The carbon footprint of AI gets the headlines, but water is where the impact lands in someone's backyard. A data centre in a drought-prone region can pull millions of litres a day from the same aquifers that farms and towns rely on. And because the supply chain starts with chip fabrication — itself a water-intensive industry — the true footprint runs from the fab to the server hall. None of this makes AI bad; it makes the accounting incomplete if we only count watts.
The hidden thirst
The most-cited academic work is the 2023 study Making AI Less Thirsty by Li, Yang, Islam and Ren (arXiv:2304.03271, published in Communications of the ACM). It estimated that training GPT-3 in Microsoft's U.S. data centres directly evaporated about 700,000 litres of clean freshwater — roughly 5.4 million litres when the off-site water used to generate its electricity is included. The same study projected that global AI demand could account for 4.2–6.6 billion cubic metres of water withdrawal in 2027, more than the annual withdrawal of several Denmarks. Those are model-based estimates, but they frame the order of magnitude: AI's water use is already country-sized in the aggregate.
What the IEA projects
The International Energy Agency put firmer numbers on the sector in its recent data-centre analysis. It estimates global data-centre water consumption was around 560 billion litres in 2023, and could rise to about 12,000 billion litres (12 trillion litres) by 2030. Importantly, the IEA breaks down where it goes: of the 2023 total, roughly two-thirds was tied to energy generation and about a quarter to cooling, with the remainder linked to semiconductor manufacturing. A single average 100-megawatt data centre can consume around 2 million litres of water per day. The coupling is tight — power and water are the same problem seen from two taps.
Where the water actually goes
There are two pathways. Direct use is cooling: servers shed heat, and in many facilities that heat is carried away by water evaporating in cooling towers. Indirect use is the electricity itself — thermoelectric power plants are water-hungry, so every kilowatt-hour a data centre draws carries a hidden water cost unless the grid is hydro, nuclear or renewable with low water intensity. Operators track this with a metric called water-usage effectiveness (WUE), litres of water per kilowatt-hour, but reporting has been inconsistent, which is why comparisons between regions are hard. Location decides everything: a data centre in a cool, water-rich grid is a very different neighbour than one in a stressed watershed.
The land nobody counts
Water is only half the unseen footprint. Hyperscale campuses now sprawl across hundreds of acres each, and the sector's land use is climbing fast — research cited by the UN University has projected a data-centre land footprint exceeding 14,500 square kilometres globally. Add the semiconductor side: leading fabs such as TSMC in Taiwan use on the order of hundreds of thousands of tonnes of water per day, and during droughts the company has resorted to trucking supplies in. So the land and water pressure is felt first where chips are made, then again where models are run — and both tend to cluster in places already short of one resource or the other.
Why this matters now
AI demand is scaling faster than efficiency gains. Carbon gives the trend its sharpest edge: Microsoft has reported emissions roughly 23% above its 2020 baseline (driven by AI data-centre construction), and Google has reported greenhouse-gas emissions up about 48% since 2019, attributing the rise to data-centre energy and supply chain. The IEA estimates data-centre electricity emissions are about 180 million tonnes of CO2 today, potentially rising to 300 million tonnes by 2035. The point is not that AI is uniquely destructive — it is that absolute footprints are rising even as per-unit efficiency improves, which is exactly the pattern that makes water and land the constraints to watch.
The takeaway
Three things move the needle. First, siting: putting new capacity in low water-stress regions with clean grids avoids the worst local harm. Second, efficiency and transparency: better cooling, recycling, and public WUE reporting make the hidden footprint visible and accountable. Third, demand: for users, heavier tasks only when needed; for industry, hardware that does more per watt. AI is not going away, but a water bill and a land footprint that no one sees are easier to ignore than ones that are measured — and measurement is the first step to managing them.
Frequently asked questions
How much water does one AI query use?
Per-query figures are tiny but not zero. The widely cited 2023 study by Li et al. estimated that training GPT-3 directly evaporated about 700,000 litres of freshwater on-site (about 5.4 million litres including off-site electricity use), and that roughly 10–50 medium-length GPT-3 responses could consume about 500 mL of water depending on when and where the model ran. Inference is far smaller per query than training, but multiplied across billions of daily interactions the total is substantial.
Why does AI need so much water?
Two reasons. Directly, data centres use water to cool servers — often by evaporating it in cooling towers. Indirectly, the electricity that powers those servers is itself water-intensive at many power plants. The IEA notes that of the roughly 560 billion litres of water data centres consumed globally in 2023, about two-thirds was tied to energy generation and about a quarter to cooling.
Is this just data centres, or the chips too?
It is both. Semiconductor manufacturing is itself water-hungry: leading fabs, such as TSMC in Taiwan, use on the order of hundreds of thousands of tonnes of water per day and have had to truck in supplies during droughts. So the footprint runs from the mine and the fab that make the chips all the way to the data centre that runs the model — and to the land those campuses occupy.
Does renewable energy fix the water problem?
Partly. Clean electricity cuts the carbon and the indirect water tied to fossil power plants, and it is a major part of the answer. But it does not remove the direct cooling water data centres evaporate, nor the water used to make the chips. Water stewardship — siting in low-stress regions, recycling, and efficient cooling — has to be addressed on its own.
What can users actually do?
Individually, the lever is demand: fewer, more deliberate prompts and heavier tasks (long documents, images, video) only when needed. At the industry level, the real fixes are efficient hardware, transparent water reporting, and siting new capacity away from water-stressed watersheds. Many operators now publish water-usage-effectiveness metrics, which makes accountability possible.
This page is an informational compilation. For reference only — please refer to each source's official documentation.
Images: Pexels (free license) · Photos by contributors on Pexels.
Privacy Policy · Contact