The water bottle is a useful lie.
It’s not inaccurate — a 100-word ChatGPT email does consume roughly 519 milliliters of water, per peer-reviewed research from Pengfei Li, Shaolei Ren, and colleagues at UC Riverside. That’s about one standard bottle. The comparison works because it makes an invisible industrial cost feel graspable. Users don’t see cooling towers. They see a text box.
But framing AI’s water problem around individual prompts is a bit like measuring ocean pollution by the straw. The bottle is the entry point. The actual story is the infrastructure behind it — and where that infrastructure is now being built.
One Bottle. Multiply by a Billion
ChatGPT crossed 1 billion monthly active app users in May 2026 — the fastest any app in history has reached that milestone, according to Sensor Tower data reported by Reuters. It also carries 900 million weekly active users, a figure OpenAI confirmed in February, and which has more than doubled in the 12 months prior. That’s not the only AI product people use daily. ChatGPT is one of dozens now embedded in search, office software, customer service pipelines, coding assistants, and personal productivity apps. AI water consumption in data centers doesn’t scale linearly with user counts — it scales with compute load, which grows faster still.
The UC Riverside team estimates global AI demand could require between 4.2 and 6.6 billion cubic meters of water withdrawal annually by 2027. A separate framing puts that at roughly 1.1 to 1.7 trillion gallons per year — close to half the United Kingdom’s total annual water withdrawal. And the 519-milliliter figure per email assumes a single response, not a conversation. Ten to fifty exchanges in a sustained session consume water at roughly the same per-exchange rate. The bottle quickly becomes a shelf of bottles.
Google’s 2024 Environmental Report puts the company’s total water consumption at approximately 8.1 billion gallons across global operations, with data centers accounting for about 95% of that. The 2024 figure was higher than 2023, which was higher than 2022, which was higher than 2021. The company names AI workload growth as the primary driver each time. Microsoft and Meta have disclosed similar year-over-year acceleration. Amazon, which operates the world’s largest cloud infrastructure, publishes no aggregate water figures.
The Three Pipes Nobody Talks About
The per-prompt number captures only one of three water pathways.
Server cooling is the visible one. AI chips run hot, and most data centers still rely on evaporative cooling — water enters the system, absorbs heat, and leaves as vapor. The Lawrence Berkeley National Laboratory’s 2024 Data Center Energy Usage Report estimated that US data centers consumed approximately 17.4 billion gallons through direct cooling in 2023. That number is striking. The indirect figure — water consumed by the power plants supplying electricity to those same facilities — was approximately 211 billion gallons. Twelve times larger than the direct figure. As AI workloads intensify, both scale in proportion.
Chip manufacturing is the third pipe. Advanced AI accelerators require ultra-pure water throughout fabrication. A data center doesn’t report that in its cooling disclosures, but a fab supplying its chips may sit in the same drought-stressed basin and draw from the same groundwater. The full supply chain picture, when counted together, is considerably larger than what any single company’s sustainability report captures.
Understanding exactly how much water ChatGPT uses across each of these pathways requires accounting for data center location, cooling technology, grid energy mix, and model size — which is why the per-query estimates across different studies vary so widely.
They’re Building in the Desert. On Purpose.
The location problem may be the sharpest part of this story.
A Guardian analysis found that roughly two-thirds of the 809 data centers currently planned across the US sit on land classified as drought-stricken over the past year, based on NOAA’s National Integrated Drought Information System. Bloomberg reached the same conclusion independently: approximately two-thirds of data centers built since 2022 have been placed in high water-stress regions. Five states account for 72% of facilities in the most strained areas.
Arizona and Texas dominate both lists. Both states have attracted significant data center investment for obvious reasons — cheap land, available power, and permitting speed. Both states have also faced persistent drought conditions and long-term groundwater depletion. The water a data center consumes in Phoenix doesn’t come from somewhere with slack capacity. It competes with agriculture, municipal supply, and communities already managing a shortage.
MSCI’s geospatial analysis of roughly 14,000 data center assets worldwide estimates that about one in four existing facilities may experience more frequent water-scarcity periods by 2050 — particularly in Chile, Brazil, Mexico, Turkey, and Australia, where rapid digital buildout is intersecting with worsening climate conditions.
The Pledge Problem
Tech companies have responded to water criticism primarily with two instruments: commitments and offsets.
Google and Microsoft have both announced goals to become “water positive” by 2030, meaning they’d replenish more water than they consume. Google disclosed replenishing approximately 4.5 billion gallons in 2024, while also reporting that its total consumption nearly doubled between 2021 and 2024. Replenishment and consumption are not the same accounting line, and critics point out that restoration projects routinely operate in geographies different from the watersheds being depleted. A company can restore water in a coastal watershed while a community near its Arizona campus still competes for dwindling groundwater.
There’s no federal regulation requiring AI companies to disclose water consumption figures. No standardized metric exists. The result is that corporate sustainability reports use different methodologies, cover different parts of the supply chain, and are largely unverifiable by outside parties.
Efficiency Exists. The Question Is Whether It Scales Fast Enough.
The industry isn’t ignoring the problem technically. Closed-loop liquid cooling and direct-to-chip cooling systems eliminate most evaporative water loss. Google has disclosed that each Gemini prompt averages approximately 0.26 milliliters of water, orders of magnitude less than the UC Riverside estimates for GPT-4. The difference reflects newer cooling architecture, more efficient chips, and a different scoping methodology. It’s a real improvement, not a PR number.
But efficiency gains are being swamped by volume growth. Data center capacity is expanding faster than cooling technology is being retrofitted. The facilities coming online in the next two years were designed and permitted years ago, largely with conventional evaporative systems.
The water story is ultimately an infrastructure planning story, not a consumer behavior story. Asking users to write shorter prompts won’t move the macro numbers. The relevant decisions sit with data center siting, cooling system procurement, grid decarbonization, and — eventually — regulatory disclosure requirements that don’t currently exist.
The bottle made the problem legible. The pipeline behind it is what actually needs solving.
Related: Is Character AI Bad for the Environment? The Hidden Carbon Cost of Chatbots
