The Truth Behind Data Center Water Usage

Brendan Nelson, PE
Aug 6
8 min read

Over the past year, it seems like every few weeks there's another headline about data centers and water. One article claims AI is "drinking rivers." Another suggests data centers are contaminating local drinking water. Politicians, media outlets, and social media have all weighed in, and as with many technical topics, the conversation has become increasingly simplified. As engineers, we know that almost every system is built around tradeoffs. There is rarely a perfect solution, only the best solution for a particular application. Cooling a data center is no different.

The goal of this article isn't to defend or criticize the data center industry. Instead, I'd like to explain how these facilities are actually cooled, where water is used, and why understanding the engineering is important before drawing conclusions.

Water Is Everywhere

Before talking specifically about data centers, it's worth taking a step back and remembering just how much water our society uses every day. The average American family uses 320 gallons per day, on all sorts of things, including watering grass. Annually, Americans use just over 3 trillion gallons of water just to water our lawns (1).

Agriculture is by far the largest consumer of freshwater in the United States. Millions of acres of crops are irrigated every growing season, including corn that is ultimately used for food, livestock feed, and ethanol production. Golf courses require significant irrigation to maintain healthy turf, over 750 billion gallons per year (2). Manufacturing facilities use water for countless industrial processes. Power plants use water to produce electricity. Hospitals, universities, airports, and commercial buildings all rely on water for plumbing systems and HVAC equipment.

Even constructing a building requires large quantities of water. Concrete production, dust control, and site work all consume water long before the building ever opens its doors. None of this is inherently good or bad. Water is simply one of the most useful resources available to engineers. The important questions aren't whether water is being used, but how it's being used. Is it being consumed? Is it being recycled? Is it discharged? Could another design reduce that usage? There is a big difference between water withdrawal vs water consumption.

Some articles estimate that large AI data centers use hundreds of thousands, or even millions, of gallons of water per day. While these figures may be technically correct depending on the cooling system and methodology used, the word "use" is often misleading. Many readers understandably assume that this water is consumed once and then discharged, when that simply isn't how most modern data centers operate. For example, one article (3) estimates that a single AI query uses 16 ounces of fresh water. At 10 million ChatGPT queries per day, that equates to approximately 160 million ounces, or 1.25 million gallons of water per day, roughly 456 million gallons per year. That is certainly a significant quantity of water, but it is still small compared to many other large-scale uses of water in the United States, such as residential lawn irrigation, agriculture, and golf course irrigation.

The modern AI data centers I have worked on utilize direct-to-chip cooling with closed-loop hydronic systems and air-cooled chillers. The cooling water, typically mixed with propylene glycol for freeze protection, remains inside the piping system and continuously circulates to remove heat from the servers before rejecting that heat outdoors through the chillers. It is not continually discharged or replaced; it performs the same job over and over again, much like the coolant in your vehicle or the chilled water loop serving a hospital or office building. This distinction is important because there is a significant difference between containing tens of thousands of gallons of water within a cooling system and consuming tens of thousands of gallons every day.

Take a car radiator: Every mile you drive, the cooling system "uses" coolant to carry heat away from the engine. That doesn't mean the coolant disappears. It stays inside the closed system, circulating continuously and being reused thousands of times over the life of the vehicle. Many modern data centers cool their servers in much the same way.

Why Are Data Centers Under Scrutiny?

As artificial intelligence continues to grow, so does the demand for computing power. Companies are building larger and larger data centers to support cloud computing, AI model training, and internet infrastructure. Along with that growth has come increased public attention. Some recent headlines have questioned whether data centers use excessive amounts of water. Others have suggested that nearby communities have experienced groundwater issues after large facilities were constructed. Representative Alexandria Ocasio-Cortez, for example, recently questioned EPA officials regarding reports of contaminated well water near a newly constructed data center in Georgia.

Those concerns deserve to be investigated. If a community believes its drinking water has been affected by nearby construction, engineers and regulators should determine exactly what happened. However, it's important to distinguish between construction impacts, groundwater management, and the cooling systems operating inside the completed building. Those are related topics, but they are not the same thing.

A large construction project may require dewatering, extensive excavation, stormwater management, or significant groundwater withdrawal depending on local conditions. Those activities have the potential to affect nearby wells and should always be carefully evaluated. That is very different from assuming the water circulating inside the completed data center somehow mixes with local drinking water. Essentially, those issues are construction related, not specifically data center construction related. Any building could have caused contaminated well water.

Why Do Data Centers Need Cooling?

At its core, a data center is simply a building full of computers. Every server consumes electricity, and nearly every watt of electricity eventually becomes heat. Unlike lighting or motors that perform mechanical work, servers spend almost all of their energy processing information, and that energy ultimately leaves the equipment as heat. A useful way to think about it is this: a 100 MW data center is also a 100 MW electric heater.

That heat has to go somewhere. If it isn't removed continuously, server temperatures rise rapidly, equipment performance begins to decrease, and hardware reliability suffers. Cooling isn't a luxury in a data center, it is one of the building's primary functions.

Why Engineers Love Water

If you've read HVAC Fundamentals #2: Heat, Air, and Thermal Comfort, you'll probably remember our discussion about why water is such an effective medium for transferring heat. Water has an unusually high specific heat, meaning it can absorb a tremendous amount of energy before its temperature changes significantly. It is dense, inexpensive, readily available, non-toxic, and relatively easy to pump through piping systems.

Compared to air, water can transport vastly more heat while moving a much smaller volume of fluid. This is why hospitals, universities, airports, manufacturing facilities, and countless commercial buildings all use hydronic systems. Data centers, in particular, have seen huge innovations and leaps in energy density. This energy density makes air a worse medium for cooling, and increases the need for water, as discussed later.

Not Every Data Center Is Cooled the Same Way

One of the biggest misconceptions in the current public discussion is the assumption that every data center uses the same cooling system. Historically, many large facilities relied on cooling towers. In these systems, water circulates continuously between the building and the cooling tower. Most of that water is reused over and over again, but a portion evaporates as heat is rejected to the atmosphere. That evaporation is not accidental; it is the mechanism that makes the cooling tower work. Cooling towers are extremely energy efficient, but they also consume water continuously through evaporation, not just during the initial system fill.

Most newer facilities take a different approach. Every large data center I've personally worked on has used air-cooled chillers with completely enclosed chilled water systems. In these facilities, water mixed with propylene glycol circulates continuously through a closed piping network. The fluid absorbs heat from the building, transfers that heat to the chillers, and then returns to repeat the process.

The important point is that the water remains inside the system. It is not continually discharged. It is not continually replaced. Much like the cooling system in your vehicle, the same fluid continues circulating throughout normal operation. While maintenance and occasional makeup water are still required, these systems do not consume cooling water in the same way that evaporative cooling systems do. That distinction is often missing from public discussions. Often times the water will come pre mixed from offsite in trucks to fill the system.

AI Is Changing the Way We Cool

Artificial intelligence has fundamentally changed the way engineers think about data center cooling. Not long ago, server racks commonly dissipated somewhere between 5 and 15 kilowatts of heat. Air cooling worked well because the amount of heat produced by each rack remained within the practical limits of what moving air could remove.

Today's AI hardware is very different. Modern AI racks routinely exceed 100 kW, and some next-generation systems are approaching several hundred kilowatts per rack. At those power densities, the challenge isn't simply installing larger fans or moving more air. The limitation is the air itself.

As we discussed in HVAC Fundamentals #2, air is actually a relatively poor medium for transporting thermal energy. It works extremely well for conditioning occupied spaces because it's already everywhere around us, but it simply cannot move heat as efficiently as water. We're beginning to reach the practical limit of what air alone can remove from modern computing equipment. That's why the industry is rapidly transitioning toward direct liquid cooling. Rather than cooling the room and removing hot air, coolant is delivered directly to the processors through cold plates. Heat is captured almost immediately, transferred into a liquid cooling loop, and ultimately rejected outside the building. Depending on the manufacturer, this may involve equipment such as CRAC units, CRAH units, CDUs, or CCUs, but the overall objective remains the same: move enormous quantities of heat efficiently and reliably.

A Real-World Example

One of the data centers I was involved with consisted of 22 individual data halls. Each hall supported approximately 10 MW of IT load. To put that into perspective, each data hall rejected roughly 34 million BTU per hour of heat. Each hall also contained approximately 6,500 gallons of water and another 3,750 gallons of a 36% propylene glycol solution as part of its cooling system.

Across ten operating halls, that represents roughly 65,000 gallons of water dedicated to cooling the compute equipment alone, not including the administrative portions of the building or domestic plumbing systems. That sounds like an enormous amount of water, but that water wasn't flowing down a drain every day. It wasn't being dumped into a river. It wasn't disappearing. It remained inside the closed cooling system, circulating continuously as it transported heat from the servers to the chillers. From an engineering standpoint, that isn't different than the chilled water systems found in hospitals, universities, airports, or many large commercial buildings.

The Real Engineering Challenges

None of this is meant to suggest that data centers are free from criticism.

Large AI campuses require enormous electrical infrastructure. They can create noise from chillers, generators, and transformers. They occupy significant amounts of land, require extensive utility coordination, and represent major construction projects with all of the environmental considerations that accompany projects of that size.

Those are real engineering challenges deserving of thoughtful discussion. Water usage also deserves careful consideration, particularly when evaluating evaporative cooling systems in regions where water resources are limited. Engineers constantly balance energy efficiency, first cost, operating cost, reliability, maintainability, and environmental impact. Choosing a cooling system is rarely as simple as maximizing one variable.

Final Thoughts

The purpose of this article isn't to convince anyone that data centers are perfect. They aren't. Instead, I hope it demonstrates that discussions surrounding water use are far more nuanced than many headlines suggest. Some facilities use cooling towers and consume water through evaporation. Others rely almost entirely on closed-loop chilled water systems. Some are transitioning toward direct liquid cooling because air is reaching its practical limits as a heat transfer medium.

These are engineering decisions driven by physics, climate, economics, and reliability, not by politics. As engineers, we have a responsibility to explain these systems accurately. Public conversations about AI infrastructure, sustainability, and water resources are important, but they should begin with a clear understanding of how these facilities actually operate. Only then can we have meaningful discussions about the tradeoffs involved and focus our attention on solving the challenges that truly matter.

Thanks for reading, and stay curious

-BN

TL;DR: Data centers use water, but not nearly as much as many other common processes in the United States. More importantly, the word "use" requires nuance. Many modern AI data centers rely on closed-loop cooling systems, where water continuously circulates to transport heat rather than being consumed and discharged. Understanding the difference between water contained in a system and water consumed by a system is key to understanding the discussion around AI and water usage.