Every prompt travels to a data centre where powerful chips consume electricity and cooling systems use water. As artificial intelligence expands, its invisible infrastructure could complicate the fight against climate change.

AI may exist on our screens, but its environmental footprint is physical. Every answer depends on chips, electricity, cooling systems and data centres whose rapid expansion is raising difficult questions about water, energy and climate change.

Your question has a physical address

You type a question into an AI chatbot and watch the answer appear almost instantly. There is no smoke, engine or factory floor in sight, so the process can feel almost weightless.

Yet the response was produced inside a physical building, sometimes thousands of kilometres away. Rows of specialised computer chips processed the request, cooling equipment carried away the heat, and a power station or renewable-energy project supplied the electricity.

The “cloud” is not floating above us. It is a rapidly expanding collection of data centres connected to power grids, water systems and communities on the ground.

AI runs on electricity, not magic

Artificial intelligence systems depend on specialised chips capable of performing enormous numbers of calculations. Those processors require electricity, while servers, networking equipment and cooling systems add to the total demand.

The International Energy Agency puts the relationship plainly: “There is no AI without energy.” As AI becomes part of search engines, offices, schools, hospitals and entertainment, the infrastructure required to support it is growing rapidly.

The IEA’s 2026 assessment estimates that global data-centre electricity consumption could rise from about 485 terawatt-hours in 2025 to approximately 950 terawatt-hours in 2030. That would place data centres at around 3% of global electricity demand, while consumption by AI-focused facilities is expected to triple.

Three percent may not sound overwhelming when viewed across the entire world. The concern is the speed of the increase and the fact that demand is often concentrated in particular cities, states and electricity networks.

Training a model is only the beginning

Creating a large AI model requires a power-intensive process known as training. Vast quantities of data are passed through thousands of chips as the system learns statistical relationships between words, images, sounds or other information.

Training receives much of the public attention because a single advanced model can require weeks or months of computing. However, the environmental cost does not end when the model is released.

Every summary, image, search and chatbot conversation requires another process called inference. One request may use relatively little electricity, but billions of requests repeated across the world can turn everyday use into a major source of demand.

More capable systems may also perform longer chains of calculations before responding. AI-generated images and videos can require significantly more computing than a short text answer, making it misleading to treat every prompt as environmentally equal.

America’s power grid is already feeling the pressure

The United States sits at the centre of the AI data-centre boom. Technology companies are building enormous facilities and searching for locations with available land, electricity, water and connections to high-capacity transmission networks.

A report prepared by Lawrence Berkeley National Laboratory for the U.S. Department of Energy estimated that data centres consumed about 176 terawatt-hours of electricity in 2023. That represented approximately 4.4% of all U.S. electricity use.

The report projected that data centres could account for between 6.7% and 12% of national electricity consumption by 2028. AI servers are expected to be one of the main drivers of that growth.

National percentages do not reveal the full pressure placed on individual communities. A new facility may demand as much power as a small city, requiring additional transmission lines, substations and generating capacity in a relatively short period.

Why artificial intelligence also needs water

Computer chips generate large amounts of heat while operating. If that heat is not removed, equipment can slow down, malfunction or suffer permanent damage.

Some data centres use outside air or closed-loop cooling systems, but many rely partly on evaporative cooling. In those facilities, water absorbs heat before some of it evaporates into the atmosphere.

The Berkeley Lab report estimated that U.S. data centres directly consumed roughly 17 billion gallons of water in 2023. That figure does not include all the water used indirectly by power plants generating their electricity.

Water demand varies considerably between facilities. The climate, cooling technology, time of year, electricity source and availability of reclaimed water can all change the footprint.

This matters because some data centres are being built in places already experiencing drought or competition for water. A facility may appear modest in a global calculation while becoming highly significant to farmers, households and ecosystems sharing the same local supply.

The viral “bottle of water per prompt” claim needs context

A widely repeated claim suggests that a short AI conversation consumes an entire bottle of water. It originated partly from a 2023 academic study that modelled how approximately 20 to 50 GPT-3 questions and answers could be associated with 500 millilitres of water under particular operating conditions.

That was a scenario-based estimate, not a measurement that applies to every chatbot, question or data centre. The amount can change depending on the model, response length, server efficiency, location, temperature and source of electricity.

This does not mean AI’s water footprint is imaginary. It means a single dramatic number can distract from the larger issue: millions of servers operating continuously across facilities that do not always disclose comparable water data.

The climate impact depends on what produces the electricity

A data centre powered by wind, solar, hydropower or nuclear energy has a different carbon footprint from one dependent on coal or natural gas. The same AI task can therefore produce different emissions depending on where and when it is processed.

The IEA expects renewable energy to meet nearly half of the additional electricity required by data centres through 2030. However, its energy-supply analysis also predicts that natural gas and coal will help satisfy part of the increase, particularly where clean generation and grid connections cannot expand quickly enough.

There are other emissions before a model even begins operating. Manufacturing chips, constructing buildings, producing concrete and steel, and installing backup generators all contribute to the climate footprint of AI infrastructure.

Buying renewable-energy certificates can help finance cleaner electricity, but annual purchases do not always mean a data centre is running on carbon-free power every hour. What matters is whether new AI demand is matched by additional clean energy when and where that electricity is required.

Communities may pay costs that users never see

The price displayed for an AI subscription does not necessarily include every cost created by the infrastructure behind it. Utilities may need to build power lines, substations or generating capacity, raising questions about how those investments are divided between technology companies and other electricity customers.

Water can create similar tensions. A data centre may bring construction, tax revenue and jobs, but residents may still ask whether an industrial facility should receive large water allocations during drought restrictions.

These debates are not an argument against every data centre. They are a demand for transparency about who receives the economic benefits and who carries the environmental and infrastructure costs.

Efficiency is improving, but demand keeps growing

Technology companies say they are making chips more efficient, purchasing renewable energy and developing cooling systems that consume less water. Google reported in its 2025 Environmental Report that it reduced data-centre energy emissions by 12% in 2024 despite rising demand.

Microsoft says its direct-to-chip cooling technology can save more than 125 million litres of water annually at an individual facility. Its sustainability report also describes greater use of reclaimed water, renewable energy and lower-carbon construction materials.

Those improvements matter, but efficiency does not automatically reduce total consumption. When AI becomes cheaper and faster, companies may place it inside more products, people may use it more frequently, and total demand can continue rising even as each task becomes more efficient.

The climate question is therefore not only whether the next chip uses less electricity. It is whether efficiency can improve faster than the world’s appetite for AI expands.

AI could also become a climate tool

Artificial intelligence is not only an energy consumer. It can help electricity operators forecast demand, balance renewable power, identify methane leaks, improve building efficiency and discover new materials for batteries and clean-energy technologies.

The IEA estimates that widespread adoption of existing AI applications could reduce emissions by an amount equal to roughly 5% of energy-related emissions in 2035. Its climate analysis warns, however, that this potential will not be realised automatically.

AI cannot lower emissions simply because a company describes its product as intelligent. Climate benefits depend on deploying the technology in places where it produces measurable savings larger than the energy and materials required to operate it.

What Next?

Technology companies should publish clearer, independently verifiable information about the electricity, water and emissions associated with their models. Reporting should distinguish between training and everyday use while showing how impacts vary by location and source of power.

New data centres can be directed toward regions with adequate grid capacity and lower water stress. Operators can use reclaimed water, less water-intensive cooling, more efficient models and clean electricity added to the grid specifically for new demand.

Governments must also protect communities from carrying infrastructure costs created primarily for private facilities. Water permits, electricity rates and data-centre incentives should be examined publicly before long-term agreements are signed.

Individual users can avoid unnecessary generations and choose smaller tools when those tools can complete the job. Yet the largest responsibility belongs to the companies designing models, selecting data-centre locations and deciding how the electricity will be produced.

The Bottom Line

AI is neither an invisible climate villain nor an automatic environmental saviour. It is a powerful industrial technology whose consequences depend on energy sources, water choices, efficiency, regulation and the purposes for which it is used.

The answer appearing on a screen may feel effortless. The electricity, water and infrastructure behind it are anything but.