The Thirst of AI: Why Trusted Water Metrics Are the Next Frontier for Data Centers

We picture the cloud as something ethereal — weightless data moving through the air. The physical reality of AI is the opposite: it’s anchored to the ground, drawing on local power grids and, just as critically, on local water. Power dominates the headlines. But water is the hidden ingredient keeping the AI revolution from overheating — and right now, almost no one can tell you exactly how much of it AI actually uses.

That last part is the real problem. As AI compute scales, trusted, verifiable water metrics have gone from a niche sustainability checkbox to a strategic necessity — for operators, for the communities hosting these facilities, and for the industry’s license to keep building.

The Hidden Water Footprint of AI

High-performance AI servers run extremely hot, and they have to be cooled or they fail. Operators face a trade-off: air cooling, which is water-light but electricity-heavy, or evaporative cooling, which is far more energy-efficient but consumes large volumes of fresh water. As racks have gotten denser, more facilities have leaned on water.

The scale is real. According to the Lawrence Berkeley National Laboratory’s 2024 assessment, U.S. data centers directly consumed about 17.4 billion gallons of water in 2023 and consumed an additional 211 billion gallons indirectly, through the electricity that powered them. A single large facility can consume up to 5 million gallons a day — on par with a small town. Researchers at UC Riverside have estimated that a 100-word email generated with GPT-4 can “drink” roughly a bottle of water (about 519 milliliters) once cooling and electricity are both counted.

That last figure is useful for intuition, but treat it the way you’d want all AI water numbers treated: as an estimate with assumptions. It bundles on-site and off-site water and swings dramatically with a data center’s location and season; a query served in cool, hydro-powered Oregon and one served in the Arizona desert are not the same. Which is exactly the point of this post. The headline numbers are big, but the methods behind them are murky, and most corporate environmental reports still don’t separate AI workloads from everything else. Today, it’s nearly impossible to state the true water cost of training and running a model with confidence.

First, Define What You’re Measuring: Withdrawn vs. Consumed

Before any metric can be trusted, one distinction has to be clear — and most coverage skips it. Withdrawal is the total water a facility takes from a source. Consumption is the portion it never gives back, lost mainly to evaporation in cooling. The water that is returned, called blowdown, comes back more concentrated than it left.

This matters because the two numbers can differ by an order of magnitude, and reporting one while implying the other is how well-intentioned disclosures mislead. A serious water metric has to state both, at the level of the individual site. Anything less isn’t measurement — it’s marketing.

The Metric Maze: The Problem with WUE

The industry standard is Water Usage Effectiveness (WUE), introduced by The Green Grid in 2011, which divides the liters of water a facility consumes by the kilowatt-hours of energy its IT equipment uses. WUE is a genuinely useful tool — but on its own, it can hide more than it reveals.

Averaging masks impact. A company can publish one blended, enterprise-wide WUE that mixes water-hungry evaporative sites with zero-water air-cooled ones. The average looks fine while a single facility quietly stresses one local watershed. Water is a local resource; a global average is the wrong unit.
Most operators don’t even track it. Older industry surveys found that fewer than a third of U.S. data center operators measure their water use at all — and current data suggests little has changed. When Texas asked data centers to report water usage, only about a third responded in 2024, and just 17% in 2025, despite the survey being mandatory. You cannot verify what is never recorded.
Embedded water is ignored. Basic WUE counts only the water used on-site. It excludes the water consumed to generate the facility’s electricity — the 211 billion gallons noted above, more than ten times the direct figure — and the water used to manufacture the chips inside. A single semiconductor fab can consume on the order of 10 million gallons of ultrapure water a day, roughly the daily use of 33,000 households. None of that shows up in a typical WUE number.

Why Verifiability Is a Strategic Imperative

Without trusted, standardized reporting, managing AI’s water demand becomes guesswork. Verifiable metrics matter for three reasons:

Protecting local watersheds. Data centers act like giant soda straws, pulling from highly localized sources. Regional planners can only balance economic development against water infrastructure if they have accurate, site-level data to plan from.
Community accountability. Blowdown discharge is more concentrated than the intake water, carrying elevated dissolved solids like calcium and chloride along with treatment chemicals. Depending on the volume, the treatment, and the receiving watershed, that discharge can affect drinking-water taste, stress aquatic ecosystems, and reduce agricultural yields. Transparent reporting keeps the burdens of the AI boom from landing quietly on the communities least able to absorb them.
Making efficiency targets real. You can’t manage what you don’t measure. Aggressive, verifiable WUE targets — set against local climate and balanced against power use — are what actually force operators to adopt closed-loop cooling, immersion cooling, blowdown reuse, and non-potable water sources. Without trustworthy measurement, an efficiency goal is just a press release.

What “Trusted” Actually Requires

Here’s where most of the conversation stops short. The fix for a measurement problem is not better cooling — that’s a result of good metrics, not a substitute. A trustworthy water-accounting regime needs four things: a standardized framework that reports both withdrawal and consumption, at the individual-site level rather than the corporate average; AI-specific disclosure, so the footprint of model training and inference can be separated from ordinary IT load; third-party verification, so the numbers are audited rather than self-attested — the same way financial and carbon figures increasingly are; and watershed-level context, because 5 million gallons a day means one thing beside the Columbia River and something very different in the high desert.

This is the unglamorous infrastructure of trust — and it’s the precondition for every efficiency technology that follows.

The Bottom Line

AI’s water story is really the same story as its power story: demand is scaling faster than the systems we use to account for it. The industry can no longer treat water as unlimited and untracked. Building a durable AI ecosystem — one that communities will actually permit and expand — depends on absolute transparency. Standardized, independently verified water metrics aren’t good PR. They are the foundational data required to scale AI without drinking our most precious resource dry.

At AuerBridge, we sit at exactly this intersection of power, water, data centers, and mega-manufacturing. No matter the scale, we can deliver valuable research and insights. Research for your firm is available here. If you’re planning the next step in infrastructure for your organization and want to pressure-test the power, water, or mega-manufacturing side before you commit, let’s talk.

Sources: LBNL 2024 / Berkeley Lab (17.4B direct, 211B indirect), UC Riverside “Making AI Less Thirsty” (519 mL), Texas mandatory water-survey response rates, npj Clean Water (<1/3 operators track), cooling-tower blowdown water quality, WEF — semiconductor/fab water use.