The Hidden Cost of ChatGPT: How Much Water and Electricity Does Your AI Chat Actually Use?

You typed a question into ChatGPT this morning. Maybe you asked it to summarise an email, write a birthday message, or explain something you were too embarrassed to Google in plain words. It felt free, instant, and weightless, the way typing into a search bar always has. Somewhere hundreds of kilometres away, though, that single question pulled real electricity off a real power grid, and helped evaporate a measurable amount of real water into the air to keep a server rack from overheating.

Exactly how much water and electricity, though, turns out to be a genuinely contested question, one where even the people building these systems and the researchers studying them from outside don't fully agree. That disagreement itself is worth understanding, especially for a country like India, where a lot of this new AI infrastructure is being built in cities already worried about running out of water.

What OpenAI Itself Says a Single Query Costs

In mid-2025, OpenAI CEO Sam Altman published a blog post with a specific, concrete claim: the average ChatGPT query uses about 0.34 watt-hours of electricity and roughly 0.000085 gallons of water, which works out to around 0.32 millilitres, a little over a fifteenth of a teaspoon. To put the electricity figure in everyday terms, that's about what a high-efficiency lightbulb uses running for a couple of minutes.

Independent researchers have found numbers in a broadly similar range for the direct, on-site cooling water alone. Research institute Epoch AI estimated a single GPT-4o query at approximately 0.0003 kWh of energy, closely matching OpenAI's own figure. On the surface, this looks like a genuinely reassuring, settled answer: tiny, almost negligible per use.

Why Independent Researchers Tell a Different Story

Here's where the story gets genuinely more complicated, and worth sitting with rather than picking whichever number sounds most convenient. OpenAI's figure counts only the water evaporated directly on-site at the data centre for cooling. It doesn't count the water used further upstream, at the power plants generating the electricity that runs the servers in the first place, and thermal power generation itself uses significant water for its own cooling.

When researchers factor that upstream water in, the numbers climb substantially. A widely cited 2023 study from UC Riverside, led by researcher Shaolei Ren, originally estimated the full water footprint, direct and indirect combined, at around 519 millilitres for a 100-word email, roughly a full bottle of water. That specific figure went viral and became one of the most widely repeated statistics about AI's environmental cost. Ren has since revised that estimate downward, to roughly 15 millilitres per GPT-4 prompt, including about 5ml of direct, on-site cooling water, reflecting improvements in model efficiency and updated methodology. Even this revised, lower figure is still tens of times higher than OpenAI's own on-site-only number, because it's measuring a genuinely different, broader thing.

Why the Numbers Disagree So Much, and What That Actually Means

This isn't really a case of one source being right and another being wrong. It's a case of different studies measuring genuinely different boundaries around the same activity, and rarely making that boundary clear in the headline number. OpenAI's 0.32ml figure counts direct, on-site evaporative cooling only. UC Riverside's figures, both the original and the revised one, count that same direct cooling plus the indirect water used to generate the electricity itself, a meaningfully larger scope. Neither approach is dishonest. They're simply answering slightly different questions, and most casual coverage of this topic doesn't make that distinction clear before quoting a single, seemingly precise number.

It's also worth being honest that OpenAI hasn't published the detailed methodology behind its own 0.34 Wh and 0.32ml figures, which means the number, while plausible and broadly consistent with some independent estimates, can't be fully independently verified the way an academic study's figures typically can. A reasonable, honestly uncertain summary of where the evidence currently sits: a single typical text query likely uses somewhere in the range of a fraction of a millilitre to around 15 millilitres of water when the full picture is counted, and somewhere between roughly 0.3 and a few watt-hours of electricity, depending heavily on model size, query complexity, and exactly what's being measured.

Why the Per-Query Number Matters Less Than the Scale

Whichever end of that range turns out to be most accurate, the more meaningful number isn't really the tiny amount used by any single question. It's what that tiny amount adds up to across a global user base sending billions of queries a day. Even taking OpenAI's own, more conservative electricity figure, roughly 1 billion daily queries at 0.34 Wh each works out to somewhere in the region of 340 megawatt-hours a day just for that one product's day-to-day operation, alongside separate, much larger amounts of electricity used for training the underlying models in the first place.

A single query genuinely is close to negligible on its own. A global population treating AI chat as a daily habit, the way search engines became one, is a different scale of question entirely, and it's the second question that's actually driving the infrastructure boom currently reshaping electricity grids and water systems in multiple countries, India included.

The Part That Should Matter More to India: Where This Infrastructure Is Being Built

This is where the story becomes genuinely more pressing for an Indian reader than an abstract per-query millilitre figure. India's data centre capacity is projected to roughly double from around 0.9 GW in 2023 to approximately 2 GW by 2026, and the Union Budget 2026 extended a tax holiday through 2047 specifically to attract foreign companies building data centres in the country. A 2026 white paper from the Council on Energy, Environment and Water, a Delhi-based think tank, estimated India's data centres consumed roughly 150 billion litres of water in 2025 alone, a figure projected to more than double by 2030.

The locations chosen for this infrastructure make the concern sharper still. Data centres cluster in Mumbai, Bengaluru, Chennai, Hyderabad, and the Delhi-NCR region, cities chosen for their power access, fibre connectivity, and talent pools, but also cities already confronting acute water stress. WRI India's research counts 278 data centres nationally, with more than half located in regions classified as under water stress, and separate analysis from Planet Tracker found 50 Indian facilities already sitting in "extremely high" water-stress zones. S&P Global has projected that 60 to 80% of India's data centres will face high water stress at some point this decade.

Hyderabad and Bengaluru: Two Real, Specific Cases

Two examples make this abstract tension concrete. Hyderabad, one of India's most active data centre hubs, faces a projected water deficit of roughly 870 million litres per day by 2027, according to reporting on the region, and yet Amazon continues actively expanding its data centre operations there. Karnataka's IT Minister, Priyank Kharge, told the state assembly in March 2026 that each megawatt of data centre capacity requires roughly 25 million litres of water a year, a figure that scales quickly as India's capacity climbs toward gigawatt-level ambitions. Bengaluru's data centres alone are estimated to consume over 26 million litres a year, in a city that recently experienced what local reporting described as its worst water crisis in nearly five centuries.

Pune tells a similar story on the electricity side, with Microsoft actively building out AI facility capacity in a city that saw consistent water shortages and public protests against local officials the previous year. This isn't a hypothetical future tension. It's an already-visible mismatch between where digital infrastructure is expanding fastest and where the physical resources to support it are already under genuine strain.

Why This Actually Matters, Beyond the Environmental Headline

It would be easy to file this under a distant, abstract environmental concern that doesn't touch daily life. But the water these data centres draw on doesn't come from a separate, dedicated supply built just for them. It draws from the same regional water systems that supply drinking water, agriculture, and industry to everyone else living nearby, in cities where India's Central Water Commission and various state governments have already flagged serious future shortfalls, independent of AI's growing footprint. None of India's fifteen current state-level data centre policies, according to reporting on the issue, currently include mandatory water stress mapping as a condition for environmental clearance, meaning this expansion is largely happening without a formal, binding requirement to account for the local water reality first.

There's a genuine tension worth naming honestly here, not to argue against AI infrastructure existing at all, but to argue for it being built with real accountability. India also stands to gain enormously from hosting more of this infrastructure domestically, in jobs, digital sovereignty, and reduced dependence on foreign-hosted compute, particularly given that India currently generates close to 20% of the world's data but hosts only around 3% of global data centre capacity. The reasonable position isn't "no data centres." It's that the water and electricity cost of this expansion deserves the same serious scrutiny and planning that any other large industrial development would face, rather than being treated as an afterthought behind the more exciting AI headlines.

A Few Honest Ways to Think About Your Own Usage

This isn't really a story where individual restraint solves the larger structural problem, the scale is driven overwhelmingly by infrastructure decisions and corporate expansion plans, not any one person's chat habits. Still, a few genuinely useful things worth knowing.

Longer, more complex queries, and image or video generation specifically, use meaningfully more resources than a short text answer. If you're running the same simple question through an AI tool five times to tweak the phrasing slightly, that adds up faster than a single, well-thought-out prompt.

The bigger environmental lever isn't your personal query count, it's which companies you support and how they source their infrastructure. Some providers are investing more seriously in liquid cooling, water recycling, and siting new facilities away from already water-stressed regions than others, and that corporate-level choice matters far more at scale than any individual's daily habits.

Be appropriately skeptical of any single, precise-sounding number you see shared online, including the ones in this article. Given how much genuine disagreement exists even among serious researchers and the companies themselves, a number presented with total confidence and no source is usually oversimplifying a genuinely contested, evolving picture.

The Honest Bottom Line

A single ChatGPT query, by any credible estimate, uses a genuinely small amount of water and electricity, somewhere between a fraction of a millilitre and a small handful of millilitres, and a fraction of a watt-hour to a few watt-hours. That part of the popular conversation around AI's environmental cost has sometimes been genuinely overstated by early, viral figures that have since been walked back by their own authors. What hasn't been overstated is the aggregate picture: billions of queries a day, running on infrastructure being built rapidly in some of the world's most water-stressed cities, India's own AI hubs very much included. The individual question you ask ChatGPT this evening is close to weightless. The industry being built to answer it, especially where it's being built, genuinely isn't.

Frequently Asked Questions

Q1. How much water does a single ChatGPT query actually use?

Estimates vary meaningfully depending on what's being measured. OpenAI's own figure, cited by CEO Sam Altman, puts a single query at roughly 0.32 millilitres, counting only direct, on-site cooling water. Independent research from UC Riverside, which also counts the water used upstream to generate the electricity powering the query, has estimated a broader figure of roughly 15 millilitres per prompt after revising down from an earlier, more widely cited estimate of 519 millilitres for a 100-word email.

Q2. Why do different sources give such different numbers for AI's water and electricity use?

The main reason is that different studies measure different boundaries around the same activity. Some figures, like OpenAI's, count only the water evaporated directly on-site at a data centre for cooling. Others, like UC Riverside's research, also include the indirect water used at power plants to generate the electricity the data centre consumes, a meaningfully larger scope that produces a higher total figure for what's ultimately a broader measurement.

Q3. Why is this a particular concern for India specifically?

India's data centre capacity is expanding rapidly, roughly doubling toward 2 GW by 2026, and much of this infrastructure is concentrated in cities including Mumbai, Bengaluru, Chennai, and Hyderabad that already face significant water stress. Research bodies including WRI India and S&P Global have found more than half of India's current data centre capacity sits in water-stressed regions, with projections suggesting 60 to 80% could face high water stress at some point this decade, raising real concerns about competing demand between digital infrastructure and existing drinking water and agricultural needs.

Q4. Does India's rapid data centre growth also mean more electricity demand?

Yes, water and electricity demand rise together in this context, since more electricity is required to power AI hardware, and more electricity generation typically requires more water for cooling at power plants. India currently generates close to 20% of the world's data but hosts only around 3% of global data centre capacity, a gap driving significant new infrastructure investment, alongside growing pressure on regional power grids in the same cities already facing water stress.

Q5. Can an individual meaningfully reduce AI's environmental footprint by using it less?

Individual usage has a comparatively small effect on the larger picture, since the scale of AI's environmental footprint is driven primarily by infrastructure decisions, corporate expansion plans, and where new data centres are sited, rather than any single person's query volume. That said, longer or more complex queries, along with image and video generation, do use meaningfully more resources than a short text response, and repeatedly rerunning similar prompts adds up faster than a single well-considered one.

This tension between AI's rapid growth and its real-world resource cost connects to other questions about how responsibly this technology is being built and deployed. AI Agents Are Here: Why Indian Companies Are Quietly Replacing Freshers looks at another side of AI's rapid expansion in India, and Humanoid Robots Are Coming to Indian Factories covers a related wave of infrastructure-heavy AI investment reshaping the country.