You Just Asked ChatGPT to Write an Email. Here’s What It Drank.
You opened your laptop, typed a quick prompt — “write a professional follow-up email, about 100 words” — and hit enter. In the seconds it took ChatGPT to respond, somewhere inside a humming data center, roughly 519 milliliters of fresh water evaporated into the air. That’s not a metaphor. That’s not rounding. That’s almost a full standard water bottle, consumed by the infrastructure behind a single, forgettable email draft.
We’ve spent years debating AI’s carbon footprint. Carbon matters — but it has a rival problem that almost nobody is talking about: water.
—
Why Servers Are Thirsty
To understand AI’s water hunger, you need to understand what happens inside a data center. When thousands of processors run complex neural network computations simultaneously — as they do every time a large language model generates a response — they produce enormous amounts of heat. That heat has to go somewhere, or the hardware fails.
The most common solution is evaporative cooling: water is circulated through cooling towers, where it absorbs heat and then evaporates into the atmosphere. Unlike electricity, which leaves a carbon trail you can track and offset, that water simply disappears. It doesn’t return to a reservoir or a river. It’s gone.
Modern data centers rely on this process around the clock. The more computationally intensive the workload — and few workloads are more intensive than running a large AI model at scale — the more water the cooling system consumes.
—
The Numbers That Should Stop You Mid-Scroll
A 2023 study from researchers at UC Riverside and the University of Texas at Arlington put a precise number on this: generating 100 words with ChatGPT consumes approximately 519 ml of water. Let that sit for a moment. Most conversations with ChatGPT span multiple exchanges. A 10-turn back-and-forth? You’ve just evaporated over five liters.
Scale that to the platform’s user base — over 100 million active users — and the math becomes staggering.
Individual data centers compound the picture even further:
- A single large facility can consume up to 5 million gallons of water per day
- That’s enough to supply the daily water needs of a city of 30,000–50,000 people
- Microsoft’s global water consumption increased by 34% between 2021 and 2022, a period that coincided directly with its major AI infrastructure buildout
- Google’s water use surged by 20% in the same timeframe — and both companies have acknowledged AI workloads as a primary driver
The world’s largest cloud providers each operate dozens of these facilities. The cumulative draw is immense — and it’s accelerating.
—
A Global Problem With a Very Local Face
Here’s where the story turns from surprising to genuinely alarming: it’s not just how much water AI uses — it’s where.
Data centers cluster around cheap land, tax incentives, and reliable power grids. Many of the world’s largest concentrations sit in regions already flagged as high water-stress zones by the World Resources Institute’s Aqueduct tool — Mesa, Arizona; Northern Virginia; parts of the Netherlands; West Texas.
These are places where aquifers are already depleting, where farmers compete with municipalities for water rights, where “drought conditions” isn’t a weather event but a way of life. When a data center in drought-stressed Arizona pulls millions of gallons daily, that water isn’t coming from a surplus. It’s drawn from the same strained system that local communities and agriculture depend on.
The invisible water bill isn’t an environmental abstraction. It has a ZIP code. And in many cases, that ZIP code is already running dry.
—
What Can Actually Be Done
The good news: this isn’t an unsolvable problem. Researchers and policymakers are beginning to map a credible path forward. A roadmap aligned with Cornell University’s sustainability research frameworks points to several actionable levers:
- Smart geographic siting: Locating new data centers in low water-stress regions — Nordic countries, coastal areas with seawater cooling potential, or zones with abundant renewable water — can dramatically reduce freshwater drawdown without sacrificing performance
- Alternative cooling technologies: Air cooling, liquid immersion cooling, and closed-loop systems that recycle water rather than evaporate it are commercially available today — they’re more expensive upfront but represent genuine, measurable reductions in consumption
- Standardized water disclosure: Most AI companies don’t publish per-model or per-query water data. Pushing for reporting frameworks analogous to carbon disclosures would empower consumers, regulators, and investors to make informed decisions
- User and developer awareness: Some AI tasks are far more water-intensive than others. Knowing that generating a 1,000-word document carries a heavier footprint than a quick fact lookup is the first step toward more conscious use
—
The Conversation We’re Not Having
Carbon is countable. You can see it in emissions reports, offset it with a reforestation pledge, and feature it in a corporate sustainability deck. Water is harder to see — it evaporates, literally, into thin air.
But the next time you prompt an AI assistant, it’s worth remembering: that response didn’t just cost electricity. It cost fresh water — in a world where 2 billion people already lack access to safe drinking water.
The hidden water bill is real, it’s growing, and almost no one is reading it. It’s time we started.