Learn AI /Data and Infrastructure /Energy and water use of AI

Energy and water use of AI

Starter
Data and Infrastructure
Policy officers
Managers & front offices

Also called: environmental footprint, carbon footprint of AI, sustainability.

Detail

In five seconds

What it is

Energy and water use of AI means the electricity and water consumed in building and running AI systems; emissions are a separate impact.

In plain words

AI needs real machines in real buildings. Those machines use electricity, create heat, and may need water for cooling. The wider environmental impact depends on how much processing is done and where and how the system is run.

An everyday example

Imagine a team in Arusha, Bogotá and Vienna helping group survey forms that raise the same issue. If they print every form, keep the lights on late, and run extra fans because the room gets hot, the job uses more electricity and materials. If they also bring in water for cooling because the room overheats, resource use rises further. The useful output is the grouped list of issues, but the office still used power, water and supplies to produce it. A better process, such as sorting only the needed pages and using a cooler room, can reduce the impact without changing the goal.

Where the story stops working

An office team is slow and visible, while AI runs in data centres and networks that users do not see directly. Also, AI impacts depend heavily on hardware efficiency, electricity source, cooling design and model architecture, which have no close match in the paper-sorting story. The analogy helps show that useful outputs still require physical resources, but it does not measure those impacts accurately.

Remember thisAI uses electricity across training, inference and supporting infrastructure, and may use water for cooling. Those resource uses are related to, but not the same as, carbon emissions.

This term is contested

The broad idea is clear, but exact measurement methods and accounting boundaries are genuinely debated. Different assessments may include or exclude training, inference, cooling overhead, network traffic, hardware manufacture and end-of-life impacts, so published figures are not always directly comparable.

Understand it

Why it matters

This affects budgets, procurement, risk and credibility. Choosing a model or hosting option without asking about energy, water and emissions can create avoidable cost and environmental impact. For leaders and policy staff, the key question is not 'AI or no AI' but 'what level of capability is needed, under what safeguards, and at what operational trade-off?' Clearer terms help organisations compare options, ask vendors better questions and avoid treating carbon alone as the whole story.

From AI request to environmental footprint

After viewing this, a reader should be able to explain how an AI task leads to electricity use, heat, cooling needs, and broader environmental impacts, and why carbon footprint is only one part of the picture.

01
AI task
A model is trained or asked to generate an output. More tasks, larger models or repeated retries usually mean more processing work.
Draw a rounded rectangle labelled 'AI task' with small icons for training and answering a prompt.
02
Hardware runs
Chips and servers carry out calculations. This stage uses electricity and depends on hardware efficiency.
Draw a server-and-chip box connected by an arrow from 'AI task'. Add a small meter icon labelled 'electricity'.
03
Heat builds up
Computing creates heat. As workload rises, cooling becomes more important to keep equipment working safely.
Draw a box with a thermometer icon and red heat waves, linked from the hardware box.
04
Cooling and operations
Cooling systems, buildings and networking add more resource use. Some sites use significant water for cooling, while others use less.
Draw a box with a fan, pipe and network icon. Add labels 'cooling', 'water', and 'network'.
05
Environmental footprint
The result can include energy use, water use, emissions and hardware impacts. Carbon footprint is one part of this wider picture.
Draw a final box with four small labels: 'energy', 'water', 'emissions', 'materials/e-waste'. Use a bracket to show emissions are a subset of environmental footprint.

A left-to-right flow with five boxes. Start with an AI task, then computing hardware, then heat and cooling, then resource use, then broader impacts. Read the arrows as cause and effect: each stage creates the conditions for the next, and different design choices can reduce or increase the final footprint.

Hovering or tapping a box highlights the next box to show the causal chain; a toggle switches the final label between 'narrow view: carbon only' and 'broader view: environmental footprint'.

Describe this without the diagram

The diagram is a five-step flow. First, an AI task starts the process, such as training a model or answering a prompt. Second, hardware like chips and servers performs calculations and uses electricity. Third, those calculations create heat. Fourth, cooling systems, networking and other operations use additional resources, and some facilities also use water for cooling. Fifth, the combined result is the environmental footprint: energy use, water use, greenhouse gas emissions, and hardware-related impacts such as material use and electronic waste. A note explains that carbon footprint is only one part of the broader environmental footprint.

With reduced motion: A static left-to-right flow chart with five labelled boxes and arrows: AI task, hardware runs, heat builds up, cooling and operations, environmental footprint. A note under the final box says 'carbon footprint is one part of environmental footprint'.

AI systems rely on physical infrastructure: chips, servers, data centres, networks and cooling equipment. Training a model and using it later both require electricity. Some facilities also use water, often for cooling. That means AI has resource use and environmental effects beyond the screen.

These related terms are not identical. Energy and water use of AI refers to direct resource consumption. Carbon footprint of AI refers to greenhouse gas emissions linked to that energy use and the wider lifecycle. Environmental footprint is broader still: it can include emissions, water, materials, electronic waste and land use. Sustainability asks whether the system can meet its purpose while keeping those impacts, costs and trade-offs manageable over time.

What goes in? Requests from users, data moved across networks, and computing work on processors such as GPUs. What happens? Machines perform calculations, generate heat, and cooling systems remove that heat. What comes out? Useful results, but also electricity demand, possible water consumption, emissions that depend on the local power mix, and hardware wear over time.

The amount varies a lot. A small model used efficiently may have a modest footprint. A larger model, frequent use, repeated retries, or hosting in a region with carbon-intensive electricity can raise it. This is why people often compare model size, hardware type, deployment choice and how many outputs are really needed.

These terms are often confused with compute. Compute means the amount of processing work. It helps explain resource use, but it is not the same thing. Two systems can use similar compute yet have different emissions or water use because location, hardware efficiency, cooling method and electricity source differ.

How “Sorting paper forms in an office” maps onto the real thing

In the story	In the AI system
Staff sorting and re-sorting forms	Processors doing repeated calculations for training or inference
Lights, fans and a hot room	Electricity use and heat generated by computing equipment
Water brought in to cool the room	Water used directly or indirectly in cooling data-centre operations
Grouped list of issues	The AI system's output or service delivered to users

A worked example

An information management officer at a regional hub compares two ways to group 12,000 screened, non-sensitive entries into themes for a progress report.

Starting from: Option A sends all 12,000 entries to a large hosted model. Option B removes duplicates and out-of-scope entries, then sends 8,000 entries to a smaller approved model.

The officer sets the same requirement for both options: 10 issue groups, short summaries and an accuracy check against a human-coded sample.
The comparison records the number of entries, model runs and retries, rather than treating every digital option as having the same resource use.
Available provider information is checked for electricity use, storage and networking overheads, cooling arrangements and any disclosed allocation for model training.
If both options meet the quality requirement, the officer selects the smaller workflow and records where provider data are insufficient for a full comparison.

Result: Option B sends 4,000 fewer entries for model processing and requires fewer repeat runs, so its expected inference workload is lower. The office does not claim a lower total energy or water footprint without comparable operational data, or a lower carbon footprint without emissions data and a stated accounting boundary.

In a United Nations workplace

An evaluation officer is coding screened, open-ended answers for a needs assessment annex before the monitoring report goes for clearance at close of business. The office is deciding whether to use spreadsheet rules, a smaller approved model or a larger hosted model.

Even this bounded task uses electricity for inference, storage, networking and facility overheads, while some hosting facilities also use water for cooling. Removing duplicates, limiting repeat runs and choosing a model proportionate to the task can reduce avoidable resource use.

Before you try thisNeeds assessment responses may contain personal or otherwise sensitive information. Raw responses must not be uploaded to a tool that has not been approved for that material. Minimise the data first, check for re-identification risk, follow the office's data protection guidance and retain human review of the coding.

Where you may hear it

At an ICT governance board reviewing whether a proposed model is proportionate to the task
During a vendor demo in a procurement meeting, when hosting and cooling arrangements are queried
In an evaluation office planning repeated AI-assisted coding runs
During clearance of a sustainability report that distinguishes electricity, water and emissions

Avoid confusion

Often confused with

Common misconceptions

Common misconception: Carbon footprint, environmental footprint and sustainability all mean the same thing.
Correction: They overlap, but they are different scopes. Carbon footprint focuses on greenhouse gas emissions. Environmental footprint is broader and may include water, materials and electronic waste. Sustainability is broader again because it also asks whether the system remains workable and responsible over time, including operational trade-offs.
Common misconception: Only training a model matters; using it later has little impact.
Correction: Training can be very resource-intensive, but repeated everyday use can also add up, especially at large scale. A model serving millions of requests, generating long outputs or being called many times in a workflow can have substantial ongoing energy demand.
Common misconception: If two systems use the same amount of compute, they have the same environmental impact.
Correction: Compute is only part of the picture. Environmental impact also depends on hardware efficiency, how the data centre is cooled, where the electricity comes from, how much networking is involved and how long the hardware lasts. Similar processing work can therefore lead to different emissions or water use.

Quick check

Which statement best shows the difference between these related terms?

Go deeper

The simple version

The technical version

The footprint of an AI system arises across training and inference, plus storage, networking and facility overheads. Direct operational impacts include electricity consumed by processors and supporting infrastructure, and sometimes water used in cooling. Carbon footprint refers specifically to greenhouse gas emissions associated with that resource use and potentially other lifecycle stages, depending on the accounting boundary used. Environmental footprint is a broader category that can include energy demand, water consumption, embodied impacts in hardware manufacture, material extraction and end-of-life electronic waste. Sustainability is the wider judgement about whether a system delivers sufficient value relative to these impacts and other constraints over time. This is why boundary-setting matters: a comparison may count only model inference, or it may include training, idle capacity, transmission, cooling overhead and hardware lifecycle. The term is commonly confused with compute, but compute is a measure of processing work, not a full environmental outcome.

Reviewed external resources

Primary sourceAI Risk Management Framework (opens in a new tab)NIST
This NIST framework helps policy, risk and governance readers think about AI impacts and trade-offs in a structured organisational way, even though it is not a single-topic guide to energy or water.
Intermediatelink checked 2026-07-31
Primary sourceOECD AI Principles (opens in a new tab)OECD
These OECD principles give a high-level policy frame for trustworthy and responsible AI and suit leadership readers who want the wider governance context.
Intermediatelink checked 2026-07-31
EncyclopaediaEnvironmental impact of artificial intelligence (opens in a new tab)Wikipedia
This Wikipedia page gives a broad introductory overview of environmental impacts linked to AI and suits readers who want a quick starting map of the topic.
Intermediatelink checked 2026-07-31
VideoVideo explanations: Energy and water use of AI (opens in a new tab)YouTube search
This YouTube search results page offers multiple video explanations on energy and water use of AI and suits readers who prefer to compare several visual explainers rather than rely on one source.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Energy and water use of AI (opens in a new tab)Reddit search
This Reddit search results page shows community discussions and questions about AI energy and water use and suits readers who want to see common public confusions and debates, but posts vary in quality and should be checked carefully.
search results, not one pagelink checked 2026-07-31

Where this came from

Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.

The one thing to remember

AI uses electricity across training, inference and supporting infrastructure, and may use water for cooling. Those resource uses are related to, but not the same as, carbon emissions.