How much of a problem is AI’s water use?
AI data centres are becoming a focus of local protests in drought-prone parts of the US, where residents fear their water, energy and pollution impacts. Although their overall water use remains small compared with agriculture and some manufacturing, rapid AI expansion could intensify local shortages unless companies improve cooling, energy sources and site selection.
Researchers estimate data-centre cooling used 66 billion litres of water in 2023, under 1% of US consumption, but forecasts put total annual US data-centre water demand at 731 billion to 1,125 billion litres by 2030. Older claims that a chatbot query used 500 ml are already outdated as models become more efficient; Google estimated five drops for a typical Gemini query in 2025. Experts say renewable power, less water-intensive cooling and locating facilities in water-rich regions could cut AI’s future water footprint by up to 86%.
- AI water use is growing but remains nationally limited.
- Local impacts matter most in drought-stricken regions.
- Better siting and renewable power could sharply reduce demand.
New here? Start with this
AI systems run on computers housed in large data centres, which need cooling to stop them overheating. Some cooling methods use water directly, while the electricity that powers the computers can also require water to produce.
The main players are the technology companies building and operating these centres, and the communities near proposed or existing sites. Agriculture and some industries use far more water overall, but the impact can be concentrated in places already short of water.
As demand for AI grows, more data centres may increase pressure on local water and energy supplies. Their effects depend partly on where they are built, how they are cooled and what powers them.
Both sides, in good faith
The strongest fair case each way — we don't pick a winner.
The case for
AI data centres are concentrating genuine local water stress in communities already vulnerable to drought, creating an equity problem even if national percentages seem modest. With consumption projected to grow 11–17 fold by 2030, waiting for voluntary efficiency improvements risks locking in high water demand in precisely the regions least able to sustain it, warranting proactive intervention through regulation and site selection now rather than after damage occurs.
The case against
AI water consumption remains negligible at under 1% of US total use, and efficiency is improving rapidly—queries have dropped from 500 ml to five drops in just a few years. Established technical solutions can reduce future impacts by 86%, making this a manageable problem through innovation and better siting practices rather than restrictive policy. Agricultural water use dwarfs AI's consumption and offers fewer efficiency alternatives, so resources would yield greater impact elsewhere.