated, with the tool and version |
| Location | Country or grid region, and the carbon intensity used |
| Emissions | kg CO₂e |
| Inference | Energy per 1,000 requests at a stated batch size and sequence length |
| Boundary | Whether cooling and data-centre overheads (PUE) are included |
Be clear about uncertainty. Estimates from power models can differ from metered values, and provider figures, such as Google's reported 0.24 Wh for a median Gemini text prompt, use their own methods and boundaries. Our guide to measuring AI energy and the article measurement is the bottleneck go further.
How do I set a compute budget?
Decide in advance how much computing a project may use, and track it like money. A budget makes trade-offs visible and stops endless tuning.
- Estimate GPU-hours for the main experiments, then add a margin for exploration.
- Use small-scale pilots to rule out ideas before scaling up.
- Prefer efficient searches for hyperparameters, such as early stopping of poor runs, over large grid searches.
- Reuse checkpoints and pretrained models where they fit.
- Schedule flexible jobs for times or regions where the grid is cleaner.
- Report results against compute, for example accuracy per kWh, not only best accuracy.
The GREENER principles from Lannelongue and colleagues (Governance, Responsibility, Estimation, Energy and embodied impacts, New collaborations, Education and Research) give a framework for groups and institutions.
How does efficiency affect reproducibility?
Reporting energy and compute makes work easier to reproduce, and efficient methods make it cheaper to reproduce. A result that needs thousands of GPU-hours can be checked by few groups.
- Publish code, configurations and seeds, and say how many runs a result is based on.
- Report variance across runs. A single accuracy figure from one run of a non-deterministic process can mislead.
- Release smaller or distilled versions of models where possible so others can build on them.
- Share negative results on efficiency methods, which save others from repeating them.
Where can mathematics contribute?
Mathematics helps beyond faster algorithms: in deciding what to measure, how much to trust a model, and how AI use spreads through society. The sandpit stressed that mathematicians have to show, with tools and evidence, what they can improve.
- Theory of efficiency: random matrix theory for pruning and sparsity, approximation theory for small models, optimisation theory for faster training.
- Numerical analysis: precision, stability, error bounds and solvers inside AI pipelines.
- Uncertainty quantification: how far a surrogate can be trusted, especially outside its training data, which decides whether it can replace an expensive physical model.
- Metrics and benchmarks: agreed, comparable measures of energy and quality, and quality metrics with efficiency built in.
- Scenario and systems modelling: how efficiency gains change total use (the rebound effect), game theory of why organisations feel pushed into AI use, and the balance between edge and central computing.
- Choosing the method: frameworks for when a problem needs AI at all and when a classical algorithm already solves it well.
How do I find funding and collaborators?
Combine an efficiency research question with a real application partner who cares about cost or speed. Funders and industry respond better to energy savings that also bring faster, cheaper or more reliable results.
- National research councils such as EPSRC in the UK, the NSF in the US and their equivalents fund methods research; check calls on AI, digital research infrastructure and net zero.
- European programmes such as Horizon Europe and Eureka clusters such as ITEA fund collaborative projects with industry. GreenCode, an ITEA project, is one example in green AI and software energy.
- Short industry projects, placements and co-funded studentships are low-cost ways to start a collaboration.
- Research software engineers and HPC centres often have green-computing expertise and training.
Interdisciplinary teams work best: put hardware, software, numerical and domain experts in the same room. The research directory lists groups and projects in green AI.
What PhD routes are there?
There are PhD projects in efficient machine learning, numerical analysis, sustainable and low-carbon computing, HPC and hardware-software co-design. Many are not labelled "green AI", so search for those topics as well.
Routes include:
- doctoral training centres and programmes, such as UKRI Centres for Doctoral Training in the UK;
- doctoral networks under the EU's Marie Skłodowska-Curie Actions;
- industry co-funded or sponsored studentships; and
- projects attached to larger research grants.
See the PhD directory for current opportunities and the courses directory for short courses on efficient AI, green software and research computing.
What to do next
- Add CodeCarbon or similar logging to your next experiment, or estimate your last project with the Green Algorithms calculator.
- Add an energy and compute table to your next paper, including tuning and failed runs.
- Set a compute budget for your current project and review it monthly.
- Try one efficiency method, such as quantisation or distillation, on a model you already use, and report the trade-off.
- Check what exists before building: search the research directory and recent deliverables.
- Find a partner with a costly AI or simulation workload and propose a joint project.
- Look for training in the courses directory, and share the PhD directory with students.
Frequently asked questions
How do I report the energy use of my experiments?
Log energy with a tool such as CodeCarbon while experiments run, or estimate it afterwards with the Green Algorithms calculator from hardware, run time and location. Report kWh, CO₂e, hardware, hours and grid region, and include failed and tuning runs.
Does quantisation hurt accuracy?
Often only slightly, but it depends on the model and task. UNESCO and UCL report energy savings of up to 44% from compression such as quantisation while keeping accuracy; check on your own benchmarks.
Is green AI a real research field or a side issue?
It is a growing field spanning machine learning, numerical analysis, hardware and computing systems. Studies of the energy of training and inference date back to at least 2019, and the EU AI Act now asks providers of general-purpose models to document energy use.
Can I do a PhD in green AI?
Yes. Look for projects in efficient machine learning, numerical methods, sustainable computing or HPC. Our PhD directory lists current opportunities.
Sources
- Carbon emissions and large neural network training, Patterson et al. (2021)
- Energy and policy considerations for deep learning in NLP, Strubell, Ganesh and McCallum, ACL (2019)
- Power Hungry Processing: Watts Driving the Cost of AI Deployment?, Luccioni, Jernite and Strubell, ACM FAccT (2024)
- Green Algorithms: quantifying the carbon footprint of computation, Lannelongue, Grealey and Inouye, Advanced Science (2021)
- GREENER principles for environmentally sustainable computational science, Lannelongue et al., Nature Computational Science (2023)
- CodeCarbon, CodeCarbon (2026)
- Smarter, smaller, stronger: resource-efficient generative AI and the future of digital transformation, UNESCO and University College London (2025)
- Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time, Wortsman et al., ICML (2022)
- Measuring the environmental impact of delivering AI at Google scale, Elsworth et al., Google (2025)
- AI Act, Annex XI: technical documentation for providers of general-purpose AI models, European Commission AI Act Service Desk (2024)
- Green AI and energy-aware computing, GreenCode (ITEA) (2026)
Related
Efficient AI training and inference
Techniques for developers and researchers to cut AI energy: pruning, sparsity, quantisation, distillation, KV cache, batching, mixed precision and scheduling.
How to measure the energy and carbon of AI
A practical guide to metering or estimating the energy and carbon of AI work: tools, system boundaries, per-task metrics and honest uncertainty.
The maths that makes AI more efficient
Pruning, sparsity, sketching, random matrix theory and physics-informed models: how mathematics cuts the computation, memory and energy AI needs.
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.