The Hidden Water Cost of AI: How Data Centres Are Draining India Dry

Ask a chatbot to draft an email and it feels weightless — a few seconds, a page of text, no smoke, no exhaust. But AI water usage is real, and it is one of the least visible costs of the machines we now reach for a dozen times a day. Every prompt travels to a data centre where thousands of processors run hot, and keeping them cool takes electricity and, very often, water. The output arrives clean. The bill lands somewhere else — usually far from the person who typed the prompt, and increasingly, in India, on communities already short of water.

The number that stops people

A figure has been circulating for a couple of years now: that a single exchange with an AI model can consume something on the order of a small bottle of water. The exact number is contested, and it should be. It depends on the model, the hardware, the local climate, and how the cooling is done. Anyone quoting a hard, universal figure is selling a certainty that does not exist.

What is worth understanding is the order of magnitude, and why water enters the picture at all. A large AI model runs on racks of specialised chips that draw a great deal of power and throw off a great deal of heat. That heat has to go somewhere. Many data centres use evaporative cooling — water is evaporated to carry heat away, exactly as sweat cools skin — and that water is gone, not returned to the tap.

There is a second, quieter draw. The electricity powering the servers is itself often generated at thermal or hydro plants that consume water to run. So even a data centre that sips water directly may sit at the end of a supply chain that drinks heavily. When researchers estimate the water behind AI, they usually try to count both: the water cooling the building, and the water behind the power feeding it.

Put those together and the honest summary is this: a page of AI-generated text is not free of physical cost. Once you count cooling and generation, it plausibly carries a small but non-trivial water and energy footprint — and multiplied by billions of queries a day, that is no longer small at all. Treat any single number as an estimate. Treat the direction as settled.

The output arrives clean. The bill lands somewhere else — usually far from the person who typed the prompt.

India's data-centre boom meets its water stress

India is in the middle of a data-centre construction wave. Mumbai, Chennai, Hyderabad, and the belt around Delhi are all absorbing new capacity, driven by cheap connectivity, state incentives, and the global scramble to host AI workloads close to a billion-plus users. An AI data centre in India is now a routine line item in state investment announcements.

Set that against the other India — the one measured by the Central Ground Water Board, where large parts of the country are officially water-stressed and where the water table under several major cities has been falling for years. Chennai, memorably, ran its reservoirs nearly dry in 2019. Many industrial corridors already ration water to farms and households through the dry months.

The tension writes itself. Data centres want to cluster where the fibre, the power, and the customers are — which is often exactly where the water is tightest. Indian outlets have reported on the strain that large water-cooled facilities can place on local supply, and on the friction when an always-on industrial user competes with agriculture and residents for the same aquifer. I am not going to invent a specific site or a specific volume here, because the verified public numbers are still thin. But the structural problem is not speculative: you cannot pour a heat-intensive, water-cooled industry into a water-scarce geography and expect no one to notice.

This is the same pattern visible in the wider story of automation and the economy — the one running through the Indian IT layoffs in 2026, where the productivity gains and the costs land on very different people.

“The cloud” is a physical thing — and it enclosed a commons

We were taught to call it the cloud, and the word did a lot of quiet work. A cloud is airy, ownerless, everywhere and nowhere. It is the opposite of a warehouse full of humming metal drinking from a municipal water main. The metaphor was a kind of anaesthetic.

This is not a new trick but a very old one. Across ten thousand years, every technology revolution runs the same five-step move: a shared resource is quietly enclosed, its cost pushed onto people who never agreed to pay it, and the gain booked by whoever owns the new machine. The same move, a new machine, every time. Water and air — the original commons — are simply the latest things to be drawn behind a private meter and rebranded as somebody's operating expense.

Seen that way, “the cloud” is not a place in the sky. It is an enclosure of shared electricity, shared water, and shared atmosphere, wrapped in a word designed to make you forget it has a physical address. This is a close cousin of what critics call technofeudalism — an economy where a few platform owners collect rent on infrastructure the rest of us depend on but do not control.

Who carries the cost, who books the profit

So does AI use water? Yes — and the more useful question is whose water. The value of an AI service flows to its owners and shareholders, most of them far from the data centre. The water and the strained grid stay local. A farmer whose borewell drops a few metres has no invoice to point to, no line connecting his dry season to the server hall down the road. The cost is real but diffuse; the profit is concentrated and precise. That asymmetry is the whole game.

It rhymes with other Indian debates about who owns the value in a digital system. The infrastructure around Aadhaar and UPI raised versions of the same question — a public commons built at public cost, with the upside and the control not always landing where the cost did. The AI environmental impact story is that argument in physical form: instead of data flowing one way and value the other, it is water and power flowing one way and profit the other.

The cost is real but diffuse; the profit is concentrated and precise. That asymmetry is the whole game.

None of this makes AI uniquely villainous. Every heavy industry externalises something. But AI is unusual in how completely its costs are hidden from its users — you feel the convenience instantly and never see the reservoir. A technology that hides its price this well is a technology whose price deserves a closer look.

What accountability could actually look like

The good news is that this is a solvable problem, and none of the fixes require abandoning AI. They require making the hidden cost visible and then pricing it honestly. A few directions:

• Disclosure. Data-centre operators could be required to publish water and energy draw the way large factories report emissions — audited, standardised, per facility. You cannot manage what no one is allowed to measure, and right now most of these numbers are private.
• Siting rules. Water-scarce districts could treat a large data centre the way they would any thirsty industry — subject to a real water-availability assessment before approval, not waved through on the promise of jobs and prestige.
• Honest water pricing. When industrial water is priced near zero, waste is rational. Pricing that reflects genuine scarcity gives operators a reason to conserve rather than a reason to pump.
• Waterless and warm-water cooling. The engineering already exists — air-side economisation, closed-loop and immersion cooling, tolerating higher inlet temperatures. These trade some efficiency for a much smaller water footprint, and they get chosen far more often when water actually costs something and disclosure actually happens.

The through-line is simple: the reason AI's water cost feels invisible is that we have arranged, quite deliberately, not to look. The cloud metaphor, the private meters, the phrase “operating expense” — all of it works to keep the reservoir out of frame. Pull it back into frame and the debate becomes ordinary: an industry, a resource, a scarce commons, and a fair question about who pays.

India, more than almost anywhere, cannot afford to answer that question by accident. The country is building the data centres and running short of the water in the same decade, often in the same cities. Deciding on purpose who carries the cost — before the borewells run dry rather than after — is not anti-technology. It is just refusing, this one time, to let the same old move happen quietly again.