greenai.guide

Home › Green AI glossary

Green AI glossary

In short

This glossary defines the terms people meet when they start reading about the environmental cost of AI, in plain English and in alphabetical order. Each term has its own link, so other pages on the site can point straight to it.

Key points

  • Training builds a model once; inference is each use of it afterwards.
  • Efficiency means doing the same with less; sufficiency means asking whether the task needs doing with AI at all.
  • PUE compares a data centre's total energy with the energy its computers use; 1.0 would be perfect.
  • The rebound effect is when efficiency makes something cheaper and so increases total use.

Green AI brings together computer science, mathematics, engineering, energy and policy, and each field has its own vocabulary. Words such as "model", "efficiency" and even "AI" mean different things to different people. A 2026 UK green AI sandpit of mathematicians, computer scientists and industry found that the terms people most wanted pinned down were the basic ones, which is why this glossary starts from scratch. Terms are in alphabetical order. If a term you need is missing, tell us.

Agentic AI

AI that carries out a task in several steps on its own, such as searching, writing code, running it and fixing errors, rather than giving a single answer. Because each step uses the model again, agentic AI can use many more tokens than a chat, and agents that get stuck in loops can waste a lot. See agentic AI and token waste.

AI

Short for artificial intelligence: computer systems that do tasks we associate with human intelligence, such as recognising images, understanding language or making predictions. In everyday speech in 2026, "AI" usually means large generative models such as chatbots, but the term also covers far smaller systems like spam filters and weather models, so it is worth saying which you mean.

AI literacy

The knowledge and skills to understand, use and judge AI sensibly: what it can and cannot do, how to check its output, and its costs, including its environmental ones.

Automatic differentiation

A technique that lets a computer work out exactly how a calculation's output changes as each input changes. It is how neural networks learn (where it is called backpropagation), and it is also used to optimise scientific simulations.

Benchmark

A standard test used to compare systems fairly, for example how accurate models are on the same set of questions, or how much energy they use on the same task. Benchmarks are only as useful as the tasks and measurements they contain.

Carbon-aware computing

Running computing jobs at times or in places where electricity is cleaner, for example delaying a large training run until the wind is blowing. It suits work that does not need an instant answer.

Carbon intensity

How much carbon dioxide (or equivalent greenhouse gas) is emitted for each unit of electricity, usually in grams of CO₂e per kilowatt-hour. It varies by country and by hour, depending on how much power comes from fossil fuels or renewables.

Data centre

A building full of computers, storage and network equipment, with the power and cooling to keep them running. Most AI chatbots run in data centres, not on the user's device.

Diffusion model

A type of generative AI, widely used for images and video, that learns to turn random noise step by step into a picture. Its many steps help explain why generating images uses much more energy than short text.

Distillation

Training a small "student" model to imitate a large "teacher" model, so that the small one can do a task almost as well at a fraction of the computing cost.

Edge AI

AI that runs on or near the device where the data is, such as a phone, camera or factory sensor, rather than in a distant data centre. It can save network energy and protect privacy, but small devices multiplied by millions add up too. See edge or cloud.

Efficiency

Getting the same useful result for less input: less energy, time, water or hardware per task. Efficiency is essential to green AI but not enough on its own, because of the rebound effect.

Embodied carbon

The emissions from making, transporting and eventually disposing of equipment, as opposed to running it. For AI this includes manufacturing chips, servers and data centre buildings.

Energy label

A simple rating that shows how much energy a product uses, like the A to G labels on fridges in Europe. Several groups are developing equivalents for AI models, such as Hugging Face's AI Energy Score, and the EU AI Act requires providers of general-purpose AI models to document their known or estimated energy use. See energy labels for AI.

Fine-tuning

Taking a model that has already been trained and training it a little more on specialist data, so it performs better on a particular task. It costs far less than training from scratch.

FLOPs

Floating-point operations: the basic arithmetic steps (additions and multiplications of decimal numbers) a computer does. Counting them is a common way to estimate how much computing a model needs. FLOPS with a capital S usually means operations per second.

FPGA

A field-programmable gate array: a chip whose circuits can be reconfigured after it is made. FPGAs can be very efficient for a specific task and can be reprogrammed as models change, but they are harder to program than GPUs.

Generative AI

AI that creates new content, such as text, images, audio, video or code, in response to a request. Chatbots and image generators are generative AI.

GPU

A graphics processing unit: a chip originally designed for computer graphics that can do many calculations at once. GPUs are the main hardware for training and running modern AI.

Green AI

AI developed and used in a way that keeps its energy, carbon, water and material footprint as low as possible, and used where it adds real value. It covers both making AI more efficient and using it only where it is the right tool. See what is green AI?.

Greenwashing

Making misleading claims that a product or organisation is better for the environment than it is. In the EU, Directive (EU) 2024/825 bans generic environmental claims to consumers that cannot be proved. See honest green claims.

Inference

Using a trained model to produce an output, such as answering a question or generating an image. Every prompt is an inference, so for popular tools inference energy keeps adding up for as long as the model is in use.

Jevons paradox

The observation, made by the economist William Stanley Jevons in his 1865 book The Coal Question, that more efficient steam engines led Britain to burn more coal, not less. It is the extreme case of the rebound effect, where efficiency increases total use.

KV cache

Short for key–value cache. When a language model writes an answer one token at a time, it stores intermediate results (the "keys" and "values" of its attention calculation) so it does not have to recompute them at every step. It saves a lot of computing but uses memory.

Large language model

A generative AI model trained on huge amounts of text to predict and produce language, often shortened to LLM. Most chatbots are built on one.

Latency

The delay between asking for something and getting the answer. Some green AI techniques trade a little latency for lower energy, for example batching requests or waiting for cleaner power.

McNamara fallacy

The mistake of treating what can easily be measured as all that matters, and ignoring what cannot. In green AI it is a warning against judging systems only by convenient numbers while overlooking harms, such as water use or social effects, that are harder to count.

Mixed precision

Doing some calculations with fewer digits of precision than others, where the result is not harmed. Lower precision uses less memory and energy.

Mixture of experts

A model design made of many specialist sub-networks ("experts"), where a router sends each input to only a few of them. It uses only part of the model for each answer, which saves computing, although all of it is needed in memory and during training.

Model

In AI, a model is the trained system that turns inputs into outputs: the learned set of numbers (parameters) and the structure that uses them. In science and engineering, "model" can also mean a mathematical description of a system, such as a climate model, so it helps to say which you mean.

Model merging (souping)

Combining two or more models, often by averaging their parameters, to get one model with the strengths of each without training a new one from scratch. "Model soups" was the name used by Wortsman and colleagues in 2022.

Multi-agent system

A set of AI agents that work together or pass tasks between them. They can be powerful, but if they get stuck talking to each other they can burn through large numbers of tokens with no useful result.

Open-weight model

A model whose trained parameters (weights) are published so anyone can download and run it, even if its training data and code are not. Open-weight models can be run locally and adapted, and their energy use can be measured independently.

Orchestration

Coordinating several models, tools or pre-built components to complete a task, rather than having one large model generate everything. See orchestration over generation.

Parameters

The numbers inside a model that are adjusted during training and together store what it has learned. Large language models have billions of them; Meta's Llama 3.1 family, for example, comes in 8, 70 and 405 billion parameter sizes. More parameters usually means more computing per answer.

Physics-informed neural network

A neural network trained to respect known laws of physics, such as conservation of energy, as well as fitting data. Building in that knowledge can mean less data, fewer parameters and less training.

Prompt

The instruction or question a person gives to a generative AI model. Clear, specific prompts usually get a good answer first time, which saves both time and energy.

Pruning

Removing parts of a trained model, such as individual weights or whole connections, that contribute little to its output. A pruned model is smaller and cheaper to run, often with little loss of accuracy.

PUE

Power usage effectiveness: a data centre's total energy use divided by the energy used by its computing equipment. A PUE of 1.0 would mean no energy goes on cooling or power losses; the lower, the better. See hosting and data centres for typical values and their sources.

Quantisation

Storing a model's numbers with fewer digits, for example 8-bit or 4-bit numbers instead of 16- or 32-bit ones. This shrinks the model and cuts the energy of running it, usually with only a small loss in quality.

Rebound effect

When an efficiency gain makes something cheaper or easier, so people use more of it and some or all of the saving is lost. If use grows so much that total consumption rises, it is the Jevons paradox. See efficiency is not sufficiency.

Routing

Sending each request to the most suitable model, for example a small, cheap model for easy questions and a large one only for hard ones. Getting the threshold right matters: a model that is too small may fail and force a second, larger attempt.

Scope 1, 2 and 3 emissions

The three categories of the Greenhouse Gas Protocol. Scope 1 is direct emissions from sources an organisation owns or controls; scope 2 is emissions from the electricity, heat or cooling it buys; scope 3 is all other indirect emissions in its value chain, such as the manufacture of the hardware it buys.

Small language model

A language model with relatively few parameters, designed to run cheaply, sometimes on a laptop or phone. For narrow tasks a well-chosen small model can match a much larger one. See choosing the right model.

Software carbon intensity

A method from the Green Software Foundation, also published as an ISO standard, for calculating the carbon emitted per unit of work a piece of software does, such as per user or per request.

Sovereign AI

AI capability, such as models, data and computing infrastructure, that a country or region controls itself rather than relying on foreign providers. It raises its own sustainability questions; see sovereign AI and sustainability.

Sparsity

When many of the numbers in a model, or many of its activations for a given input, are zero. Computing can skip the zeros, so a sparse model can be much cheaper to run if the hardware and software support it.

Spiking neural network

A neural network that, like the brain, passes information as brief pulses (spikes) and stays mostly idle between them. On suitable hardware they promise very low energy use, especially at the edge.

Sufficiency

Using only as much of a resource as a need really requires, including asking whether AI is needed at all. It is the partner to efficiency: efficiency asks "how can we do this with less?", sufficiency asks "do we need to do this?".

Surrogate model

A fast, approximate model trained to stand in for a slow, expensive calculation, such as part of a climate or engineering simulation. A good surrogate can save a lot of energy, provided its accuracy is checked.

Sustainable AI

A broad term for AI that can be developed and used without exhausting the resources it depends on, often including social and economic effects as well as environmental ones. It overlaps with green AI and the terms are often used interchangeably.

Token

The chunk of text a language model reads and writes, often part of a word. OpenAI's rule of thumb for English is that one token is about four characters, or about three-quarters of a word; other languages often need more tokens for the same meaning. More tokens means more computing.

TPU

A tensor processing unit: a chip designed by Google specifically for the calculations used in machine learning.

Training

Building a model by showing it large amounts of data and repeatedly adjusting its parameters to reduce its errors. Training a large model can take months on thousands of chips; Meta reports 39.3 million GPU hours to train its Llama 3.1 family.

Uncertainty quantification

Working out how confident we can be in a model's output, for example giving a range rather than a single number. It helps decide when a cheaper model is good enough and when a result needs checking.

WUE

Water usage effectiveness: the water a data centre uses on site, in litres, divided by the energy used by its computing equipment, in kilowatt-hours. Lower is better, though a cooling design that saves water may use more electricity, and the reverse.

Frequently asked questions

What is the difference between green AI and sustainable AI?

People use the terms in overlapping ways. Green AI usually focuses on cutting the energy, carbon, water and materials AI uses; sustainable AI is often wider, also covering social and economic effects and whether resources can be renewed as fast as they are used.

What is a token in AI?

A token is the chunk of text a language model reads or writes, often part of a word. OpenAI's rule of thumb for English is about four characters, or three-quarters of a word, per token.

What is the difference between training and inference?

Training is building a model by learning from data, usually once and at great cost. Inference is using the finished model to answer a question or make an image, which happens every time someone uses it.

Sources

  1. What are tokens and how to count them?, OpenAI Help Center (2025)
  2. Uptime Institute global data center survey 2025, Uptime Institute (2025)
  3. Data center metering and resource guide, Lawrence Berkeley National Laboratory (2017)
  4. GHG Protocol corporate standard, Greenhouse Gas Protocol (2004)
  5. W. Stanley Jevons, The Coal Question, 1865, Yale University, Energy History (1865)
  6. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time, Wortsman et al. (arXiv:2203.05482) (2022)
  7. Annex XI: technical documentation for providers of general-purpose AI models, European Commission, AI Act Service Desk (2024)
  8. Directive (EU) 2024/825 on empowering consumers for the green transition, European Union (2024)
  9. Software Carbon Intensity specification, Green Software Foundation (2024)
  10. ISO/IEC 21031:2024 Software carbon intensity (SCI) specification, International Organization for Standardization (2024)
  11. Llama 3.1 8B model card, Meta (Hugging Face) (2024)

Join us

Help cut AI's impact on the planet

Get new guides, funding calls and directory listings by email, about once a month. Or add your product, course, PhD or research to the directories.

No tracking. Unsubscribe in one click. Other ways to join · Privacy