A search engine query used to be the smallest, most forgettable action on the internet: type a few words, get an answer, move on. AI chat has quietly replaced a lot of those searches, and with it, replaced a near-weightless action with one that draws on something physical: water, drawn from a river or an aquifer, and electricity, generated somewhere on a grid that still burns a good deal of gas and coal to keep up.

None of this is visible from a laptop or a phone. The infrastructure behind AI, like the infrastructure behind electricity itself, only becomes noticeable once someone adds up the numbers.

What a data centre actually runs on

Every AI query, image or block of code is processed in a data centre: rows of servers running hot enough that they need constant cooling, usually water-based, to avoid failing.

Only in the US, the Department of Energy has recorded data centre energy demand tripling over the past decade, with projections of it doubling or tripling again by 2028 as AI infrastructure keeps expanding.

Water enters the picture in two ways:

  • Direct consumption is the water used on-site for cooling.

  • Indirect consumption is the water used by power stations to generate the electricity those servers draw, which in most countries is the larger share by far.

The International Energy Agency estimated that data centres worldwide consumed about 560 billion litres of water in 2023, projected to reach around 1,200 billion litres by 2030, with roughly two-thirds of that coming from power generation rather than the cooling systems themselves. Around 80% of the water drawn into a data centre's cooling system evaporates rather than returning to the local supply, according to water researcher Shaolei Ren at UC Riverside.

So, how much water does one AI query actually use

This is where the numbers get genuinely difficult to pin down, and where a lot of viral claims have overstated things in one direction while the industry has, until recently, said very little in the other.

The most widely cited figure, from a UC Riverside study, put the water cost of a roughly 100-word AI response at around 500 ml, the equivalent of a small bottle of water, though that figure depends heavily on which model, data centre and local climate is behind the request.

Extend that per-prompt logic to the scale AI now runs at, and the aggregate figures grow quickly. A UN University report published in June 2026 estimated 2025 data centre water consumption at roughly the equivalent of 1.8 million Olympic swimming pools, with a high-AI-growth scenario reaching 9.3 trillion litres of AI-related water use by 2030. A separate peer-reviewed estimate for 2025 alone put AI systems' water footprint specifically at somewhere between 312 and 765 billion litres.

Company disclosures, still patchy but improving, fill in some of the gap. Google reported consuming 10.9 billion gallons of water in 2025, up 34% on the previous year and more than double its 2021 figure, attributing most of that increase to AI infrastructure growth; it says it replenished 78% of that volume through stewardship projects. Amazon published its first water figure in June 2026: 2.5 billion gallons consumed globally, at a water usage effectiveness of 0.12 litres per kilowatt-hour, well below the industry average of 0.84. Microsoft has said it is now water positive, five years ahead of its own target. The gap between these companies' figures is partly genuine efficiency and partly a reminder that "water positive" and "water usage effectiveness" are defined differently enough that comparing headline claims across providers is harder than it should be.

Who is actually checking

Despite all this, meaningful oversight barely exists. More than 200 AI-related laws now exist across over 100 countries, and the large majority focus on privacy, bias and security rather than water or energy. The UK has no AI-specific legislation at all; its 2023 white paper on AI regulation explicitly places questions of environmental sustainability outside its scope. The EU's AI Act does require developers to disclose known or estimated energy consumption, but only when the EU's AI Office requests it, and related environmental codes of conduct remain voluntary rather than compulsory. In the US, a bill requiring a full federal study of AI's environmental impact was introduced in the Senate in 2026 and has not yet been voted on. Even where reporting rules already exist, compliance is inconsistent: an analysis in Texas found 83% of the state's 341 data centres had not met mandatory water reporting requirements.

The UN Secretary-General launched a global disclosure initiative in June 2026, asking major AI developers to publish carbon, water and land-use figures. It's a voluntary call, which sums up where things currently stand: the data mostly exists, scattered across corporate sustainability reports published on each company's own schedule, with no shared standard forcing an apples-to-apples comparison and no regulator checking the numbers add up.

What this means for everyday use

None of this makes AI unique among resource-hungry technologies. Cloud storage, video streaming and cryptocurrency have all faced versions of the same reckoning. What sets AI apart is the sheer speed of its adoption, layered onto a data centre buildout that was already accelerating before generative AI existed, in places that don't always have water or grid capacity to spare.

For an everyday user, that turns a fairly abstract policy question into a genuinely practical one: given how little any of this is currently regulated or disclosed at the point of use, most of the responsibility for using AI thoughtfully currently sits with the person doing the prompting, not with a system built to flag it for them.

What you can actually do

A few changes make a real, if modest, difference, and add up when repeated across millions of users:

  • Batch your prompts. Ten small back-and-forth questions typically cost more in aggregate than one well-considered one. Think through what you actually need before typing.

  • Match the tool to the task. A calculator, a search engine or a quick mental estimate often answers something perfectly well without invoking a large model at all. Save AI for tasks that genuinely benefit from it.

  • Be deliberate with image and video generation. These tend to be far more resource-intensive than text, and "regenerate" is an easy habit to fall into without noticing how many attempts have piled up.

  • Favour providers that disclose their numbers. Water usage effectiveness and renewable energy sourcing vary enormously between companies. Where a provider publishes this data, it's at least possible to compare; where none exists, that absence is worth noticing too.

  • Support disclosure requirements, not just efficiency pledges. Voluntary commitments are easy to make and hard to verify. Written to your local representative or supporting organisations pushing for mandatory reporting has more effect on the industry as a whole than any single person's usage habits.

  • Ask before automating. Not every workflow needs an AI layer added to it. Sometimes the most sustainable option is simply not building the extra step in the first place.

At EnergieBee, we think a lot about the resources that go unseen behind everyday technology.

We created an app to help people understand how resources are being used and wasted in households. We arrived to this creation inspired by nature. We keep observing how nothing is wasted neither neglected in Mother Nature's balance.

Is our goal to keep this practice and inspire others to make time and space to stay mindful and active when it comes to live in harmony with the planet, before is too late to find a solution at all.