AI's Hidden Thirst

The water AI is quietly drinking

The AI prompt you just sent? That used about 500 ml of water. The answer that came back? About a litre.

AI is consuming freshwater at a scale most of us haven't started thinking about. The Environmental and Energy Study Institute (EESI), a US non-profit focused on clean energy policy, has put out some strikingly clear numbers on this.

To put that in perspective: a typical Indian household of four people uses about 540 litres a day — or 1,97,100 litres a year. That one data centre used water that could have served 557 such households for an entire year. And it's medium-sized. Larger ones can use water equivalent to over 10,000 households annually.

The US has over 5,400 data centres. The total water use runs into billions of litres a day — and this isn't grey water. It is accessible freshwater, drawn from rivers, lakes, or pulled from aquifers. The reason? Freshwater has fewer dissolved minerals, so it corrodes cooling systems less and keeps them running longer.

Why freshwater, and where does it go?

Data centres use water in three broad ways — on-site operational needs, coolant systems, and chip manufacturing. The cooling systems work much like a large air conditioner or industrial heat exchanger. Liquid coolant absorbs heat from GPU vaults, and that heat is transferred to water in the exchange. This hot water then goes into evaporators — similar to what thermal and nuclear power plants use.

Some water is lost to evaporation. The rest is cooled, treated, and returned to the aquifer. But the water lost to evaporation is a net loss that would not have occurred if the data centre had not been there.

f

How small is that 0.5–1%?

Only 3% of Earth's water is freshwater. Of that, just 0.5–1% is accessible and safe for human use — frozen glaciers, deep underground reserves and contaminated sources account for the rest. The water that data centres compete for is the same water people drink, cook with, and grow food on.

OpenAI alone: 500 million litres a day

A University of California Riverside study found that every 100-word AI prompt uses 500 ml of water for cooling. OpenAI has stated that approximately 1 billion messages go into its systems every day. At 500 ml per prompt, that is 500 million litres — every single day, from one company. Then there is Anthropic, Google DeepMind, Gemini, Perplexity, and dozens of others.

India: a collision waiting to happen

A paper by Dr A Shaji George — Data Centres and Water Crisis in India: Why Digital Infrastructure Could Drain Our Wells Dry by 2030 — published on ResearchGate — brings this closer home.

A 100 MW data centre consumes water equivalent to what 6,500 households need. For perspective, the META–Reliance data centre announced in June is 168 MW. India already faces severe water stress — Delhi alone is deficient by 1,100 million litres a day. Hyderabad is similarly stretched.

It is not coincidental that META chose Jamnagar for its 168 MW data centre. One key reason is that Reliance Industries recycles water at its plant there — making it one of very few locations in India where a data centre of that scale is even feasible responsibly.

India's 270-plus data centres — plus the large ones being commissioned — are projected to consume 60 million litres a day, drawn from areas that are already water-stressed. This is a collision between two legitimate needs: digital infrastructure and human access to freshwater.

The trade-off we have to make

None of this is an argument against AI or data centres. They improve lives in real ways — healthcare diagnostics, agricultural advisories, climate modelling, education access. The point is to understand the cost at the other end of the wire.

Every industry goes through a phase of rapid growth and high resource consumption before the reckoning hits. The auto industry had a century of burning fossil fuels before existential pressure pushed it toward hybrid and electric. Data centres are not far behind that curve. The question is whether the reckoning comes too late for the aquifers.

Solutions exist — liquid immersion cooling, air-side economisers, recycled water loops, locating data centres near treated effluent sources. Some are being adopted. The pace needs to be much faster, particularly in water-scarce countries like India where both digital growth and water stress are accelerating simultaneously.