Put the question to ChatGPT itself, “how much water do you use?”, and the chatbot will answer politely, at length, and without committing to a single number. It will speak of estimates, of variables, of studies that differ. This evasiveness is not modesty; it is an honest reflection of a genuine mystery. For the first three years of the AI boom, nobody outside a handful of technology companies could say with any authority how much water a conversation with a chatbot actually consumes. The question sounds absurd until one learns what is happening behind the glowing interface: in windowless buildings on the edges of cities, enormous quantities of fresh water are being evaporated every second to keep the machines from overheating. The numbers that have slowly emerged, from university researchers, corporate disclosures and, finally, OpenAI itself, tell a story that is more interesting than any viral headline suggests. It is a story of teaspoons and lakes, of dry reservoirs in Chennai, and of a question India can no longer afford to treat as trivia.
The Short Answer: From a Teaspoon to a Water Bottle
The first official figure arrived only in mid-2025, when OpenAI’s chief executive Sam Altman disclosed that an average ChatGPT query consumes roughly a third of a millilitre of water, about one-fifteenth of a teaspoon, along with about 0.34 watt-hours of electricity. It was the first time the company behind the world’s most popular chatbot had put a number on the record, and the figure was framed to reassure: a single question, after all, is barely a sip.
The academics who forced this conversation take a less comforting view. In 2023, researchers at the University of California, Riverside, led by Associate Professor Shaolei Ren, published a now-famous study titled “Making AI Less Thirsty”. They estimated that GPT-3, the model then powering ChatGPT, effectively “drank” a 500-millilitre bottle of water for every 10 to 50 medium-length responses, depending on where and when the servers were running. Training that same model, they calculated, consumed roughly 700,000 litres of water in Microsoft’s American data centres. Other European researchers have since put the per-prompt figure as high as 10 to 25 millilitres, and independent calculations for a medium-length GPT-4o response have landed near 3.5 millilitres.
Why do the estimates differ by two orders of magnitude? Partly because models and cooling systems have grown far more efficient since 2023; partly because a server in cool Ireland needs far less cooling water than one in hot, humid Chennai; and partly because some counts include only water evaporated on-site at the data centre, while others add the water consumed upstream at power plants generating the electricity. The honest summary is this: a casual chat costs a few drops to a few teaspoonfuls per exchange, no more, and the figure is real, measurable, and growing.
Why a Chatbot Needs Water at All
The mechanism is straightforward. Every prompt sent to ChatGPT triggers billions of calculations on specialised processors inside data centres, and every calculation produces heat. Left uncooled, a hall of such processors would cook itself within minutes. The cheapest way to displace that heat at scale is evaporation: water flows through cooling towers, absorbs the heat, and escapes into the atmosphere as vapour. Data centres also evaporate additional water to humidify the air and control static electricity. On top of this “direct” consumption sits an indirect cost, since the electricity feeding the facility is itself often generated by thermal power plants that withdraw and consume water. A chatbot, in other words, never drinks directly, but the infrastructure that thinks on its behalf drinks continuously, and the hotter and drier the city around it, the thirstier it becomes. Engineers estimate that an identical facility in Mumbai or Chennai can require 20 to 30 per cent more cooling water than one in Germany or Ireland.
From a Teaspoon to a Lake
A fraction of a millilitre means nothing until it is multiplied by the scale of adoption. By OpenAI’s own account, roughly 800 million people now use ChatGPT every week, firing off billions of queries a day. Multiply a third of a millilitre by that volume and the teaspoon becomes a reservoir: independent estimates have placed ChatGPT’s global water consumption at well over 100 million litres per day. The corporate disclosures tell the same story in slower motion. Microsoft, whose servers helped train and run OpenAI’s models, reported that its global water consumption jumped by 34 per cent in a single year, to about 6.4 billion litres in 2022, a spike researchers linked directly to AI workloads. Google, which runs its own rival models, consumed more than 23 billion litres (6.1 billion gallons) across its data centres in 2023, a 17 per cent year-on-year increase. Every time a person opens ChatGPT, or any online alternative to ChatGPT, to draft an email or settle an argument, a few drops of that global total evaporate somewhere, and with hundreds of millions doing so daily, the arithmetic turns startling.
Why India Cannot Afford to Look Away
For India, this is not an abstract accounting exercise. The NITI Aayog’s Composite Water Management Index warned as early as 2018 that 600 million Indians already face high to extreme water stress. Chennai lived the warning in June 2019, when, after two failed monsoons, its four main reservoirs ran nearly dry and the city declared “Day Zero”. Yet it is precisely in such cities that the data centre industry is expanding fastest, drawn by cheap land, undersea cables and digital demand. More than half of India’s data centres sit in water-stressed regions, with Mumbai alone hosting dozens of facilities and Chennai remaining a major hub despite its history of crisis. An analysis by S&P Global expects 60 to 80 per cent of India’s data centres to face high water stress this decade, and projects the sector’s water consumption to more than double, from roughly 150 billion litres in 2025 to about 358 billion litres a year by 2030. The bitter irony is hard to miss: the hottest cities, where cooling demands the most water, are often those where water is scarcest, and where farmers and households compete for every pipeline and borewell.
The Backlash Has Already Begun
Communities elsewhere have drawn their lines early. In 2023, during Uruguay’s worst drought in decades, residents of Canelones protested a proposed Google data centre that would have used 7.6 million litres of water a day, roughly the domestic consumption of 55,000 people, with demonstrators calling it “pillage of our water”. Google ultimately redesigned the project around air cooling before breaking ground. In Santiago, Chile, a court partially reversed the authorisation for Google’s Cerrillos data centre over water concerns, prompting the company to pause and rethink the design. In Memphis, USA, Elon Musk’s xAI was found to be drawing about a million gallons of water a day for its “Colossus” supercomputer before committing to recycled wastewater. The pattern is consistent: the industry expands quietly until the arithmetic of local water becomes public, and then it adapts, often late, and only after resistance.
Keep It in Perspective, But Keep Asking
Perspective matters, and honesty cuts both ways. A single ChatGPT query’s water cost is trivial beside the roughly 130 litres that go into a cup of coffee or the nearly 2,700 litres embedded in a cotton t-shirt. No individual need feel guilty about asking a chatbot a question, and vilifying ordinary users distracts from where the leverage actually sits: with the corporations building the facilities and the regulators who set the rules for them. Both Microsoft, which has pledged to be “water positive” by 2030, and Google, which has promised to replenish 120 per cent of the water it consumes, now face the task of proving those commitments against rapidly rising baselines. The deeper problem is transparency. The world’s most scrutinised technology produced no official per-query water figure until 2025, and estimates from independent researchers still vary several-fold. What citizens and journalists can do, 4+67and increasingly do, is simple: keep asking the question, and insist on answers specific enough to be checked. The number ChatGPT could not give about itself is the number the public is entitled to know.
Conclusion
How much water does ChatGPT use? By its maker’s own admission, about a fifteenth of a teaspoon per question; by academic estimates, sometimes a water bottle’s worth per few dozen exchanges; and at global scale, hundreds of millions of litres every day, evaporating mostly in places where water was never spare. None of this is a reason to abandon a genuinely useful tool, but it is a reason to retire the illusion that digital services are weightless. The cloud, it turns out, has a water table. In a country where 600 million people live with water stress and where the data centre industry is set to more than double its consumption by 2030, that is not an environmental footnote; it is a planning question for the decade ahead. The chatbot may never tell us exactly how thirsty it is. The lakes, reservoirs and cooling towers on which it depends already do.
Madeline Miller love to writes articles about gaming, coding, and pop culture.
