How Much Water Does an AI Data Center Really Use?
CONSERVATION JULY 5, 2026
At A Glance (TL;DR)
A large AI data center can use up to 5 million gallons of water a day — enough for a town of 50,000 people — thanks to an old cooling method called evaporative cooling. But a real shift is underway. The newest AI chips run so hot they require a fundamentally different approach: liquid cooling, which circulates coolant in sealed loops and can operate with near-zero fresh water consumption. Energy demand from data centers is still rising fast and won't slow down soon. Water, though, is a different story — and the trajectory is better than most people realize.
A large AI data center can use up to 5 million gallons of water a day — enough to supply a town of up to 50,000 people.
That water has been needed for the cooling systems that keep servers from overheating, and this technology has been industry standard for decades. But there’s a shift underway. The same AI hardware boom driving that water demand is also forcing a fundamental rethink of how data centers cool themselves — and for once, the economics and the environment are pointing in the same direction.
Why Data Centers Use So Much Water
Most data centers built before 2020 use evaporative cooling, the same principle as sweating. Warm air from the servers passes over water, the water evaporates and carries heat away, and cooled air cycles back in. It works. It is relatively cheap to build. And it burns through fresh water continuously, at massive scale.
Two-thirds of data centers built since 2022 are located in water-stressed regions, primarily Texas and Arizona. In the Phoenix area alone, data center water use is projected to jump 870% as planned facilities come online, reaching nearly double what the entire city of Flagstaff currently uses.
Servers generate enormous heat, and water evaporation has been the most efficient way to manage it at scale. Until recently.
The Shift Happening Right Now: Liquid Cooling
The industry is moving toward liquid cooling and not primarily because of environmental pressure. The newest AI chips simply run too hot for air or evaporative systems to handle.
Think of it like the radiator in your car. Instead of blowing air over hot components, liquid cooling circulates a coolant like water, refrigerant, or a dielectric fluid directly through cold plates mounted on the server hardware. The heat transfers into the liquid, which carries it away to a heat exchanger, cools down, and circulates back in.
The critical difference from evaporative cooling: in a properly designed closed-loop liquid system, that coolant does not evaporate. It recirculates in a sealed circuit. In theory, you can run an entire data center this way with near-zero fresh water consumption.
This is sometimes called a water-free or closed-loop data center and they are no longer prototypes. Microsoft, Google, and Meta are actively deploying liquid cooling in new builds, and in some cases retrofitting existing facilities.
Did Nvidia Accidentally Solve the Water Problem?
Sort of.
Nvidia's GB200 NVL72 — the current gold standard for AI training infrastructure, generates more than 120 kilowatts of heat per rack. Nvidia mandates liquid cooling for the system, with strict specifications for coolant flow rate, inlet temperature, and pressure. Deviation from those specs triggers automatic throttling that can reduce AI performance by 60%. There is no air-cooled alternative.
Operators who want to run cutting-edge AI have no practical choice but to build liquid cooling infrastructure. When the dominant chip vendor's product requires a fundamentally different cooling approach, the entire supply chain adapts. Facility designs, power delivery, and thermal management are all being rebuilt around liquid cooling as the default.
Nvidia did not set out to solve a water crisis. The performance demands of modern AI just happened to make the water-intensive approach obsolete.
What About Energy? That Is a Harder Story.
Water may be on a better trajectory. Energy is more complicated.
According to the Lawrence Berkeley National Laboratory's 2025 data center energy report, total U.S. data center electricity consumption is projected to reach 325 to 580 terawatt-hours by 2028, up from 176 TWh in 2023 — as much as 12% of all U.S. electricity consumed nationally. The DOE's own assessment echoes these findings. A single hyperscale facility already uses roughly 100 megawatts, equivalent to powering 100,000 homes.
Efficiency gains in chips get absorbed by expanded demand. More capable AI requires more training runs. More users generate more queries. The math does not net out to less energy — it nets out to more, faster.
There’s no question that AI's energy footprint is going to be large. Utilities, grid planners, and policymakers are actively working on nuclear, renewables, and grid expansion but there is no clean resolution in the near term. What consumers can realistically influence is pressure on operators to source that energy from clean sources, and to make consumption data public and verifiable.
Dr. Michael Webber, an energy expert and mechanical engineering professor at UT Austin, said in a recent PBS interview: “We don't know which way it's going to go right now. Are the data centers and hyperscalers…going to invest to make electricity more affordable for all of us? Or are they going to compete with us for electricity and drive our rates up?”
Your Questions, Answered
How are data centers reducing water use?
The main shift is from evaporative cooling, which continuously consumes fresh water, to liquid cooling systems that circulate coolant in closed loops. Some operators also use recycled water, harvested rainwater, or air-cooled heat exchangers in cooler climates. New facilities in water-stressed regions are increasingly being designed with liquid cooling as a baseline requirement rather than an upgrade.
What is liquid cooling for data centers?
Liquid cooling removes heat from servers by circulating a coolant directly through or adjacent to the hardware. It is far more efficient than air cooling and, in a closed-loop system, does not require continuous fresh water input. The coolant recirculates rather than evaporates, making it viable in regions where water scarcity would otherwise rule out large data center operations.
What is a water-free or closed-loop data center?
A closed-loop data center uses sealed cooling circuits that do not release water vapor into the atmosphere. Heat transfers to the coolant and is then dissipated through air-cooled exchangers or similar dry methods — no fresh water drawn from local supplies. These facilities can operate indefinitely in water-scarce regions with minimal environmental water impact.
Did Nvidia solve AI's water problem?
Not intentionally. But Nvidia's GB200 hardware generates more heat per rack than any previous generation, making liquid cooling a hard requirement. By making liquid cooling non-negotiable for performance, Nvidia accelerated an industry-wide shift away from water-heavy systems faster than environmental regulation likely would have.
What can I do about data center water use?
Individual query volume has minimal direct impact. The more meaningful lever is collective: support companies that publicly disclose water usage, favor services with verifiable renewable energy and water recycling commitments, and pay attention to data center development in your region — particularly if you live in a water-stressed area.
The Bigger Picture
You interact with AI infrastructure every day — every search, video stream, cloud backup, and chatbot query runs through it. Understanding what that actually requires is not about guilt. It is about being informed.
In his PBS interview, Dr. Webber says the costs of noise, water, and electricity are showing up in people's backyards now, while the payoff like better cancer treatment, safer roads, a smarter grid arrives later, and often for someone else. “We feel the impacts today, the benefits come later. That mismatch in time is a huge challenge.”
He framed the real task as a collective one, on society and everyone developing AI: “Is this a tool we can live with? If so, do we need rules? What are those rules going to be to ensure it's safe for our children and everybody else?”
The water story is genuinely getting better. The technology is moving in the right direction, the incentives are aligning, and new facilities are being designed with water efficiency as a requirement rather than an afterthought. The energy story will take longer. But an informed public is part of what moves that story too.
Sources:
