Why Big Tech's AI Water Numbers Don't Add Up

Every time you type a prompt into ChatGPT, something happens thousands of miles away that most people never think about. A server heats up. A cooling system kicks in. Water, real water pulled from a local supply, evaporates into the air.

People use ChatGPT for far more than writing help now, including budgeting and personal finance questions, and our guide to the best ChatGPT prompts for personal finance in 2026 covers that in detail. Every one of those prompts still runs on the physical infrastructure this article is about.

By the middle of 2026, the water side of that infrastructure is no longer a side note. Arizona froze a tax break that helped build its data center industry. Microsoft says it reached a major sustainability target a year early. And in rural Louisiana, hydrologists are running groundwater models on a Meta campus that has not even opened yet. This is where all three stories come from, and what the underlying numbers actually show.

What is an AI data center?

An AI data center is a large facility packed with servers, GPUs, and networking equipment built to run AI workloads. Training a model like GPT-4, serving ChatGPT queries, or running Google's Gemini assistant all happen inside these buildings.

They are not ordinary offices with a few servers in a back room. A modern hyperscale AI data center can cover hundreds of thousands of square feet and draw as much electricity as a small city. The companies that operate them, Amazon Web Services, Microsoft, Google, Meta, and Equinix, are spending at a genuinely historic scale. Between January and August 2024, Microsoft, Meta, Google, and Amazon collectively spent $125 billion on AI data centers. The same four companies are expected to spend up to $700 billion in capital expenditures in 2026 to fuel their AI buildouts. For a closer look at how the routers, switches, and networking gear inside these buildings actually fail once they are running at scale, our piece on the hidden side of network devices covers the failures ISPs rarely explain.

That level of infrastructure requires an enormous amount of power. Power generates heat. Heat requires cooling. Cooling requires water. That is the chain.

Why data centers need water

The servers inside a data center generate heat constantly. Left unchecked, that heat would destroy the hardware, so data centers use cooling systems to keep temperatures in a safe range.

The dominant method is evaporative cooling. Hot air from the server floor passes over or through water. The water absorbs the heat and evaporates, carrying the heat away. Approximately 80 percent of the water drawn into an evaporative cooling system is lost to evaporation. The rest returns to local water systems, sometimes at higher temperatures and with chemical residues from the cooling process.

That discharged water is not harmless. Cooling water that returns to the supply carries a higher concentration of dissolved solids, including calcium, chloride, and silica. These can affect the taste of drinking water, lower crop yields, and harm aquatic life.

AI workloads make this worse than traditional cloud computing. Power densities for advanced AI racks have been scaling fast. In 2023, the average sat between 25 kW and 40 kW per rack. By 2025, that figure had climbed well past 100 kW, with some frontier training clusters approaching 200 kW per rack. More heat per rack means more cooling demand per square foot.

Image Credit: Leonardo AI

How much water does a data center use

The numbers depend on the size of the facility, the cooling method, and the local climate. The scale is not small.

A typical 100 MW AI data center consumes 1.5 to 3.0 million cubic meters of water per year for evaporative cooling. The average American uses about 80 to 100 gallons per day, so a single 100 MW facility can match the residential water use of a city of 50,000 people.

Data centers are already one of the top 10 water-consuming industries in the United States. Current growth rates suggest that by 2030, AI data centers could drain between 731 and 1,125 million cubic meters of water annually, equivalent to the household water usage of 6 to 10 million Americans.

Nationally, the trajectory is steep. Direct US data center water use sits around 17.4 billion gallons annually, according to the Lawrence Berkeley National Laboratory's 2024 federal report, which projects that figure will reach 38 to 73 billion gallons by 2028, driven primarily by AI training.

That figure only covers water used directly at the facility. The same Lawrence Berkeley report estimated that an additional 211 billion gallons were consumed indirectly through the electricity required to power those same data centers in 2023, roughly 12 times larger than the direct figure.

Where the water actually comes from

Almost every article on this subject, including most sustainability reports, treats a gallon of water as one uniform thing. It is not. A data center pulling from a municipal drinking water system competes directly with residents for the same tap. One pulling from a groundwater well competes with farmers and other well owners drawing on the same aquifer, often invisibly, since aquifer decline can take years to show up in the data. One using treated wastewater effluent is not touching a drinking supply, but it is removing water that would otherwise recharge a river or be reused for irrigation downstream.

Companies do sometimes highlight the source as a mitigation. Meta has said its water use figures are drawn largely from municipal and groundwater permits rather than pristine drinking sources, and in Louisiana, the company has argued that its projected draw is comparable to what the land used when it was irrigated farmland. That is a genuine point in Meta's favor, and also not the full picture, because reclaimed or agricultural water diverted to cooling is still water no longer available for the crop or the river it used to serve.

Groundwater is the quietest version of this problem. A well can be pumped for years before the surrounding water table shows measurable decline, and by the time it does, the drawdown can extend well beyond the property line. That dynamic sits at the center of the Meta case study later in this article, where independent hydrologists modeled exactly that kind of delayed, spreading impact.

Source: US Environmental Protection Agency water withdrawal categories; Lawrence Berkeley National Laboratory 2024 Data Center Energy Usage Report

Two data centers can report the exact same annual consumption figure and represent very different levels of real-world risk, depending entirely on where that water was pulled from.

How does AI use water compared to traditional computing

Traditional data centers that run web servers or store files also use water for cooling. What changed is the density and intensity of AI workloads.

A standard Google Search query uses roughly 0.3 watt-hours of energy. Per-query energy estimates for ChatGPT-class queries range from approximately 0.3 to 3 watt-hours, roughly 3 to 10 times more than a Google search, depending on model size and serving infrastructure.

More energy means more heat. More heat means more cooling. More cooling means more water. The math is direct.

Image generation is estimated at around 23 mL of water per generated image, based on energy consumption data from the UC Riverside and UT Arlington research team. AI workloads running today are not just slightly more demanding than what came before. They represent a structural shift in how much thermal load a single building has to handle.

How much water does ChatGPT use per query?

This is where things get genuinely confusing, because the honest answer depends entirely on what is being measured.

The most widely cited academic figure comes from a research paper by Pengfei Li and colleagues at UC Riverside and UT Arlington, titled Making AI Less Thirsty. Their figure for generating a 100-word email with GPT-4 works out to approximately 519 milliliters of water, close to a standard bottle.

OpenAI CEO Sam Altman addressed this directly in a June 2025 blog post called The Gentle Singularity. Altman wrote that the average ChatGPT query uses about 0.34 watt-hours of electricity and roughly 0.000085 gallons of water, close to 0.3 milliliters, about one-fifteenth of a teaspoon. His figure accounts only for direct operational water at the data center level, not the water used in electricity generation or chip manufacturing.

The gap between 0.3 mL and 519 mL is not a measurement error. It is a scope decision. The 519 mL estimate includes indirect water use, the water consumed by power plants generating electricity for the data center. Altman's figure only counts water physically flowing through pipes at the facility itself, and OpenAI has not published the underlying methodology for outside verification.

Both numbers are defensible within their own definitions. The disagreement is about which costs should count.

Image Credit: Leonardo AI

Google has disclosed a comparable direct figure for its Gemini assistant, putting on-site water use per query in a similar fractional-milliliter range to Altman's figure of  ChatGPT. That is a direct, facility-level figure and, like Altman's, it does not include the power plant water.

For most users, the practical takeaway is this: a single query is negligible. ChatGPT has well over 100 million active users. Multiply any per-query number by billions of daily queries, and the real story shows up in the aggregate, not in any one prompt.

The ChatGPT data center: what OpenAI actually runs on

OpenAI does not operate its own data centers. ChatGPT runs on Microsoft's Azure cloud infrastructure. The water and power footprint of ChatGPT is embedded inside Microsoft's broader data center operations, which is part of why isolating a single company's contribution is harder than the headline numbers suggest.

A training run for GPT-4 in West Des Moines, Iowa, consumed 11.5 million gallons of water in July 2022 and 13.4 million gallons in August 2022. Across its five facilities in that city, Microsoft used 68.5 million gallons of water in 2024, making it the region's largest single user.

That is the cost of training a single large model. Inference, actually running ChatGPT queries for hundreds of millions of people, adds to that continuously.

According to Google's 2025 Environmental Report, its data center water consumption grew from 4.3 billion gallons in 2021 to 7.7 billion gallons in 2024, close to an 80 percent increase in three years. Across the same period, Microsoft's last fully disclosed aggregate figure, 1.69 billion gallons for fiscal year 2022, marked a 34 percent rise from the year before. These are not companies being reckless. They are companies growing at a pace that their sustainability pledges are still trying to catch up to.

Image Credit: Leonardo AI

Which AI data center companies use the most water

The hyperscalers, companies that own and operate massive data center fleets, are the biggest consumers. Here is where the publicly available data lands.

Google

Google's data centers consumed 7.7 billion gallons of water in 2024, up from 6.1 billion in 2023 and 4.3 billion in 2021. A single facility in Council Bluffs, Iowa, consumed about 1 billion gallons in 2024 alone, more than any other Google site. Including offices, Google's total 2024 water consumption reached roughly 8.1 billion gallons, a 28 percent jump from the prior year that the company says is equivalent to watering 54 golf courses annually in the arid Southwest. In 2024, Google replenished 4.5 billion gallons through water stewardship projects, bringing its freshwater replenishment rate from 18 percent in 2023 to 64 percent, with a stated goal of 120 percent by 2030.

Microsoft

Microsoft's last fully disclosed aggregate water figure was 1.69 billion gallons for fiscal year 2022, a 34 percent increase from the year before. Since then, the company has reported efficiency metrics rather than a fresh total consumption number, until a significant announcement in 2026. Microsoft said it reached water positive status for fiscal year 2025, meaning its water replenishment projects returned more water than its global operations consumed, five years ahead of its original 2030 target. According to Data Center Dynamics reporting on the announcement, Microsoft's average water usage effectiveness across its owned fleet improved from 2.3 liters per kilowatt hour in the early 2000s to 0.27 liters per kilowatt hour in 2025, a drop of nearly 90 percent. The company's primary cooling method is direct air cooling with evaporative assist, which uses water only when outside temperatures climb above 85 degrees Fahrenheit, and its newest data center designs, piloting in Phoenix, Arizona, and Mt. Pleasant, Wisconsin, since August 2024, use closed-loop, chip-level cooling that eliminates evaporative water loss at the facility level entirely.

Image Credit: Leonardo AI

Equinix

Equinix, which operated 268 data centers worldwide in 2024, reported withdrawing 1.4 billion gallons and consuming 1.2 billion gallons, or 85 percent of what it withdrew.

Meta

Meta consumed 813 million gallons of water company-wide in 2023, its most recent fully reported figure, with roughly 95 percent of that attributable to data centers. Meta's newest and largest facility, under construction in Richland Parish, Louisiana, is the subject of its own case study later in this article, because the gap between what that site is legally permitted to draw and what Meta says it will actually use is one of the clearest examples of a pattern that shows up across the entire industry.

The water rights problem nobody quotes in gallons

Every article on this topic, including the one you are reading up to this point, quotes current consumption. Far fewer quote the permitted maximum, which is usually much higher, because permits get sized for future expansion phases that the public rarely hears about at initial approval.

Meta's Richland Parish data center is the clearest public example. According to state records reported by NOLA.com, the facility is registered to consume more than 23 million gallons of water per day, or 8.4 billion gallons per year. If it drew that full amount for a year, it would use more water than all of Google's disclosed data centers combined in 2023. Meta says the actual figure will be far lower, between 500 and 600 million gallons per year once the campus is operational, an average of about 1.5 million gallons a day.

Both numbers can be true at once. The 8.4 billion figure is the legal ceiling. The 500-600 million figure is Meta's stated operating plan. The distance between the two is the risk that water researchers at LSU and Tulane flagged when the permit became public. Frank Tsai, director of the Louisiana Water Resources Research Institute at LSU, ran a 17-year simulation assuming Meta withdrew its maximum registered amount every day, and found that groundwater levels could drop more than 65 feet in some areas beneath the site, extending well beyond the facility's roughly 70 football field footprint. There is currently no state regulatory body in Louisiana tasked with monitoring the facility's actual draw against that ceiling over time.

Image Credit: Leonardo AI

Meta disputes that this is likely to happen, noting the property was previously irrigated cropland that used a comparable volume of water for agriculture, and that the company has pledged to be a water-positive company-wide by 2030. Both of those points are worth holding alongside the hydrologists' modeling, not instead of it. A similar pattern played out at a Meta data center in Georgia, where nearby residents reported their household taps running dry, a case Meta has disputed responsibility for, according to reporting cited by the Sierra Club.

The dynamic is not unique to Meta. Arizona offers a real-time example of how the incentive structure around water rights actually works. Governor Katie Hobbs called for a full repeal of the state's 13-year-old data center sales tax exemption in her January 2026 State of the State address, but the final budget deal, signed in June 2026, settled for a three-year moratorium on new applications rather than a repeal, running through June 2029. In the two weeks before the pause took effect, data center developers filed 113 new applications with the Arizona Commerce Authority, according to Axios Phoenix reporting, nearly matching the 123 applications the state had received across the entire prior 13 years. That surge is the same grandfathering logic seen in water permitting: once a company secures the right or the incentive, a later policy shift rarely claws it back.

The water you cannot see: embodied water in AI chip manufacturing

Almost every article on this topic stops at operational water use. Few trace the water consumed before a server ever boots.

Every AI training run depends on thousands of high-end GPUs, and those chips carry a water debt long before they reach the data center. TSMC, the world's dominant semiconductor manufacturer, consumes roughly 150,000 metric tonnes of water per day across its Taiwan operations. An average chip fab uses approximately 10 million gallons of ultrapure water daily, equivalent to the daily household consumption of 33,000 Americans, according to the World Economic Forum. Some of that manufacturing pressure traces back to a separate bottleneck in the AI supply chain, the high bandwidth memory shortage we detailed in our report on the AI chip war and the hidden HBM bottleneck, since fabs racing to meet memory demand are also the ones expanding water-intensive capacity fastest.

Image Credit: Leonardo AI

The water intensity scales with chip sophistication. When TSMC moved to 16 nm process nodes in 2015, water usage per unit increased by over 35 percent, according to S&P Global Ratings. More advanced nodes require more ultrapure water per wafer because fabrication involves more steps, each requiring rinsing. It takes roughly 1,400 to 1,600 gallons of municipal water to produce 1,000 gallons of ultrapure water for chipmaking.

TSMC's first Phoenix, Arizona fab already uses about 4.75 million gallons of water per day. Once all three planned Phoenix fabs are operational, combined daily demand is projected at roughly 17.2 million gallons, before a planned Industrial Reclamation Water Plant, expected online around 2028, reduces that draw through recycling rates the company targets at 90 percent or higher.

The 2021 drought in Taiwan illustrated what this dependency looks like under stress. Water shortages forced Taiwan's water authority to reduce agricultural supplies to prioritize chip production. From 2015 to 2019 alone, TSMC's total water consumption surged by 70 percent as chip complexity increased.

When water efficiency metrics do not tell the whole story

WUE, Water Usage Effectiveness, is the industry's primary tool for measuring how efficiently a data center uses water. Developed by The Green Grid and reported in liters per kilowatt-hour, a lower number signals better efficiency. Zero is the theoretical ideal, achievable only in fully air-cooled facilities.

The metric sounds authoritative, but there are structural problems with how it gets reported that most coverage skips entirely.

WUE only counts water at the facility boundary. A data center that purchases chilled water from a third-party municipal cooling utility does not count that water in its WUE, even though the same water is being consumed to serve that facility's cooling needs. Different operators can draw that boundary differently, and no uniform regulatory standard enforces a consistent definition.

Companies also report annual average WUE, which masks seasonal spikes. A facility in Phoenix might run a WUE of 0.5 L/kWh in January and 4.0 L/kWh in August. The annual average looks acceptable. The local aquifer in summer does not. Monthly or quarterly reporting gives a far more accurate picture of real-world water stress on local communities, but annual disclosure remains the industry norm.

Some colocation operators lease space to hyperscalers and do not report water consumption in their own disclosures. The hyperscaler reports server-level efficiency. The colo reports building-level performance. Nobody publishes the combined figure. Amazon does not disclose total water consumption at all. Microsoft has not published facility-by-facility water figures. Only Google provides individual data center numbers, and even those lack context on facility size and cooling technology used.

Claims of zero-water or water-free cooling also deserve scrutiny. Most such claims refer to eliminating evaporation at the facility level. The heat absorbed by closed-loop liquid cooling systems still has to go somewhere, often to dry coolers or rooftop heat exchangers that may use water depending on ambient temperature. Calling a facility zero-water because the cooling tower sits indoors rather than outside does not change the underlying physics.

Source: The Green Grid WUE framework; Lawrence Berkeley National Laboratory 2024 Data Center Energy Usage Report

An impressive WUE score can coexist with serious local water stress, depending entirely on how the measurement boundary is drawn.

Why cutting water use can raise the water bill somewhere else

Most coverage treats a switch to dry cooling or liquid cooling as a clean fix. It is not quite that simple. Air-cooled and dry-cooling systems use less on-site water but pull more electricity, especially in hot climates where chillers have to work harder. That extra electricity has its own water cost at the power plant, particularly in regions where the grid still leans on coal, natural gas, steam turbines, or nuclear, all of which use meaningful cooling water themselves. A facility can lower its on-site WUE and raise its total water footprint at the same time, and almost no mainstream coverage runs both numbers side by side.

The basic trade-off is straightforward once it is stated plainly. Evaporative cooling trades electricity for water. Dry cooling trades water for electricity. Which one is the better choice depends on the local grid's fuel mix, not on the technology itself. A facility going water-free on cooling in a coal-heavy grid region can increase its total water draw once the power plant's water use is counted, even as its on-site figure improves.

Climate matters as much as the grid. Dry cooling loses efficiency in hot, humid conditions, sometimes forcing supplemental water use anyway through adiabatic pre-cooling to hit the same server temperatures. A simple rule of thumb holds in most cases: a water-scarce region paired with a clean grid tends to favor dry cooling, while a water-abundant region paired with a fossil-heavy grid tends to favor evaporative cooling. Most real sites sit somewhere in between, which is exactly why a single-metric marketing claim, such as a company announcing it cut its WUE by 40 percent, should always prompt a follow-up question about what happened to electricity use and the grid's water intensity over the same period.

Why data centers keep getting built in water-stressed regions

Coverage of this issue focuses almost entirely on what AI companies consume. The more revealing question is why they keep building in dry places when wetter places exist, and 2026 gave a clear real-world answer.

Water availability sits below several other variables in most site selection decisions. Land cost, fiber proximity, permitting speed, labor market depth, and tax incentives consistently rank higher. Water-rich regions in the upper Midwest and Great Lakes area were slower to offer competitive incentive packages, so they lost builds to Arizona, Texas, and Virginia.

Image Credit: Leonardo AI

Arizona's incentive structure has been particularly aggressive, and 2026 is the year that started to change. The state enacted data center sales and use tax exemptions in 2013 and extended them through 2033 in a 2021 renewal. Companies qualifying for the exemption saved roughly 9 cents on every dollar spent on equipment in the Phoenix area. The incentive cost Arizona $38.5 million in foregone tax revenue in fiscal year 2025 alone, up from $1.4 million in 2020, with the Grand Canyon Institute projecting it would exceed $60 million annually by fiscal year 2027 if left unchanged.

Governor Katie Hobbs, who voted for the original exemption as a state senator in 2013, called for its full repeal in her January 2026 State of the State address, saying the incentive had already achieved what it set out to do. Arizona now ranks in the top 10 nationally for data centers, with close to 100 facilities operating and roughly 86 more planned or under construction. The Republican-controlled legislature did not agree to a full repeal. The budget signed in June 2026 instead froze new applications for three years, from July 2026 through June 2029, a compromise Hobbs says will save the state about $57 million that is being redirected to childcare, food assistance, and rural hospitals. Arizona is not alone. Illinois paused its own data center tax credits starting July 2026, Ohio has taken similar action, and Texas Governor Greg Abbott has outlined 2027 priorities that include requiring water-efficient cooling systems and repealing the state's sales tax exemptions for data centers, according to tracking by MultiState.

The regulatory window dynamic compounds the underlying problem. Once a company has a permit and a shovel in the ground, water restriction changes rarely force a relocation. Developers who moved fast in Phoenix, Goodyear, and Mesa before any municipal water caps took effect are effectively grandfathered in. There is still no federal mandate requiring water impact assessments before data center construction, and environmental review requirements vary by state, often waived for economic development projects above a certain job threshold. The financial incentives are real and immediate. The water costs are diffuse, slow-moving, and land on communities that had no vote in the original site selection decision.

What water positive pledges actually cover

Several of the largest AI infrastructure operators have made headline commitments to water positivity. Microsoft says it reached that status for fiscal year 2025. Google has committed to replenishing more water than its data centers consume by 2030. Amazon made a similar pledge at re: Invent 2024. These commitments deserve scrutiny before they are treated as finished achievements.

Image Credit: Leonardo AI

Replenishment credits, not local restoration

Most water positivity commitments are met through purchasing water replenishment credits, not through on-site physical restoration. Water consumed in Mesa, Arizona, can be offset by funding a watershed project in a different state or country entirely. The local aquifer in Mesa receives no direct benefit from that offset. A Kairos Fellowship investigative report found that Google is using annual savings from water stewardship projects elsewhere to offset actual consumption figures in its sustainability reporting, a structure that Microsoft's 2026 water positive claim relies on as well.

Renewable energy has its own water cost

Running on 100 percent renewable energy says nothing about the water consumed to generate that power. Nuclear, geothermal, and some solar thermal plants are water-intensive. Hydropower loses water through reservoir evaporation. The link between energy source and water intensity is not as clean as the marketing around it suggests.

Liquid cooling cuts evaporation, not the whole footprint

Liquid cooling reduces on-site evaporative water use, sometimes to near zero, and Microsoft's transition to chip-level closed-loop systems is genuinely meaningful progress. But closed-loop systems still require makeup water for losses over time, and the heat the fluid absorbs still has to be rejected somewhere, often to outdoor dry coolers that may use water at high ambient temperatures. Immersion cooling uses dielectric fluids that carry their own environmental handling requirements at the end of life. Treating the transition as a complete solution overstates what the engineering currently delivers.

Efficiency gains are being outpaced by growth

The data so far does not fully support the idea that efficiency alone will offset AI's growth. Google's WUE improved between 2021 and 2024, while its water consumption rose roughly 80 percent over the same period. Microsoft's WUE improved by nearly 90 percent since the early 2000s, and its most recent disclosed aggregate consumption figure still rose year over year before the company stopped publishing a fresh total. Efficiency gains are real. Capacity additions have been outpacing them by a substantial margin.

Sources: Data Center Dynamics on Microsoft's 2026 water positive announcement; Google Environment Report; Kairos Fellowship data center water analysis; The Green Grid WUE framework

Claims that do not survive scrutiny

The water positive pledges above get plenty of pushback already. These five claims circulate just as widely, including in otherwise careful coverage, and hold up less well than they first appear to.

Desert locations are usually blamed as the root cause, but the pattern is broader than that. Virginia, Georgia, and other water-rich states have already seen local water disputes and building pauses of their own, because the actual driver is competing local demand, not regional climate on its own. Loudoun County, Virginia, the densest data center cluster in the world, saw its facilities use 1.6 billion gallons in 2023, roughly 10 percent of all water consumed in the county, after growing 250 percent in four years, and that is one of the wettest parts of the country.

On-site water recycling is often described as a fix for evaporative loss, but treating cooling tower blowdown for reuse concentrates minerals and salts into a brine that still needs disposal, usually to a sewer system, a deep injection well, or an evaporation pond, each carrying its own environmental footprint. Recycling reduces the volume of fresh water drawn in, it does not eliminate the waste stream the process creates.

Isolating AI's water use from the rest of the cloud is harder than the per-query statistics suggest. Shared hyperscale facilities run AI training, AI inference, and ordinary cloud workloads on the same racks and cooling loops, so figures like the 519 mL and 0.3 mL numbers earlier in this article are modeled allocations, not metered measurements. No standardized methodology currently exists to split AI's share from everything else running through the same building.

The largest, most scrutinized companies are not necessarily the biggest consumers in absolute terms, only the most visible ones. Google, Microsoft, and Meta report because disclosure rules and investor pressure require it. The much larger population of smaller colocation operators and enterprise-leased data halls running AI workloads faces no comparable reporting requirement and stays effectively invisible in this conversation.

A declining company-wide WUE does not automatically mean local water risk is falling. A global average can improve while individual watersheds get worse, because the metric is a portfolio average blending facilities in low-stress and high-stress regions alike. Richland Parish is a useful check on that assumption, since Meta's company-wide sustainability numbers can look strong even as one specific aquifer absorbs a concentrated new draw.

Thermal density and the cooling threshold AI has already crossed

This section is technical. If the basics of data center cooling are already familiar, this is where the current AI infrastructure gets genuinely different from anything built before it.

The relationship between rack density and cooling water demand is not linear. Going from 10 kW to 40 kW per rack does not quadruple water use; it more than quadruples it, because cooling efficiency drops as heat load increases and ambient temperatures rise. The physics of heat transfer gets less forgiving at higher densities, not proportionally harder.

At around 100 kW per rack, air cooling reaches a thermodynamic limit regardless of airflow volume. The air simply cannot carry enough heat away fast enough. This is a physical ceiling, not a design preference or cost tradeoff. Facilities built or planned for air cooling before AI workloads arrived are now structurally underequipped, and many data centers considered modern in 2019 cannot safely host current-generation GPU clusters without retrofitting.

Image Credit: Leonardo AI

The challenge is worse in mixed-use colocation facilities. A colo running a combination of traditional cloud workloads at 10 to 20 kW per rack alongside AI GPU clusters at 100 to 200 kW per rack cannot cool uniformly. Hot spots develop. Water-based spot cooling has to be retrofitted around existing raised floor infrastructure, conduit runs, and structural columns, often at high cost and with real performance compromises, because the building was never designed for the thermal geography that AI creates.

NVIDIA's GB200 NVL72 rack, which houses 72 GB200 GPUs in a single cabinet, operates at densities that effectively make liquid cooling a baseline requirement rather than an upgrade option. The transition from optional to required happened faster than most facility planners expected, and several colocation operators signed long-term leases with hyperscale customers at density commitments they are now technically unable to meet without unbudgeted capital investment.

There is also a temperature quality problem that gets almost no coverage. Waste heat reuse, one of the industry's most promising solutions, depends on producing output water hot enough to be useful for district heating or industrial processes. Facilities optimized purely for low WUE at moderate rack densities often produce warm water at 35 to 40 degrees Celsius, while district heating networks typically need water at 60 to 80 degrees Celsius to be useful. At higher rack densities with direct liquid cooling, output water temperatures can reach 60 degrees or above, making it genuinely reusable. The data centers positioned to contribute to waste heat recovery are exactly the ones running the densest, most water-intensive AI workloads.

Source:  ASHRAE thermal guidelines; Nvidia GB200 NVL72 product specifications

Image Credit: Leonardo AI

Are data centers bad for the environment?

The honest answer depends on how the costs weigh against the benefits, and where the facility sits.

On the carbon side, the International Energy Agency estimates that data centers produced approximately 182 million tons of CO2 in 2024, about 1 percent of global energy-related emissions. A 2024 study of 2,132 US data centers found their average carbon intensity was 48 percent higher than the national average across all economic sectors, largely because many facilities sit in regions where the electricity grid still runs heavily on coal and natural gas. Some of that same grid strain is showing up in adjacent infrastructure, too, including the fixed and mobile networks we mapped in our guide to wireless communication's invisible infrastructure.

At the current growth rate for AI, an estimated 24 to 44 million metric tons of carbon dioxide will enter the atmosphere, equivalent to adding 5 to 10 million cars to US roadways, according to a 2025 report by Cornell University researchers.

On the water side, location matters enormously. A data center in water-rich Norway drawing from hydropower carries a very different footprint than one in drought-prone Arizona or New Mexico. In Loudoun County, Virginia, the densest data center cluster in the world, facilities used 1.6 billion gallons in 2023, roughly 10 percent of all water consumed in the county, after growing 250 percent in just four years.

Regulation is starting to catch up in some places. The Netherlands barred new hyperscale data centers, defined as facilities larger than 10 hectares with 70 megawatts or more of connected power, from most of the country under a zoning decree that took effect in January 2024, with exemptions limited to two designated municipalities. Singapore's Green Data Centre Roadmap set a target WUE of 2.0 cubic meters per megawatt-hour and requires a power usage effectiveness of 1.25 or better for new builds. In the US, over 190 data center bills were introduced across state legislatures in 2025, and Arizona's 2026 moratorium is one of the most concrete outcomes of that legislative wave so far.

The environmental picture is also not uniform across communities. Data centers are often located in regions with cheaper land and lower regulatory barriers, areas that tend to have lower-income populations who absorb the resulting air and water impacts. Unlike carbon emissions, the effects of a data center drawing down one region's aquifer cannot be offset by cleaner water somewhere else.

What is being done to reduce data center water usage

The industry knows the problem. Several solutions are already in deployment.

Liquid cooling and immersion cooling

Liquid and direct-to-chip cooling systems can reduce water use by up to 90 percent compared to traditional evaporative methods. Instead of using water to cool air that then cools servers, these systems bring liquid directly into contact with the chips. Closed-loop systems lose almost no water during operation. Microsoft's Fairwater design, piloting in Phoenix and Mt. Pleasant since 2024, is the most publicly documented example of this transition at scale, and it now underpins the company's 2026 water positive claim.

Geographic placement

Geographic optimization, locating facilities in naturally cooler or water-abundant regions, is becoming a key design principle for sustainable AI infrastructure, according to Arup's 2025 Foresight Report on Water-Conscious Data Centers. Finland, Sweden, and Norway have attracted significant data center investment precisely because ambient temperatures reduce cooling demand for most of the year. Satellite infrastructure is following a related logic in reverse, moving capacity off the ground entirely, which is part of what we covered in our 2026 review of Starlink's plans, pricing, and IPO track.

Waste heat reuse

European researchers published findings in March 2026 showing that data center waste heat can be redirected to power water purification and carbon capture, potentially making facilities water-positive in a direct physical sense rather than through purchased credits. That would mean a data center producing more usable water than it consumes, a genuine reversal of the current model. Viability depends on producing output water hot enough for industrial use, which current high-density AI clusters are beginning to achieve.

Scheduling and timing

Google and Microsoft are already training some AI models at night or in cooler regions to reduce cooling demand and evaporation losses. Cooler ambient temperatures reduce how hard the cooling system has to work. A 10-degree Fahrenheit drop in ambient temperature can reduce cooling water consumption by 15 to 25 percent at a facility using evaporative systems.

Industry collaboration

Microsoft, Google, Amazon, and Meta have signed on to the Data Center Innovation Initiative, led by nonprofit investor Elemental Impact, aimed at funding up to 10 startups developing technologies around cooling, energy storage, and low-carbon building materials. These solutions exist. The constraint is deployment speed relative to the pace of AI expansion.

The electricity side of the water problem

Water and electricity are linked in AI infrastructure. More power consumed means more heat generated, which means more water is needed for cooling.

National electricity consumption by data centers is projected to grow from roughly 4 percent to 12 percent of total US electricity use by 2030. In raw terms, US data center electricity consumption is expected to rise from 183 terawatt-hours in 2024 to 426 TWh by 2030.

The surge in demand has driven historic price increases in regional energy markets. The July 2024 capacity auction in the PJM regional market cleared at roughly $270 per MW-day, an 800 percent increase from the previous year's price of $30. By July 2025, prices reached the market cap of nearly $330 per MW-day.

In 2025, Microsoft, Google, Amazon, and Meta were projected to spend a combined $320 billion on AI infrastructure, more than double the $151 billion spent in 2023. That capital intensity is not confined to hyperscale campuses either. Consumer demand for AI-capable devices sits on the same growth curve, visible even in smaller stories, such as the reported $59 million in early deposits behind Trump Mobile's T1 phone launch, a reminder that AI-driven hardware demand runs from gigawatt-scale campuses all the way down to individual handsets. The power grid in many parts of the US was not designed for this kind of concentrated, high-density demand. Data center clusters in Virginia, Texas, and Arizona are already straining local grid capacity, and water scarcity in those same regions compounds the problem.

Building a full lifecycle water footprint

Everything up to this point in the article assumes a reader already understands the difference between direct and indirect water use. This section is for the smaller group that wants to know how an analyst would actually build a defensible total water footprint, rather than quoting a single company-reported figure.

Analysts who work on this professionally generally use a three-layer model. The first layer is direct operational water, the on-site cooling figure that every company discloses when it discloses anything. The second is indirect water, tied to the electricity consumed, calculated using a regional grid water-intensity factor that varies enormously by fuel mix. The third is embodied water, the water consumed in manufacturing the chips, servers, and the facility itself, the layer covered earlier in this article's chip manufacturing section.

The indirect layer is the hardest to pin down with precision. Grid water intensity changes by time of day, season, and which specific power plants are dispatched to meet marginal demand at any given moment, so a single annual average understates real variation. A data center drawing power from a coal plant at 6 p.m. on a summer evening has a very different indirect water cost than the same facility drawing solar power at noon.

The embodied layer is getting worse as a share of the total, not better, even as individual chips improve. GPU generations are refreshing faster than the servers that host them used to, which means the manufacturing water cost gets amortized over a shorter useful life, raising the effective water cost per unit of compute delivered even when the chip itself uses less power per calculation.

Some formal standardization exists for this kind of accounting. ISO 14046:2014 specifies principles and requirements for water footprint assessment using a life cycle assessment approach, and companies that pursue it often do so to strengthen their CDP Water Security disclosure scores. No AI company currently reports its data center water footprint against this full lifecycle standard.

A rough worked example makes the gap visible. Take a hypothetical 100 MW facility that discloses 2 million cubic meters of direct annual water use. Add an indirect estimate using published grid water-intensity research for its region, which might range from 8 to 20 million cubic meters, depending on the local fuel mix. Add a conservative embodied estimate based on its GPU refresh cycle, plausibly another 1 to 3 million cubic meters amortized annually. The result is a range, not a single verified number, typically several times larger than the direct figure alone, and it is a range that no company currently publishes for any real facility.

The global picture: water scarcity and AI on a collision course

The regions where AI data center growth is fastest are often the regions with the least water to spare.

Arizona has more than 370 golf courses in a desert climate, so data centers are hardly the first industry to make questionable water choices there. The scale of AI growth is different in kind. By 2030, global data center water consumption is projected to exceed 1.2 trillion liters annually, surpassing the total annual water use of London's 9 million residents.

The Li and Ren research paper projects that global AI demand will require somewhere between 4.2 and 6.6 billion cubic meters of water withdrawal annually by 2027. For reference, the entire country of Australia uses roughly 74 billion cubic meters per year across all sectors.

The IEA's Energy and AI report from April 2025 and analysis from MSCI covering 680 data center assets worldwide both flag water scarcity as a growing physical risk for the sector, not just an environmental concern but an operational one. If the water is not there, the cooling does not work. If the cooling does not work, the servers do not run.

What the numbers actually tell you

AI data centers are real infrastructure with real physical footprints. They consume water, generate heat, draw electricity, and discharge warmer, chemically altered water back into local supplies. That is how the cooling physics works right now.

The companies building this infrastructure, AWS, Google, Microsoft, Meta, Equinix, Digital Realty, and others, are not ignoring the issue. Immersion cooling, closed-loop systems, waste heat recovery, and geographic optimization are all real investments. The question is whether the pace of deployment matches the pace of expansion, and whether the legal water rights being secured today match what companies actually intend to use.

The data says efficiency alone has not closed that gap yet. Water consumption has risen faster than efficiency improvements have cut it in most disclosed cases. That may change. The engineering is improving, state-level regulatory pressure is now real and specific rather than theoretical, and the economic incentive to reduce cooling costs is significant. The gap between where the industry is headed and where it needs to be is still large enough to deserve close attention.

Every major AI company publishes annual sustainability reports, and several states now publish water permit filings that tell a different part of the same story. Reading both together, rather than either alone, is the only way to get an accurate picture.

Recent Articles from USABeam

• Real Madrid's World Cup Talent is scattered across Nine Countries. Barcelona is almost all in one
• Starlink in 2026: The Internet Network Quietly Redefining Global Connectivity
• Why the AI chip war may come down to a part nobody talks about