The Hidden Cost of Every Prompt: How AI Is Draining the World’s Water

AI query you type has a water footprint — cooling the servers that process it, and the power plants that supply them electricity. As AI adoption scales into billions of daily queries, global data centre water consumption is on track to hit 9.3 trillion litres by 2030. India, home to 18 percent of the world’s population but only 4 percent of its freshwater, is building data centres fastest in exactly the cities — Mumbai, Hyderabad, Chennai, Bengaluru — that are already running short.

THE NUMBER BEHIND THE HEADLINE

The claim that “one ChatGPT prompt equals a bottle of water” traces back to a single 2023 paper by UC Riverside researcher Shaolei Ren, which estimated that {{GPT-3 “drinks” a 500ml bottle of water for every 10 to 50 medium-length replies}}, depending on the data centre’s location and the weather that day. When the Washington Post revisited the estimate for GPT-4 in 2024, the number came out closer to 519ml for a single 100-word response — a figure roughly 30 times larger than Ren’s original per-prompt math, though he never fully explained the jump, and the underlying assumptions (a much higher energy cost per query) differ from the original study. Ren himself has called the per-query figure conservative and cautioned that real-world use could run several times higher.

The honest answer is that a single prompt’s water cost is genuinely small and genuinely hard to pin down — estimates for a routine text reply range from a fraction of a millilitre to several dozen. What scales is not the prompt. It’s the number of prompts. Training a single large model is a different order of magnitude entirely: UC Riverside estimated GPT-3’s training run evaporated roughly 700,000 litres of freshwater on-site alone, with total lifecycle water consumption — including the electricity used to power it — closer to 5.4 million litres. Later, more power-hungry models cost more: one industry estimate puts GPT-5-class training water use near 500 million litres, enough to fill 200 Olympic swimming pools.

WHY COOLING NEEDS WATER AT ALL

AI servers generate enormous heat, and the cheapest way to remove it at scale is evaporative cooling — the same principle as sweat cooling skin. About 80 percent of the water a data centre draws in simply evaporates into the air; the rest returns as warm wastewater. A single 1-megawatt facility can consume roughly 25–26 million litres of water annually just for cooling. Hyperscale campuses running at 100 megawatts or more can draw millions of gallons a day — comparable to the water needs of a small city. Newer chip generations, like NVIDIA’s Vera Rubin GPUs announced in January 2026, can run on water heated to 45°C without chillers, a genuine efficiency gain — but that innovation applies mainly to new AI-specific builds, while most existing infrastructure still relies on older, thirstier cooling methods.

There’s a second, less visible water cost too: the electricity itself. Thermal power plants — coal and gas especially — consume water to generate the power a data centre draws from the grid. In the US, this indirect water use from electricity generation was estimated at 211 billion gallons in 2023 alone, roughly 1.2 gallons for every kilowatt-hour consumed. A data centre’s total water footprint, in other words, isn’t just what happens on-site — it’s the water cost of the entire supply chain behind it.

The scale problem

A UN University report published in mid-2026 — the first UN-commissioned study to quantify AI’s water and land footprint alongside its carbon impact — estimated that global data centres consumed water equivalent to 1.8 million Olympic-sized swimming pools in 2025, and that this could rise to 9.3 trillion litres annually by 2030 as AI’s market value grows roughly 25-fold over the decade. In the US, hyperscale data centre water consumption alone could reach 16 to 33 billion gallons a year by 2028, according to Berkeley Lab projections. Among individual companies, Google disclosed 10.9 billion gallons of water consumption in its 2026 environmental report — more than four times Amazon’s disclosed 2.5 billion gallons — making Google the largest known water consumer among hyperscalers that publish figures. Meta and Microsoft have not yet published comparably detailed absolute figures, meaning much of the industry’s true footprint is still estimated rather than confirmed.

INDIA’S SPECIFIC EXPOSURE

India is not a peripheral player in this story — it is one of the fastest-growing data centre markets in the world, expanding IT capacity from 0.4 gigawatts in 2020 to 1.5 gigawatts by 2025, with Deloitte projecting another 8–10 gigawatts by 2030. Google alone broke ground on a $15 billion facility in Visakhapatnam in April 2026; Amazon has committed a further $7 billion to expand its Mumbai and Hyderabad sites; Microsoft has pledged $17.5 billion through 2029. India’s data centre water consumption is projected to more than double, from roughly 150 billion litres in 2025 to 358 billion litres by 2030.

The locations make this more than an abstract number. Mumbai, Hyderabad, Chennai and Bengaluru together account for the large majority of India’s data centre capacity — and all four are cities already confronting water stress. Hyderabad faces a projected daily deficit of up to 909 million litres within two years, even as Amazon expands its facility there; groundwater levels in parts of the city reportedly dropped by roughly a metre between January and April 2026 alone. Mumbai’s own reservoirs stood at just 23 percent of capacity by mid-2026, prompting a municipal 10 percent water-use restriction. India holds 18 percent of the world’s population but only 4 percent of its freshwater resources — among the tightest ratios of any large country — and the IMD’s 2026 long-range monsoon forecast, at 92 percent of the long-period average, was the weakest opening forecast in 25 years.

WHAT’S BEING DONE

The industry is not standing still. Liquid and immersion cooling systems can cut water use by 30–50 percent compared to evaporative systems, and several Indian facilities — including Sify Technologies’ Mumbai-Rabale campus — have adopted them. Microsoft has committed to being “water positive” by 2030; Google says it replenished 78 percent of its 2025 consumption through watershed restoration projects. Critics, including Ren, note that these commitments are typically measured globally rather than locally — replenishing water in a wet region does nothing for the aquifer beneath a drought-stressed data centre. Karnataka’s state government has publicly flagged data centres as heavy water and energy consumers, even as its 2022 policy still commits to guaranteeing them uninterrupted supply — a tension state and national policy have yet to resolve.