When you ask an online AI assistant a question, computers in a data centre do the work of producing an answer. A data centre is a building that houses computing equipment. Those computers need electricity and produce heat. Keeping them cool can also require water. This physical infrastructure is the starting point for understanding AI’s environmental impact. Li and colleagues, 2025.

The amount of electricity and water involved is not the same for every AI tool or every request. To understand why, it helps to look at how these systems run, where their resources come from and what they are being used to do.

Why does AI need electricity?

AI uses electricity both while it is being developed and while people are using it. During training, computers process data to develop a model’s capabilities. After training, computers still have to run the model each time it produces an answer. Researchers call this second stage inference. Both stages contribute to its environmental impact. Wu and colleagues, 2022.

Where the electricity comes from also matters. Electricity supplied by coal and gas has a different emissions profile from electricity supplied by wind, solar or nuclear power. The International Energy Agency shows that the sources supplying data centres vary across regions. As a result, knowing how much electricity a service uses is only part of understanding its effect on the climate. IEA, Energy supply for AI, 2025.

The equipment itself matters too. Making chips and servers creates emissions before they are switched on. Wu and colleagues’ study of AI systems at Facebook examined both manufacturing and use. Their work shows why an account of electricity use alone leaves out part of the picture. Wu and colleagues, 2022.

How can a computer use water?

Water can help carry heat away from computing equipment. Some data centres use cooling systems in which water evaporates as heat is released. Water may also be used at the power plants that generate their electricity. The amount depends on the cooling system, the electricity supply and local conditions. Li and colleagues, 2025.

This is why a statement such as “one AI question uses this much water” needs context. An estimate for one model in one location cannot automatically describe another service. The water study by Li and colleagues includes a modelled example using GPT-3. Li and colleagues, 2025.

For a community considering a new data centre, useful questions would include where its water will come from, how much it expects to use and how that use fits with other local needs. Those questions connect the technology to the place where it operates.

How large is the electricity demand?

The International Energy Agency’s April 2026 report puts global data-centre electricity use at about 485 terawatt-hours in 2025. Its central forecast reaches about 950 terawatt-hours in 2030, nearly twice as much. A terawatt-hour is a unit for measuring electricity use at this large scale. IEA, Key Questions on Energy and AI, 2026.

These figures cover data centres overall, including computing other than AI. They help explain the scale of demand facing electricity systems, but they cannot tell us how much electricity a chatbot answer uses.

For that smaller question, we need information about the particular model and task. A 2024 study by Luccioni, Jernite and Strubell found that energy use varied considerably across the models and tasks they tested. Its results help explain why there is no single number that represents every AI request. Luccioni and colleagues, 2024.

When is using AI worth the resources?

Environmental information becomes more useful when connected to a specific purpose. Consider a hypothetical library deciding whether to add an AI assistant to its website.

If visitors mainly need opening hours, a clear information page might meet that need. If they need help understanding a complicated policy, an assistant might offer something more. The library would first need to check that its answers are accurate, accessible and useful.

It could then compare the resources needed to provide that help. The comparison should include repeated attempts and corrections, rather than counting only the first answer. A service that produces an unusable answer has not completed the same task as one that successfully helps the visitor.

The approach is to identify the need, test whether each option meets it, then compare their environmental demands using the information available.

Could AI also help the environment?

AI can help protect resources by finding problems that would otherwise be difficult to detect. One example is water leaking from underground pipes.

In Singapore, researchers working with the national water agency, PUB, developed a system that combines AI with models of the water network. It analyses readings from sensors to identify unusual changes in pressure and flow. In a documented case, the system flagged a suspected leak, and a field team subsequently located it underground. Chew and colleagues, 2024.

Finding a leak gives maintenance teams a chance to repair it and stop water being lost. The environmental benefit comes from that repair. To understand how much AI contributes, the water agency would need to track whether it helps teams find and fix leaks sooner than their existing methods, how much water is saved, and the resources needed to operate the detection system.

What can readers do with this information?

This guide suggests three practical starting points:

  • Start with the task. Decide what help is needed before choosing a tool. Consider whether an existing page, conventional software or an AI service can provide it reliably.
  • Look for a clear explanation behind a number. A useful environmental estimate says which service it describes and whether it includes electricity, water, equipment manufacturing or only some of these.
  • Leave room for what is unknown. If a provider has not supplied enough information, avoid treating the missing data as proof that its impact is either small or large.

For Interflourishing, the underlying question is whether a use of AI provides a worthwhile benefit while taking its demands on people and the environment seriously. Answering that question requires attention to the particular use, the available alternatives and the communities affected.

Related reading: Should we use AI? A decision worksheet, Interflourishing’s working definition, research directions and editorial standards. Corrections can be sent through the contact page.