AI’s Hidden Water Bill: The 3.4-Trillion-Gallon Data Center Number Is Real—But It Doesn’t Mean 3.4 Trillion Gallons Were “Used Up”
The modern artificial-intelligence boom does not live entirely in the cloud.
It lives in buildings.
Enormous buildings.
Behind their walls are racks of GPUs and servers consuming electricity continuously, producing heat continuously and depending on physical infrastructure that reaches far beyond the data-center campus itself.
Power plants.
Transmission lines.
Reservoirs.
Rivers.
Cooling towers.
And water.
A major new report from sustainability nonprofit Ceres has put a startling number on one part of that hidden infrastructure.
In seven U.S. states containing roughly half of America's data-center capacity, Ceres estimates that the electricity attributed to data centers was associated with approximately 3.4 trillion gallons of freshwater withdrawals in 2024. The states are Virginia, Texas, California, Illinois, Georgia, Ohio and Arizona.
Convert 3.4 trillion gallons and you get roughly 10.4 million acre-feet.
That is greater than the approximately 8 million acre-feet of water Californians use in an average year for urban purposes, according to California's Legislative Analyst's Office.
It is also roughly 12 times the combined annual water use of Los Angeles, Phoenix and Washington, D.C., according to Ceres.
Those comparisons are extraordinary.
But there is an equally important correction.
The 3.4 trillion gallons represent estimated water withdrawals, not water consumption.
Much of that water—particularly water associated with hydropower and once-through power-plant cooling—is returned to rivers, reservoirs or other water bodies rather than permanently removed from the local water system. Ceres explicitly states that its analysis uses withdrawals rather than consumption and includes hydropower.
In fact, approximately 78% of the withdrawal volume in the seven-state analysis was associated with hydropower, where enormous quantities of water pass through turbines and are largely returned downstream.
So the viral interpretation—
“AI data centers consumed more water than every California city combined”
—is not what the study found.
A much more accurate statement is:
Data-center electricity demand in seven major U.S. data-center states was associated with an estimated 3.4 trillion gallons of freshwater withdrawals from power generation in 2024—a volume greater than California’s average annual urban water use.
That distinction makes the number less sensational.
It does not make the underlying problem unimportant.
The Ceres report exposes something the public conversation about AI often misses entirely:
A data center's water footprint does not stop at the property line.
The 3.4 Trillion Gallons Are Not All From Water Pipes Entering Data Centers
When people hear that data centers use water, they usually imagine cooling.
That is reasonable.
Computers generate heat.
High-density AI hardware generates extraordinary amounts of heat.
Data-center operators must remove that heat using combinations of:
- Air cooling
- Chilled-water systems
- Evaporative cooling
- Cooling towers
- Direct-to-chip liquid cooling
- Closed-loop systems
- Outside-air economization
Some facilities consume large quantities of water directly.
But Ceres's 3.4-trillion-gallon figure is mostly about something else.
It is about water used by the electricity system supplying the data centers.
A data center may therefore advertise impressive on-site water efficiency while still depending on electricity from water-intensive power plants hundreds of miles away.
The water footprint has moved.
It has not necessarily disappeared.
There Are Two Different Water Footprints
To understand the issue, separate data-center water use into two broad categories.
Direct water use
This is water used at the data-center site itself.
Much of it is associated with cooling.
Lawrence Berkeley National Laboratory estimated that U.S. data centers directly consumed approximately 66 billion liters—about 17.4 billion gallons—of water in 2023.
Indirect water use
This is water associated with producing the electricity the facility purchases from the grid.
Power stations may withdraw or consume water to:
- Condense steam
- Cool equipment
- Operate cooling towers
- Support thermal generation
- Generate hydropower
LBNL separately estimated approximately 800 billion liters, or roughly 211 billion gallons, of indirect water consumption associated with U.S. data-center electricity in 2023.
Notice the huge difference between that 211-billion-gallon consumption estimate and Ceres's 3.4-trillion-gallon withdrawal estimate.
Both can be valid because they are measuring different things.
Withdrawal and Consumption Are Not the Same
This distinction is absolutely central.
Water withdrawal
Withdrawal measures water taken from a source for use.
That might involve drawing water from:
- A river
- A lake
- A reservoir
- An aquifer
Some or most of that water may later be returned.
Imrul
Water consumption
Consumption measures the portion of water that is effectively removed from immediate local availability.
This can happen through:
- Evaporation
- Incorporation into products
- Transfer to another basin
- Other processes preventing immediate return
Ceres explains that it deliberately chose withdrawals because it wanted to estimate the full volume of water on which electricity-generating assets depend, treating that as an indicator of operational exposure to water shortages.
That is a legitimate water-risk question.
It is simply different from asking:
“How many gallons disappeared?”
Why a Power Plant Can Withdraw Huge Amounts Without Consuming Them
Consider a thermal power plant with once-through cooling.
Cold water is withdrawn from a river or lake.
It passes through the cooling system.
It absorbs heat.
Then most of the water is returned.
Ceres notes that once-through systems can withdraw extremely large volumes while returning most of the water at a higher temperature.
A cooling tower behaves differently.
It circulates water repeatedly and therefore may withdraw far less.
But more of that water can be consumed through evaporation.
So two power plants can create very different numbers:
Plant A:
Huge withdrawal, relatively low consumption.
Plant B:
Lower withdrawal, higher proportional consumption.
Neither metric alone tells the complete story.
Hydropower Makes the 3.4-Trillion-Gallon Figure Look Enormous
Hydropower is the biggest reason the withdrawal number becomes so large.
Ceres found that approximately 78% of power-generation withdrawal volume across its seven-state study was associated with hydropower.
Hydroelectric generation works by passing water through turbines.
The water is not normally destroyed or permanently removed from the river system.
It continues downstream.
There can still be water-related impacts.
Reservoirs can lose water through evaporation.
Drought can reduce available generation.
Reservoir operations can compete with environmental flows, irrigation, flood control and municipal needs.
But counting all water flowing through turbines as a withdrawal produces a very different concept from counting gallons permanently consumed.
Ceres acknowledges this directly in its methodology: its water-intensity factors treat hydropower as a withdrawal rather than simply an instream use.
That is why the 3.4-trillion-gallon number must be interpreted carefully.
California Dominates the Withdrawal Figure Because of Hydropower
California illustrates this issue perfectly.
Ceres estimates that electricity attributed to California data centers was associated with approximately:
1.4 trillion gallons of freshwater withdrawals in 2024.
That is about ten times the annual water use of the City of Los Angeles.
But Ceres explains that California's unusually high withdrawal figure is driven heavily by its hydroelectric fleet.
Across California's entire electricity system—not merely the portion attributed to data centers—Ceres estimated roughly 37.9 trillion gallons of freshwater withdrawals in 2024, largely because of hydropower.
That tells us something important.
A massive withdrawal number does not automatically imply an equivalently massive reduction in available drinking water.
Is the California Comparison Therefore Meaningless?
No.
It is useful as a scale comparison.
But it should not be interpreted as an apples-to-apples measure of depletion.
California's Legislative Analyst's Office says Californians use roughly 8 million acre-feet annually for urban purposes, including residential, commercial, industrial and large-landscape uses.
Ceres's 3.4 trillion gallons convert to roughly 10.4 million acre-feet.
So the numerical comparison is real.
But:
10.4 million acre-feet of estimated power-plant withdrawals
is not environmentally equivalent to:
8 million acre-feet of urban water use.
Some withdrawn power-plant water returns almost immediately.
Some urban water is also treated and returned.
Some is consumed.
Some goes to wastewater systems.
Some recharges groundwater.
The two accounting categories are different.
A good headline should make that clear.
The Seven States Contain Roughly Half of U.S. Data Centers
Ceres focused on:
- Virginia
- Texas
- California
- Illinois
- Georgia
- Ohio
- Arizona
Together, they account for approximately 50% of U.S. data-center capacity or facilities, depending on the underlying industry dataset being referenced.
These states also represent very different electricity systems.
Virginia has become perhaps America's most famous data-center hub.
Texas combines explosive data-center growth with a large fossil-fuel fleet and rapidly expanding wind and solar.
California combines major digital infrastructure with substantial hydropower.
Illinois relies heavily on nuclear generation.
Arizona faces severe water constraints while attracting enormous new computing projects.
That variation allowed Ceres to ask not only how much electricity data centers use, but:
What kind of water dependency sits behind that electricity?
These Are Not Strictly “AI Data Centers”
Another correction matters.
The Ceres estimate does not isolate only servers running artificial-intelligence workloads.
It applies state-level estimates of data-center electricity demand generally.
That includes infrastructure supporting combinations of:
- AI training
- AI inference
- Cloud computing
- Enterprise services
- Databases
- Streaming
- Search
- Social media
- Traditional web workloads
- Other digital services
AI is a major driver of the new growth.
But calling the entire 3.4-trillion-gallon estimate “AI's water use” overstates what was measured.
The more precise statement is:
The AI boom is accelerating growth in a data-center sector whose electricity already carries a huge indirect water dependency.
AI Is Still the Reason This Question Has Become Urgent
Even with that qualification, AI is transforming the scale of the problem.
The U.S. Department of Energy reported that data centers consumed approximately 176 terawatt-hours of electricity in 2023, representing about 4.4% of total U.S. electricity consumption.
DOE's 2024 analysis projected that this could rise to approximately 6.7% to 12% by 2028.
More recent Electric Power Research Institute scenarios are even higher.
EPRI's 2026 analysis estimates that U.S. data centers could consume approximately 9% to 17% of national electricity by 2030, compared with roughly 4% to 5% today.
EPRI estimates 2024 data-center demand at roughly 177–192 TWh, potentially rising to approximately 380–790 TWh by 2030.
Every additional terawatt-hour has an environmental profile.
Carbon emissions are one part.
Water is another.
Ceres Projects the Withdrawal Footprint Could Nearly Double
Applying data-center growth scenarios to the electricity system, Ceres estimates that annual freshwater withdrawals embedded in electricity for the seven states could rise from:
3.4 trillion gallons in 2024
to approximately:
4.1 trillion to 7.6 trillion gallons by 2030.
Again, those are withdrawal estimates.
Not consumption.
But they indicate increasing dependence on water-sensitive power infrastructure.
And dependence matters even when water is returned afterward.
Why Withdrawal Can Still Create Risk
Suppose a power station withdraws 100 gallons and returns 98.
One might conclude:
Only two gallons matter.
But what happens if the river becomes too low to withdraw the original 100 safely?
The plant may be forced to reduce output.
Or shut down.
Or seek alternative water.
Or compete with other users.
The operational dependency exists because the water has to be available at the moment the plant needs it.
Ceres therefore argues that withdrawal is a useful risk metric even where consumption is comparatively low.
This is particularly important during:
- Drought
- Heatwaves
- Low river flows
- Reservoir shortages
- Extreme electricity demand
Unfortunately, those events can occur simultaneously.
Heatwaves Create an Awkward Feedback
During extreme heat:
Air conditioners run harder.
Electricity demand rises.
Data-center cooling systems may work harder.
Thermal power plants may need significant cooling.
River and reservoir temperatures can rise.
Water availability may decline.
That combination can reduce the resilience of water-dependent electricity systems exactly when electricity is most valuable.
This is the type of compound risk Ceres wants utilities and data-center operators to consider.
78% of Electricity in the Study Came From Water-Using Plants
Among 3,907 active power plants examined across the seven states, Ceres identified 1,434 facilities as dependent on freshwater for generation.
Those water-dependent plants generated approximately:
1,142 TWh
out of:
1,467 TWh
of electricity in the study area in 2024.
That is almost 78% of total generation.
The remaining roughly 22% came from facilities such as:
- Wind
- Solar
- Battery storage
- Plants using saltwater
- Other low-freshwater generation
That difference matters because electricity source can dramatically change the water footprint of computing.
Ohio, Georgia and Virginia Are Especially Dependent on Thermal Generation
Ceres found that more than 90% of electricity generation in Ohio, Georgia and Virginia came from water-intensive thermal power plants.
Thermal generation includes technologies such as:
- Natural gas
- Coal
- Nuclear
- Some geothermal
- Biomass
- Petroleum
These technologies generally generate heat first, then convert that heat into electricity.
That process frequently creates a major cooling requirement.
Water often performs that cooling job.
Wind and Solar Change the Equation Dramatically
Wind turbines require almost no operational water to generate electricity.
Solar photovoltaic panels similarly require little to no water during electricity production, aside from limited cleaning and other site needs.
Ceres specifically contrasts these technologies with thermal generation and hydropower, noting that wind and solar require little to no operational water.
This means two identical AI facilities can have dramatically different indirect water footprints depending on their electricity supply.
A megawatt-hour is a megawatt-hour to the server.
It is not a megawatt-hour to the watershed.
Texas Shows the Complexity of the Energy Transition
Texas has built enormous quantities of wind and solar generation.
That lowers the water intensity of portions of its electricity supply.
At the same time, Ceres found that fossil generation—principally natural gas and coal—still accounted for roughly 90% of Texas's estimated power-sector freshwater withdrawals within its analysis.
This illustrates why simply reporting a state's renewable percentage does not describe its full water exposure.
The marginal power source matters.
Cooling technology matters.
Which plants run during peak periods matters.
Location matters.
Illinois Has a Different Problem: Nuclear Power and Water Stress
Illinois generated approximately 63% of its electricity from nuclear power in 2024, according to Ceres.
Nuclear power produces very low operational carbon emissions.
But conventional nuclear plants can require substantial cooling water.
Ceres reports that approximately 75% of Illinois nuclear generation came from facilities located in areas it classified as experiencing high water stress.
This reveals an important truth about environmental policy:
Low-carbon does not always mean low-water.
Energy choices involve multiple environmental dimensions.
Virginia Is Where the Data-Center Story Becomes Extreme
Northern Virginia is one of the largest concentrations of data centers anywhere in the world.
Ceres estimates approximately 753 billion gallons of freshwater withdrawals associated with electricity attributed to Virginia data centers in 2024.
That was approximately 21 times Washington, D.C.'s annual water use, according to the report's comparison.
Again, this is not saying Northern Virginia data centers physically pumped 753 billion gallons through their cooling systems.
The number represents their modeled share of power-generation water withdrawals.
That difference is enormous.
Arizona’s Water Context Makes Every Additional Demand Sensitive
Ceres estimates approximately 520 billion gallons of withdrawals associated with electricity for Arizona data centers in 2024—about five times Phoenix's annual municipal water use in the report's comparison.
Arizona and Illinois showed the greatest combined exposure to water stress and drought among the study states.
That does not automatically mean a new Arizona data center will cause household taps to run dry.
Water systems are far more complex.
But it means location matters enormously.
An additional gallon demanded in a water-rich basin and the same gallon demanded in an intensely stressed basin are not environmentally equivalent.
Two-Thirds of Water-Dependent Generation Was Exposed to Significant Water Stress
Ceres found that approximately 66% of electricity generated using water in the seven states came from areas facing medium-high to extremely high baseline water stress.
Water stress is not simply drought.
It describes the relationship between demand and available supply.
A region can be water-stressed even during a year without a dramatic drought if normal withdrawals already approach available resources.
That can increase competition among:
- Cities
- Farmers
- Industry
- Ecosystems
- Power plants
Add rapidly expanding data-center electricity demand and the pressure can increase.
Drought Adds Another Layer
Ceres also found that approximately 44% of electricity generated using freshwater came from power plants located in areas that experienced at least three months of drought conditions during 2024.
The report's plant-count metric gives a slightly different number because it asks a different question: around 46% of water-using power plants experienced at least three months of drought.
The distinction matters.
One number counts electricity generation.
The other counts facilities.
Either way, water-sensitive electricity infrastructure is already operating in areas exposed to water stress.
Hydropower Is Renewable but Not Immune to Drought
Hydropower is sometimes presented as though renewable electricity is automatically independent of water risk.
Hydropower obviously is not.
Ceres reports that approximately 39% of California's hydropower water use in 2024 occurred in basins facing medium-high to extremely high water stress.
Low reservoir levels can directly reduce generation.
Ceres points to planning by Pacific Gas & Electric that reduced expected hydro generation partly because of climate-related changes including warmer conditions, reduced snowpack and drought.
So hydropower's carbon advantage does not eliminate its dependence on hydrology.
The Most Important Water Number May Actually Be Much Smaller
This sounds paradoxical.
Ceres's 3.4-trillion-gallon figure attracts attention because it is enormous.
But for understanding depletion, the smaller consumption numbers may often matter more.
LBNL's 2024 national analysis estimated that in 2023 U.S. data centers directly consumed approximately:
17.4 billion gallons
while their electricity supply was associated with approximately:
211 billion gallons of indirect water consumption.
That means electricity-generation water consumption was roughly an order of magnitude larger than direct on-site consumption nationally under that methodology.
This reinforces Ceres's central argument without requiring the 3.4-trillion-gallon withdrawal headline.
The hidden electricity footprint is genuinely important.
Why Do Companies Talk More About On-Site Cooling?
Because it is easier to measure and control.
A data-center operator knows:
How much water enters its site.
What type of cooling equipment it uses.
Its Water Usage Effectiveness.
Whether reclaimed water is available.
Its local municipal supply.
That gives the company a clear operational number.
Indirect water is harder.
A facility connected to the grid may receive electricity produced by dozens or hundreds of generators.
Those plants change by hour.
Grid imports change.
Weather changes.
Market prices change.
Plant dispatch changes.
Calculating the water embedded in one data center's electricity therefore requires modeling.
Ceres Says Most Operators Still Focus Primarily Inside the Fence
Ceres reviewed company disclosures and concluded that most data-center operators' water strategies continue to focus primarily on cooling technologies and water used within the data center itself.
It found much more limited disclosure of water embedded in purchased electricity.
The report names Amazon, Microsoft and Apple among companies developing direct water-use targets or strategies.
But embedded electricity water is far less consistently reported.
That is the transparency gap Ceres wants to close.
Some Companies Are Beginning to Measure It
The report identifies exceptions.
Meta reports embedded water consumption associated with purchased electricity and estimates avoided water consumption linked to renewable-energy procurement.
CyrusOne calculates a metric combining site water use with water used to produce electricity, calling the latter component WUE Source.
Ceres reports that CyrusOne's source-level water metric declined substantially after increased renewable-energy procurement.
That demonstrates why water accounting can influence energy purchasing decisions.
If a company only measures cooling water, changing electricity sources may appear to offer no water benefit.
If indirect water is measured, the picture changes.
Power Companies Have the Same Blind Spot
Ceres found that many large power producers recognize data centers as major drivers of future electricity demand.
Far fewer explicitly connect that demand growth with water risk.
The report gives Ohio utility AEP as an example.
AEP anticipates approximately 12 gigawatts of new contracted load between 2026 and 2030, much of it driven by data centers, and is investing billions in grid expansion.
Yet Ceres says publicly available disclosures do not clearly connect that data-center load growth to increased water reliance.
Similarly, Dominion Energy reported 81 new data-center projects coming online in Northern Virginia between 2020 and 2024, while planning huge investments in additional generation and infrastructure.
Ceres found little corresponding discussion in readily available disclosures about what that growth means for water.
The Report Is Not a Precise Meter Reading
Another major caveat deserves emphasis.
Ceres itself calls its results directional estimates rather than precise measurements.
The researchers did not install water meters at every power plant and trace individual electrons into particular AI servers.
They constructed a model.
That model used:
- 2024 EIA power-plant data
- Reported plant water data where available
- Modeled water-intensity values where unavailable
- EPRI estimates of data centers' share of state electricity use
- WRI water-stress data
- U.S. Drought Monitor information
The result estimates scale and exposure.
It does not provide exact gallon-by-gallon attribution.
Only a Minority of Relevant Plants Reported Water Data Directly
Of the 1,434 water-dependent plants in the analysis, only 194 reported withdrawal or consumption data through the relevant EIA schedule.
For the other 1,240 plants, Ceres relied on technology-specific water-intensity estimates from World Resources Institute guidance.
That does not invalidate the analysis.
Modeling missing data is normal scientific practice.
But it increases uncertainty.
A national-average water-intensity factor may not perfectly capture:
- Plant age
- Cooling design
- Local climate
- Efficiency
- Operational patterns
Ceres explicitly lists these limitations.
California and Virginia Required More Modeling Than Some Other States
The report says states with large thermal fleets such as Arizona, Texas and Illinois had a greater share of directly reported water data.
California and Virginia, which include more hydroelectric and smaller facilities, required more estimated values.
This is particularly relevant because California produced the largest attributed withdrawal number.
The result should therefore be treated as an estimate of water-risk exposure, not a precise invoice.
The Electricity Grid Does Not Respect State Borders
Another limitation is geography.
Ceres applies state-level data-center electricity shares to state power-generation withdrawals.
But real electricity grids cross state borders.
Northern Virginia is part of PJM, a regional transmission system spanning multiple states.
Texas's ERCOT covers most, but not all, of Texas.
A particular data center may therefore consume electricity generated outside the state in which it sits.
Ceres acknowledges that this can distort state totals.
Power-exporting states may have attributed water use overstated.
Importing states—including Virginia and California—may have some indirect water use understated.
This is exactly why Ceres describes its state numbers as approximations.
The Report Also Excludes Some Water Use
The estimate does not capture everything.
Ceres excludes or incompletely captures:
- Some behind-the-meter power generation
- Water used in semiconductor manufacturing
- Construction-related water
- Other supply-chain water
- Seasonal variations in electricity demand
- Detailed site-specific water sources
So 3.4 trillion gallons is not a comprehensive lifecycle water footprint of AI.
It is a particular estimate of freshwater withdrawal exposure associated with electricity generation.
Hardware Manufacturing Has Its Own Water Footprint
Before an AI server ever begins processing a prompt, its chips have already required enormous industrial infrastructure.
Semiconductor manufacturing uses ultrapure water for repeated cleaning and fabrication steps.
Server manufacturing uses metals, chemicals and energy.
Data-center buildings consume concrete and steel.
Backup generators and batteries have their own supply chains.
None of that is captured by Ceres's 3.4-trillion-gallon electricity-withdrawal estimate.
This is another reason phrases such as “AI's total water use” should be avoided unless the boundary of the calculation is explicitly defined.
One AI Prompt Does Not Have a Universal Water Cost
You may have seen estimates claiming that a certain number of chatbot prompts require a bottle of water.
Such estimates can be useful illustrations.
They are not physical constants.
The water footprint of an AI workload depends on:
- Model
- Hardware
- Utilization
- Location
- Weather
- Cooling system
- Electricity source
- Time of day
- Grid mix
Running the same computation in two locations can create dramatically different indirect water impacts.
The server does not know the difference.
The watershed does.
Where the Data Center Is Built Matters
This may be the biggest policy implication.
A data center using 100 megawatts in a water-abundant region supplied mainly by wind and solar has a very different water-risk profile from an identical facility in a severely stressed watershed dependent on water-intensive thermal generation.
Yet both may report similar:
- Computing capacity
- Power Usage Effectiveness
- On-site Water Usage Effectiveness
Site-level efficiency metrics alone cannot capture the regional environmental context.
Ceres therefore recommends evaluating basin-level water stress when deciding where facilities and power infrastructure should be built.
Community Concerns Are Becoming a Business Risk
The AI boom has moved so quickly that many communities are encountering data-center proposals before regulatory systems have developed consistent standards for evaluating them.
Ceres reports that opposition related to water and grid capacity disrupted approximately $130 billion in data-center projects during the first quarter of 2026 alone.
That figure illustrates a larger point.
Water is no longer merely an environmental reporting issue.
It can become:
- A permitting issue
- A political issue
- A financial issue
- A grid-reliability issue
- A community-relations issue
Ignoring it can delay projects worth billions.
Better Cooling Alone Cannot Solve the Whole Problem
Suppose a technology company builds a data center with nearly zero direct cooling-water consumption.
That is valuable.
But if the facility requires gigawatts of electricity generated by water-intensive power stations, substantial indirect dependence remains.
The inverse can also happen.
A facility may use water-efficient electricity but operate an evaporative cooling system consuming substantial water directly.
The best strategy must therefore examine:
Cooling + electricity + local watershed conditions.
Not one metric in isolation.
The Energy Source Can Matter More Than the Cooling Tower
Ceres's core argument is that purchased electricity may represent the majority of a data center's operational water footprint.
That means procurement decisions become water decisions.
Buying more electricity from:
- Wind
- Solar PV
- Some low-water generation
can reduce indirect water dependence.
Buying additional power from:
- Water-intensive thermal generation
- Certain hydroelectric systems
can increase it.
Renewable procurement is therefore not only a carbon strategy.
Depending on the technology, it can be a water strategy too.
Nuclear Power Presents a More Complicated Trade-Off
AI companies increasingly discuss nuclear energy because it can provide large amounts of reliable, low-carbon electricity.
From a carbon perspective, that is attractive.
From a water perspective, conventional nuclear power can be intensive because of cooling needs.
This does not mean nuclear power is environmentally inferior overall.
It means policymakers cannot optimize one metric and assume every other metric improves automatically.
Future reactor designs using alternative cooling systems may reduce water needs.
Existing plants still need local water-risk evaluation.
Gas Is Not Automatically a Low-Water Solution Either
Natural-gas plants generally emit less carbon dioxide than coal for equivalent electricity generation.
But their water requirements vary dramatically.
Combined-cycle plants with steam cycles and cooling towers can consume meaningful quantities.
Simple-cycle gas turbines can use much less water but operate less efficiently.
Dry cooling can reduce water dependence but can increase cost and reduce performance under hot conditions.
There is no single “gas water footprint.”
Technology details matter.
Ceres Wants Water Added to Power Planning
The report recommends that utilities and power producers explicitly include water availability in decisions about new generation built for data-center growth.
That means assessing:
- Withdrawals
- Consumption
- Water source
- Basin stress
- Drought vulnerability
- Cooling technology
- Future climate conditions
For a load expected to operate 24 hours per day for decades, this is not merely environmental idealism.
It is infrastructure risk management.
Ceres Also Wants Data-Center-Level Disclosure
One of the report's strongest recommendations is greater transparency.
Ceres argues companies should disclose both:
Direct site water use
and
Water embedded in purchased electricity
for individual facilities or, when necessary, at city, county or watershed level.
That would make it easier for communities to evaluate proposed projects realistically.
A company could no longer claim near-zero water use based solely on cooling while ignoring the water dependency of its power supply.
Regulators Are Beginning to Pay Attention
Water reporting for data centers is increasingly entering state policy debates.
Ceres notes that several jurisdictions are developing requirements or guardrails around data-center energy and water impacts. Its broader data-center policy work points to states such as Utah requiring more detailed reporting of water sources, withdrawals, consumption and efficiency measures.
The direction is clear.
Data centers are becoming too large to be treated like ordinary commercial buildings.
Their infrastructure demands increasingly resemble industrial development.
AI Has Turned Digital Infrastructure Into Heavy Industry
This may be the deeper story.
For years, the digital economy was described using almost weightless language.
Cloud.
Virtual.
Online.
Digital.
Artificial intelligence makes the physical reality harder to ignore.
Training and serving large models requires:
Land.
Steel.
Concrete.
Copper.
Transformers.
Transmission infrastructure.
Power plants.
Cooling equipment.
Water.
A hyperscale AI campus may consume electricity comparable with a large city.
EPRI notes that a single 100- to 1,000-megawatt data center can use electricity comparable with approximately 80,000 to 800,000 homes.
That is not merely an IT installation.
It is industrial infrastructure.
Does AI Literally Compete With Residents for Drinking Water?
Sometimes the possibility exists.
But the answer depends heavily on location and system design.
Direct cooling can draw from municipal potable supplies.
Other facilities use:
- Reclaimed wastewater
- Non-potable water
- Closed-loop systems
- Air cooling
Indirect electricity water may come from an entirely different watershed.
Hydropower may involve river flow without substantial consumption.
A thermal plant may consume water through evaporation.
So saying “AI is taking drinking water away from residents” as a universal statement is too broad.
The better question is local:
What water does this facility and its electricity supply depend on, and what else depends on the same water?
The Local Question Is More Important Than the National Number
Three trillion gallons sounds terrifying.
Three billion sounds smaller.
Neither tells you whether a specific community is in danger.
Imagine two projects.
Project A
Uses 20 million gallons in a water-rich basin with abundant renewable supply.
Project B
Uses 5 million gallons in an aquifer already declining faster than it can recharge.
Project B may create the greater risk despite using one-quarter as much water.
That is why Ceres repeatedly emphasizes basin-level stress rather than total gallons alone.
Water is geographically constrained.
You cannot solve a drought in Arizona with unused water in Michigan.
The Viral Headline Gets One Thing Right
The phrase:
“AI has a hidden water footprint”
is fundamentally correct.
Electricity is not water-free.
Cooling is not the whole story.
Grid choices matter.
Location matters.
And corporate disclosure frequently gives an incomplete picture if it reports only water crossing the data-center property boundary.
Ceres has identified a real blind spot.
But the Viral Headline Gets Three Things Wrong
Three corrections should accompany the 3.4-trillion-gallon number every time it is shared.
1. It is not AI-only
The calculation covers data centers generally.
AI is driving rapid growth, but the 2024 baseline includes many other computing workloads.
2. It measures withdrawals, not consumption
A large share of the water is returned.
The figure does not mean 3.4 trillion gallons vanished.
3. Hydropower dominates the withdrawal volume
Approximately 78% of the study area's withdrawal volume came from hydropower.
Those details radically change how the number should be interpreted.
The Corrected Story Is Still Serious
Strip away the sensationalism and the conclusion remains significant.
U.S. data centers are becoming one of the fastest-growing electricity loads in the country.
A large majority of electricity across the seven major data-center states still comes from plants that depend on freshwater.
Much of that generation occurs in areas already exposed to substantial water stress or drought.
Most operators report direct water far more clearly than the indirect water dependency of purchased electricity.
And both electricity and water demand are likely to rise sharply as AI infrastructure expands.
No exaggeration is required.
The Real Environmental Question Is Not Whether AI “Drinks” 3.4 Trillion Gallons
AI does not drink water.
Servers do not drink water.
Power plants and cooling systems interact with water in specific physical ways.
The useful questions are:
How much is withdrawn?
How much is consumed?
Where?
From what source?
At what time of year?
Under what level of existing water stress?
How much returns?
At what temperature?
What other users depend on the same watershed?
What happens during drought?
And can the electricity be supplied using less water-intensive technology?
Those questions are less viral.
They are much more useful.
Frequently Asked Questions About AI Data Centers and the 3.4-Trillion-Gallon Water Claim
Do U.S. data centers really use 3.4 trillion gallons of water per year?
Ceres estimates that electricity attributed to data centers in seven major U.S. data-center states was associated with approximately 3.4 trillion gallons of freshwater withdrawals in 2024.
The number represents indirect power-generation withdrawals, not direct water piped into data centers.
Which seven states were included?
The study analyzed:
Virginia, Texas, California, Illinois, Georgia, Ohio and Arizona.
Together they account for roughly half of U.S. data-center capacity or facilities.
Is the 3.4 trillion gallons all consumed?
No.
This is one of the most important distinctions in the report.
Ceres measured withdrawals rather than consumption.
A significant share of withdrawn water is returned to its source.
What is the difference between water withdrawal and water consumption?
Water withdrawal means water is taken from a source for use.
Water consumption refers to water that is effectively removed from immediate local availability, such as through evaporation.
A power plant can withdraw enormous quantities while consuming only a fraction.
Why did Ceres use withdrawals?
Ceres says withdrawals better represent the total volume of water on which an electricity-generating asset depends to operate and therefore provide an indicator of exposure to water scarcity or drought.
Does hydropower count in the 3.4 trillion gallons?
Yes.
Ceres explicitly includes hydropower in its withdrawal calculation.
How much of the withdrawal volume came from hydropower?
Approximately 78% of withdrawal volume across the study area was dominated by hydropower.
Is hydropower water permanently lost?
Most water passing through a hydroelectric turbine continues downstream.
There can still be losses from reservoir evaporation and significant water-management constraints, but turbine flow should not be interpreted as water permanently consumed.
Did data centers use more water than all California cities combined?
As a volume comparison, the Ceres estimate is larger.
The 3.4 trillion gallons equal roughly 10.4 million acre-feet, while California's Legislative Analyst's Office says average statewide urban water use is approximately 8 million acre-feet annually.
However, this is not an apples-to-apples comparison because Ceres is comparing power-generation withdrawals with urban water use.
How does Ceres compare the number with major cities?
Ceres says 3.4 trillion gallons is approximately 12 times the combined annual water use of Los Angeles, Phoenix and Washington, D.C.
Are these specifically AI data centers?
No.
The analysis covers data-center electricity demand broadly.
AI is a major driver of current and projected growth, but the estimate does not isolate only AI workloads.
How much electricity do U.S. data centers use?
DOE estimated data centers consumed approximately 176 TWh in 2023, around 4.4% of U.S. electricity demand.
EPRI estimates roughly 177–192 TWh in 2024.
How much electricity could data centers use by 2030?
EPRI's 2026 scenarios estimate approximately 380–790 TWh, equal to roughly 9% to 17% of U.S. electricity consumption.
Is AI responsible for the growth?
AI is one major driver, especially large-scale training and inference.
Other drivers include cloud computing, streaming, enterprise computing and other digital services.
How much water do data centers consume directly?
LBNL estimated approximately 66 billion liters, or 17.4 billion gallons, of direct U.S. data-center water consumption in 2023.
How much water is consumed indirectly through electricity?
LBNL estimated approximately 800 billion liters, or about 211 billion gallons, of indirect water consumption associated with data-center electricity in 2023.
Why is that so much smaller than 3.4 trillion gallons?
Because the LBNL figure measures consumption, while Ceres's 3.4 trillion gallons measure withdrawals.
Hydropower and once-through cooling can create enormous withdrawal volumes while returning much of the water.
How much water did California data-center electricity withdraw?
Ceres estimates approximately 1.4 trillion gallons in 2024.
The unusually large number is heavily influenced by California's hydropower system.
How much did Virginia account for?
Ceres estimates approximately 753 billion gallons of withdrawals associated with electricity attributed to Virginia data centers.
How much did Arizona account for?
The report estimates approximately 520 billion gallons associated with data-center electricity in Arizona.
Which state had the lowest data-center-attributed withdrawal in the report?
Ceres says state-level estimates ranged down to approximately 25 billion gallons in Ohio.
Why can Ohio be low even though it relies heavily on thermal generation?
Total attributed water depends on several factors simultaneously, including the amount of data-center electricity demand, generation mix and plant water intensity.
Water intensity alone does not determine the state total.
How much of the seven-state electricity came from plants using freshwater?
Approximately 78% of electricity generation in the study area came from power plants that Ceres classified as using freshwater.
Which states were most dependent on water-intensive thermal power?
Ceres says Ohio, Georgia and Virginia each generated more than 90% of their electricity from water-intensive thermal plants.
Which states had the greatest exposure to water stress?
Ceres identified Arizona and Illinois as having the greatest combined exposure to water stress and drought within the study.
How much electricity was generated in water-stressed areas?
Approximately 66% of electricity generated using water came from areas experiencing medium-high to extremely high baseline water stress.
What percentage was exposed to prolonged drought?
Approximately 44% of electricity generated using water came from power plants in areas experiencing at least three months of drought in 2024.
Can drought shut down power plants?
It can reduce the output of some water-dependent plants.
Low water levels can affect hydropower, while thermal plants can face cooling-water constraints.
Ceres cites previous cases where water conditions contributed to generation curtailment or reductions.
Is nuclear power water-intensive?
Many conventional nuclear plants require substantial quantities of cooling water.
Ceres notes that 63% of Illinois electricity came from nuclear power in 2024 and that much of this generation was located in water-stressed areas.
Does solar power use water?
Solar photovoltaic generation requires very little operational water compared with conventional thermal generation, aside from activities such as panel cleaning.
Does wind power use water?
Wind generation requires little to no water during electricity generation.
Ceres therefore identifies wind and solar as comparatively low-water electricity sources.
Are tech companies hiding data-center water use?
“Hidden” should be used carefully.
Many companies publicly report some direct water metrics and sustainability targets.
Ceres's criticism is that indirect water associated with purchased electricity is much less consistently disclosed.
Which companies disclose indirect water metrics?
Ceres highlights Meta as reporting embedded water consumption associated with purchased electricity.
It also cites CyrusOne's WUE Source metric, which includes electricity-related water.
Do Amazon, Microsoft and Apple report water strategies?
Ceres says these companies have targets or strategies addressing direct water use.
The report argues that sector-wide reporting on indirect electricity-related water remains much less developed.
Is Google's water footprint included in the Ceres calculation?
Ceres does not calculate a company-specific Google total.
The study works primarily at state and power-plant level and applies estimates of data centers' share of electricity demand.
Can a data center use zero water?
A facility can potentially approach zero direct cooling-water consumption with air cooling or closed-loop cooling designs.
That does not automatically make its total operational water footprint zero because its electricity supply may still depend on water.
Is liquid cooling automatically better for water?
Not necessarily.
Closed-loop liquid cooling can drastically reduce direct water use.
But total impact depends on how the electricity used to run pumps, chillers and computing equipment is generated.
Does AI directly compete with drinking water?
Sometimes infrastructure can share or compete for the same water sources, particularly when facilities use municipal potable supplies or power plants rely on stressed freshwater basins.
But the relationship is highly local and cannot be generalized to every data center.
Could switching to renewable electricity reduce AI’s water footprint?
Often, yes.
Wind and solar photovoltaic generation generally require far less operational water than thermal generation.
Renewable procurement can therefore reduce both carbon emissions and indirect water dependence.
Is all renewable electricity low-water?
No.
Hydropower is renewable but inherently dependent on water availability and can create large withdrawal-accounting figures.
Some geothermal technologies can also use water.
“Renewable” and “low-water” are overlapping but not identical categories.
How accurate is the 3.4-trillion-gallon estimate?
Ceres explicitly describes its figures as directional estimates rather than exact measurements.
The model combines reported data with water-intensity estimates and state-level assumptions.
How many power plants had directly reported water data?
Of 1,434 water-dependent plants included, 194 had relevant reported withdrawal or consumption data.
The remaining 1,240 relied on modeled water-intensity factors.
Why does that matter?
Actual plants differ in cooling technology, efficiency, age and local climate.
National-average estimates cannot capture every detail.
Ceres acknowledges this limitation.
Does the analysis know exactly which plant powers each data center?
No.
This is another important limitation.
The model uses state-level generation and data-center electricity shares, while real grids frequently cross state borders.
Could Virginia’s number actually be higher?
Possibly.
Ceres notes that importing states such as Virginia and California may have some water use understated because some electricity is generated outside their borders.
Could other states be overstated?
Yes.
Net power-exporting states may have some water withdrawals attributed to local data centers even though part of their electricity is exported elsewhere.
Ceres explicitly identifies this limitation.
How much could data-center electricity withdrawals reach by 2030?
Ceres estimates approximately 4.1 trillion to 7.6 trillion gallons annually across the seven states under different growth scenarios.
Does that mean 7.6 trillion gallons will disappear every year?
No.
That projection is for withdrawals, not permanent water consumption.
Is the environmental concern exaggerated?
Some viral presentations are exaggerated because they confuse withdrawal with consumption or label the entire footprint as AI-specific.
The underlying concern is nevertheless real: data-center electricity increasingly depends on water-sensitive power infrastructure, including in water-stressed regions.
What would be the most accurate headline?
A scientifically careful version would be:
“Electricity for data centers in seven major U.S. states was associated with an estimated 3.4 trillion gallons of freshwater withdrawals in 2024, according to Ceres.”
Then immediately add:
“Most of that water was not consumed, and hydropower accounted for the majority of withdrawal volume.”
What is the biggest takeaway from the Ceres report?
It is not that AI secretly “drank” 3.4 trillion gallons of freshwater.
It is that the environmental footprint of digital infrastructure extends far beyond the cooling pipes visible at the data center itself.
A company can build an exceptionally water-efficient cooling system and still rely on electricity generated by facilities dependent on stressed rivers or reservoirs.
The server sees electricity.
The community sees a power plant.
The power plant sees a river.
And the river has other users.
That is the hidden connection Ceres is asking the AI industry to acknowledge.
The 3.4-trillion-gallon figure gets attention because it is enormous.
But the most important lesson is not the size of that number.
It is the boundary of the calculation.
If we measure only what happens inside the data-center fence, we can make enormous industrial infrastructure look almost weightless.
Once we follow the electricity upstream, the picture changes.
AI does not live in an invisible cloud.
It lives in a physical system.
And that system runs not only on silicon and electricity—
but, in many places, on water.
