TITLE: AI’s Water Bill Is Smaller Than You Think, For Now
—
Your Dishwasher and AI Have More in Common Than You’d Expect
Running a load of dishes uses about 6 gallons of water. Running a query through a large language model uses a few milliliters — roughly the amount you’d flick off your fingertips after washing your hands. That comparison won’t hold forever, but right now, in 2025, the public narrative around AI and water is running ahead of the actual numbers. And as someone who spends his days testing AI tools and cutting through the hype, that gap bothers me.
I review AI toolkits for a living. I care about what these systems actually cost — in dollars, in compute, and yes, in resources. So when I started seeing headlines treating AI data centers like they were draining the Colorado River dry, I went looking for the real picture. What I found was more complicated, and honestly more interesting, than the panic suggested.
What the Numbers Actually Say
The global AI economy currently consumes around 23 cubic kilometers of water per year. That sounds enormous until you put it next to agriculture, which accounts for roughly 70% of all freshwater withdrawals worldwide. AI’s share of global water use is real, but it is not the runaway crisis some coverage implies — at least not yet.
The “not yet” part matters a lot. That same 23 cubic kilometers is projected to grow by 129% by 2050, pushing consumption past 54 cubic kilometers annually. That trajectory is serious. Microsoft has already acknowledged internally that its data center water use is expected to more than double as AI infrastructure scales up. U.S. data center water consumption alone could reach between 150 and 280 billion gallons by 2028, according to analyst estimates.
So the story is not that AI’s water footprint is fine. The story is that the public is simultaneously overestimating the present and underestimating the future — and that confusion is making it harder to have a useful conversation about what to actually do.
Why the Misread Happens
Part of the problem is how water use in data centers gets reported. Cooling systems are the main culprit — they use water to regulate server temperatures, and the amount varies wildly depending on the facility’s location, age, and design. A data center in a hot, dry region uses far more water for cooling than one in a cooler climate that can rely on outside air. Aggregating these numbers into a single global figure flattens all of that nuance.
There’s also a motivated reasoning problem on both sides. AI backers have a clear incentive to downplay environmental costs. Critics have an incentive to amplify them. Neither group is giving you a straight read. As a toolkit reviewer, I see this dynamic constantly — vendors oversell efficiency gains, and skeptics overstate harm. The truth tends to live somewhere less dramatic.
The Part Nobody Talks About
Here’s what actually surprised me in my research: AI is also being used as a tool for water conservation. Recent studies point to real applications — optimizing irrigation systems, modeling drought patterns, detecting leaks in municipal water infrastructure. The same technology that consumes water at scale is being deployed to reduce waste in sectors that consume far more of it.
That’s not a reason to ignore AI’s footprint. But it does complicate the simple “AI bad for water” framing. If a water utility uses an AI system that cuts its leakage rate by 15%, the net water math looks very different than if you only count what the data center used to run the model.
This is the kind of tradeoff analysis that almost never makes it into the headlines, but it’s exactly the kind of thing that should inform how we build and regulate these systems going forward.
What I Actually Want From This Industry
As someone who reviews AI tools, I want vendors to be straight about resource costs. Not buried in a sustainability PDF nobody reads — front and center, the way energy ratings appear on appliances. Water consumption per query, per model run, per API call. Give me the number. Let me factor it in.
The 129% projected increase by 2050 is not inevitable. It’s a projection based on current trajectories. Trajectories change when there’s pressure to change them — from regulators, from customers, from journalists who don’t let vendors off the hook.
AI’s water footprint is underestimated in the future and overestimated in the present. Getting that distinction right is how we build systems that are actually worth using — and worth defending.
🕒 Published: