AI’s Water Problem Is More Complicated Than It Looks - Data Center Water Use

The real question is not simply how much water a data center uses. It is where, when, why and what kind.

The AI water conversation usually begins with one very large number.

Billions of gallons.

The number travels quickly through headlines, community meetings and social media. Before long, every data center looks like a giant industrial straw pointed at the local water supply.

The reality is more complicated.

Some data centers use water for cooling every day. Others use it only during the hottest hours of the year. New facilities may circulate cooling liquid through closed systems without continuously consuming fresh water.

Water can also be used indirectly by the power plants generating the facility’s electricity. It is needed to manufacture semiconductors. It may come from drinking-water systems, reclaimed wastewater, groundwater or several sources combined.

A gallon used in a water-stressed watershed carries a different risk than a gallon used where supply is abundant and renewable.

That does not make AI’s water demand unimportant.

It makes the details more important than the headline.

The Short Answer

AI data centers use water directly for cooling and humidification and indirectly through electricity generation and semiconductor manufacturing.

The amount varies significantly based on the cooling system, climate, computing density, operating practices and source of electricity.

Water risk cannot be evaluated through total gallons alone. Leaders must also examine whether the water is potable or reclaimed, how much is returned after use, when peak demand occurs and whether the local watershed can support the facility during drought and extreme heat.

The right question is not:

“How much water does AI use?”

It is:

“How much water will this facility consume, from which source, in which watershed and under what conditions?”

The Signal

In 2023, U.S. data centers directly consumed approximately 66 billion liters of water—roughly 17.4 billion gallons—according to Lawrence Berkeley National Laboratory.

Hyperscale and colocation facilities accounted for 84% of that total. By 2028, hyperscale data centers alone are projected to consume between 60 billion and 124 billion liters annually, depending on growth and cooling-system choices.

But direct consumption is only part of the picture.

The same report estimated that U.S. data centers had an indirect water footprint of nearly 800 billion liters in 2023 because many power plants consume water while generating electricity. Lawrence Berkeley National Laboratory

That indirect footprint was approximately 12 times larger than the water consumed at the facilities themselves.

The figure varies by regional electricity mix and does not account for every facility’s individual power agreement or behind-the-meter generation. But it illustrates why the AI water conversation cannot stop at the cooling tower.

Water and power are connected.

Reducing one resource at the facility can sometimes increase demand for the other somewhere else.

What Is the Difference Between Water Withdrawal and Water Consumption?

Water terminology is often used loosely.

That creates confusion.

Water Withdrawal

Water withdrawal is the total amount drawn from a source, including municipal systems, rivers, lakes and groundwater.

Some withdrawn water may eventually be treated and returned.

Water Discharge

Water discharge is the portion returned to a municipal system, watershed or other approved destination after use.

Its quality and temperature may be regulated.

Water Consumption

Water consumption is the water withdrawn but not returned to the immediate water system.

In data-center cooling, this is often water lost through evaporation.

AWS uses the same distinction in its reporting: withdrawal represents the total amount drawn, while consumption is the amount not returned, generally because it evaporated. AWS

These measurements should not be treated as interchangeable.

A facility could withdraw a large amount of water and return most of it. Another could withdraw less but consume a larger share through evaporation.

The headline number may look similar.

The effect on the watershed may not be.

Why Do AI Data Centers Need Water?

Every computation produces heat.

Traditional servers produced enough heat to require substantial cooling. AI processors increase the challenge because more computing power is being concentrated into each rack.

The higher the density, the more heat must be removed from a relatively small physical space.

Cooling systems generally move that heat through a combination of air, water, refrigerants and specialized liquid coolants.

Water May Be Used To:

• Remove heat from cooling equipment

• Support evaporative cooling towers

• Circulate through chilled-water systems

• Maintain temperature and humidity

• Carry heat away from processors

• Produce the electricity powering the facility

Water is not the product being created.

It is part of the physical process allowing the product to exist.

Liquid Cooling Does Not Always Mean High Water Consumption

The term “liquid cooling” can create the impression that a data center continuously pours water over its processors.

That is not how most systems work.

In direct-to-chip cooling, a coolant circulates through cold plates attached to high-temperature components. The liquid absorbs heat and moves it away from the processor.

The fluid can operate inside a closed loop.

Once the system is filled, the same liquid circulates repeatedly. Water does not have to be continuously withdrawn or evaporated inside the server.

The question becomes how the facility rejects the captured heat.

The System May Use:

• Cooling towers

• Dry coolers

• Mechanical chillers

• Outside air

• A closed-loop heat exchanger

• A hybrid combination of these technologies

A liquid-cooled server can therefore be connected to either a water-consuming or low-water facility cooling system.

“Liquid cooled” describes how heat leaves the processor.

It does not tell us how heat ultimately leaves the property.

The Cooling Trade-Off

Water-based evaporative cooling can use less electricity than mechanical air cooling.

Water absorbs and transfers heat efficiently. Evaporation can reduce the amount of energy needed to cool a facility during hot conditions.

But the process consumes water.

Air-cooled and dry-cooling systems can dramatically reduce on-site water use. They may also require more electricity, particularly during hot weather.

That Creates a Trade-Off:

Use more water to reduce cooling electricity or use more electricity to reduce on-site water.

The correct decision depends on climate, water availability, grid conditions, operating costs and community priorities.

A cooling system that makes sense in northern Europe may not make sense in Arizona.

A design optimized for a water-rich region may become difficult to justify in a watershed facing long-term drought.

There is no universally sustainable cooling system.

There are systems that fit their local conditions better than others.

Zero-Water Cooling Is Becoming Real

In 2024, Microsoft introduced a data-center design optimized for AI workloads that consumes no water through cooling evaporation.

The system uses chip-level cooling and circulates water continuously through a closed loop after it is initially filled. Microsoft estimates the design can avoid more than 125 million liters of water use each year at an individual data center.

The company has also acknowledged the trade-off. Replacing evaporative cooling with mechanical systems creates a nominal increase in annual electricity use, although warmer cooling temperatures and more efficient chillers help reduce the difference. Microsoft

Microsoft’s first projects using the new design include facilities in Phoenix, Arizona, and Mt. Pleasant, Wisconsin, with the data centers expected to begin coming online in late 2027.

That matters because Phoenix is exactly the type of market where water concerns can influence community acceptance.

Zero-water cooling will not eliminate every environmental impact.

It does show that data-center water demand is a design decision - not an unavoidable fixed number.

Reclaimed Water Changes the Equation

Not every cooling system requires drinking-quality water.

Reclaimed water is wastewater that has been treated and reused for non-potable purposes. It can provide a cooling source while preserving drinking water for homes, hospitals and other community needs.

AWS reported in July 2026 that it was using reclaimed water at 26 data centers and planned to expand the practice to more than 130 locations across nine jurisdictions.

The company works with utilities to develop dedicated reclaimed-water infrastructure, sometimes referred to as purple-pipe systems. AWS also said it does not use water for cooling in water-stressed regions including Phoenix, the Middle East and South Africa. AWS

Reclaimed water can create benefits for both the data center and the utility.

The facility reduces its dependence on potable water. The utility gains a customer for treated wastewater that might otherwise be discharged.

But reclaimed water is not automatically available.

Communities May Need:

• Additional treatment capacity

• Dedicated pipelines

• Storage systems

• Pumping infrastructure

• Quality monitoring

• New permits

• Capital for construction

A purple pipe on a planning document is not the same as reclaimed water available at the site.

The water must still be treated, transported and delivered.

Water Usage Effectiveness

Water Usage Effectiveness, or WUE, is a common data-center efficiency metric.

It generally measures the amount of water used for cooling and humidification relative to the electricity consumed by the computing equipment.

A lower WUE indicates that the facility uses less water per unit of computing energy.

The metric is useful.

It is not complete.

WUE May Not Tell Stakeholders:

• Whether the water is potable or reclaimed

• Whether the watershed is water-stressed

• When the water is consumed

• How much is used during extreme heat

• How much water is consumed indirectly through electricity

• Whether consumption increases as the campus expands

• How much water is returned to the local system

• Whether replenishment occurs in the same watershed

A facility with a low WUE can still create local concerns if it relies on drinking water during drought.

A facility with a higher WUE may have less community impact if it uses treated wastewater in a water-abundant region.

Efficiency metrics need geography.

The Water-Power Connection

The Power and Water pillars cannot be evaluated separately.

Thermoelectric power plants may consume water for cooling. Hydroelectric reservoirs lose water through evaporation. Natural-gas plants, nuclear facilities and other generating resources have different water requirements.

That means a data center using little water on-site may still have a meaningful indirect water footprint through its electricity supply.

The reverse can also be true.

A carefully designed evaporative cooling system may consume more water at the campus but reduce electricity demand enough to lower water use at power plants.

This is why focusing exclusively on site-level WUE can create the wrong conclusion.

Leaders should evaluate:

Direct facility water + indirect electricity water = the fuller water footprint

The ideal balance will vary by region.

Water planning is also power planning.

Water Is Local

Carbon emissions can influence the global atmosphere regardless of where they occur.

Water operates differently.

Water stress is shaped by the local watershed, aquifer, climate and competing demand.

One million gallons consumed in a region with abundant renewable supply may create limited stress. The same amount withdrawn from a declining aquifer during a drought can become a major community issue.

Timing matters as well.

A facility may report modest annual consumption while placing significant pressure on a water system during the hottest weeks of the year, exactly when residential and agricultural demand is also elevated.

This leads to the Watershed Readiness Principle:

Data-center water risk must be evaluated at the watershed level and under peak conditions not through global annual averages alone.

Communities need to understand what happens during the difficult year, not only the average one.

Water Positive Does Not Mean Water Neutral Everywhere

Several hyperscalers have committed to becoming water positive.

This generally means returning or replenishing more water than the company consumes through a combination of efficiency, reuse and community projects.

Those Projects Can Include

• Watershed restoration

• Agricultural efficiency

• Leak reduction

• Aquifer recharge

• Wetland restoration

• Wastewater reuse

• Improved community water access

These programs can create real value.

But leaders should still examine where and when replenishment occurs.

A project restoring water in one watershed does not physically replace water consumed in another. A replenishment project completed years later may not address current pressure on a municipal system.

Water-positive commitments should therefore be evaluated alongside local operating data.

The global target matters.

The local water balance matters more to the host community.

The Community Question

Water often becomes the point where a technical infrastructure project turns into a public-trust issue.

Residents may be asked to conserve water while a large industrial facility requests a new allocation. Agricultural users may already be operating under restrictions. Utilities may need to expand treatment plants, pipelines or wastewater systems.

Even an efficient data center can face opposition when information is incomplete.

Communities Should Expect Clear Answers About:

• The proposed water source

• Average and peak consumption

• Potable versus reclaimed water

• Drought-year operations

• Municipal infrastructure requirements

• Public financing and incentives

• Water discharge and treatment

• Expansion plans

• Facility-level reporting

• Local replenishment commitments

Trust cannot be built through a global sustainability report alone.

It has to be built at the watershed level.

The Market Is Responding

Water management is becoming part of the AI infrastructure supply chain.

In 2025, Ecolab introduced monitoring technology for direct-to-chip liquid cooling that tracks coolant temperature, pH and flow rates in real time. The system is designed to help operators protect computing equipment while improving cooling performance and resource efficiency. Ecolab

The company then made a much larger move.

In March 2026, Ecolab announced an agreement to acquire CoolIT Systems for approximately $4.75 billion. CoolIT develops coolant-distribution units, cold plates and direct-to-chip cooling technology for high-density AI systems.

Ecolab said the acquisition would create a more complete data-center cooling platform combining thermal engineering, water treatment, chemistry, digital monitoring and on-site service. The transaction was expected to close during the third quarter of 2026, subject to approvals. Ecolab

That is not simply a water company buying a cooling company.

It is a bet that fluid management will become mission-critical AI infrastructure.

Company Watch: Ecolab

Everyone is watching who makes the chip.

I’m watching who keeps it cool.

Ecolab has traditionally been associated with industrial water treatment, sanitation and resource management. AI infrastructure gives the company a new growth market.

The CoolIT acquisition moves Ecolab closer to the processor.

Its platform could span the entire cooling loop—from the cold plate touching the chip to the water treatment and heat-rejection systems serving the wider facility.

Ecolab has also been testing AI-enabled water management with Digital Realty. A program across 35 U.S. data centers was designed to reduce water use by as much as 15%, potentially saving 126 million gallons of potable water annually. Ecolab

This is the second-order AI opportunity.

AI creates more heat.

More heat creates demand for advanced cooling.

Advanced cooling creates demand for fluid management, monitoring, treatment and water-efficiency technology.

Ticker: ECL

Other companies worth watching include Vertiv, Schneider Electric, Trane Technologies, Johnson Controls and Xylem.

Each controls part of the system responsible for moving heat and managing water.

The Hot Take

AI does not have one water problem.

It has hundreds of local water decisions.

Some projects will create legitimate pressure on constrained municipal systems. Others will operate with closed-loop cooling, reclaimed water or minimal annual consumption.

Treating every data center as equally water-intensive is inaccurate.

Treating water as an afterthought is equally dangerous.

The next phase of the market will move beyond debating whether data centers use water.

The important question will be whether each facility uses the right water, in the right way, for the watershed where it is built.

Water Across the Eight Pillars

Water affects every part of AI infrastructure development.

Cooling choices influence power demand. Water availability changes the value of land. Pipelines and treatment systems require capital. Permitting and withdrawal rights depend on policy.

Water systems need engineers, operators and skilled technicians. Reuse and replenishment strategies shape sustainability commitments. Connectivity can allow computing workloads to move between facilities with different resource conditions.

The eight pillars operate as one system.

A water constraint can reduce power efficiency.

A power decision can expand the indirect water footprint.

A policy decision can determine whether reclaimed water infrastructure gets built.

AI readiness depends on understanding those relationships before the campus is designed.

What Leaders Should Start Asking

Communities, Developers, Utilities and Investors Should Ask:

• What is the facility’s projected annual water withdrawal?

• How much water will actually be consumed?

• What is the peak daily requirement?

• How does demand change during extreme heat?

• What cooling system will the facility use?

• Is the system evaporative, air-cooled or closed loop?

• Does liquid cooling require continual fresh-water input?

• Will the facility use potable, reclaimed, surface or groundwater?

• Is reclaimed-water infrastructure already available?

• What new pipelines or treatment systems are required?

• Who will pay for those improvements?

• What WUE is expected at full campus buildout?

• How does the facility’s power supply affect its indirect water footprint?

• What happens during drought restrictions?

• Can the campus operate with reduced water availability?

• How much water will be returned after use?

• How will discharged water be treated?

• Will water data be reported at the individual facility level?

• Do replenishment projects occur in the same watershed?

• What other users depend on the water source?

• How will future campus expansion change the numbers?

These questions turn a broad environmental debate into an infrastructure assessment.

The Golden Nugget

Water risk is not measured in gallons alone. It is measured in gallons, source, place and time.

The number matters.

The watershed gives the number meaning.

Frequently Asked Questions

How Much Water Do U.S. Data- Centers Use?

Lawrence Berkeley National Laboratory estimated that U.S. data centers directly consumed approximately 66 billion liters, or roughly 17.4 billion gallons, in 2023. Actual use varies widely by facility, location and cooling technology.

Why Do AI Data-Centers Use Water?

AI processors create substantial heat. Water can be used in cooling towers, chilled-water systems and other heat-rejection equipment. Water may also be consumed indirectly by the power plants generating electricity for the facility.

Does Liquid Cooling Consume Large Amounts of Water?

Not necessarily. Direct-to-chip liquid cooling can use a closed loop in which the same fluid circulates continuously. Water consumption depends largely on how the facility ultimately removes heat from that loop.

What is Water Usage Effectiveness?

Water Usage Effectiveness, or WUE, measures the water used for cooling and humidification relative to the electricity consumed by IT equipment. It is useful for comparing efficiency but does not fully capture local water stress, water source or indirect consumption.

Can Data-Centers Use Reclaimed Water?

Yes. Treated wastewater can be used for cooling and other non-potable purposes. However, the community may need treatment capacity, pipelines and additional infrastructure to deliver it.

Can a Data-Center Operate Without Consuming Water for Cooling?

Yes. Air-cooled, dry-cooling and closed-loop systems can operate with little or no ongoing cooling-water consumption. These approaches may create different electricity and cost trade-offs.

What Does Water Positive Mean?

Water positive generally means a company intends to replenish more water than it consumes through efficiency, reuse and watershed projects. The location and timing of replenishment should still be evaluated.

The IOI Take

The AI water debate has become too simple.

One side treats data centers as an inevitable threat to community water supplies. The other points to improving efficiency and assumes the problem has been solved.

Neither position is sufficient.

Water demand is real. Cooling design is changing. Efficiency is improving. Geographic risk remains uneven.

The strongest AI-ready communities will not avoid the water conversation.

They will become better at answering it.

They will know which water sources are available, how those sources perform during drought and which properties can access reclaimed-water systems. They will understand the relationship between cooling and power. They will establish transparent reporting expectations before projects are approved.

Most importantly, they will stop treating water as an unlimited input.

The Intelligence Economy will require cooling.

It does not have to require careless water use.

One Last Thought

Every AI request eventually becomes heat.

That heat has to go somewhere.

It can move into air, water, liquid coolant or another physical system, but it cannot disappear.

This is the part of artificial intelligence that software conversations tend to miss.

The model may be digital.

Cooling it is a local infrastructure decision.

Workforce

Policy

Capital

Sustainability