The AI Water Crisis: Inside the Thirsty Data Centers

Key takeaways

Microsoft's water use increased 34% to 6.4 million cubic meters between 2021-2022, largely attributed to AI training workloads.
Each ChatGPT conversation of 20-50 questions requires approximately 500 milliliters of water for cooling.
Data centers use evaporative cooling (high water use), mechanical cooling (high electricity use), or direct liquid cooling (most efficient).
Amazon reports water consumption only at company-wide level without facility breakdowns, while co-location facilities often go unreported.
Large model training runs lasting weeks on thousands of GPUs consume millions of liters of water for continuous cooling.
AI workloads are growing faster than other data center operations and will become a significant fraction of total water use within the decade.

How AI Uses Water

Every time a GPU runs a training job or processes an inference request, it generates heat. That heat must be removed or the chip will overheat and fail. Cooling is the fundamental reason AI data centers consume water. The connection between computation and water consumption is direct and unavoidable with current technology.

The most common cooling method in large data centers is evaporative cooling. Warm water from the data center is pumped to cooling towers on the roof, where it is exposed to ambient air. Some of the water evaporates, cooling the remaining water, which is then recirculated through the data center. The water that evaporates is consumed. It leaves the data center as water vapor and does not return.

The alternative to evaporative cooling is mechanical cooling, which uses refrigerant systems similar to those in air conditioners. Mechanical cooling uses electricity but very little water. It is more expensive to operate because electricity costs more than water. Data centers in water-scarce regions are increasingly adopting mechanical cooling despite the higher electricity costs. Data centers in water-abundant regions often use evaporative cooling because it is cheaper.

A third approach, increasingly adopted in new hyperscale builds, is liquid cooling directly applied to the chips. Water or other coolants are circulated directly through cold plates attached to the processors. This is significantly more efficient than air cooling and uses much less water than evaporative cooling. The challenge is that liquid cooling requires different data center infrastructure and is more complex to maintain. Adoption is growing but the installed base of older facilities remains predominantly air-cooled.

The water consumption of a given data center depends on its climate, its cooling method, and its size. A data center in Arizona running evaporative cooling consumes dramatically more water than the same sized facility in Iceland using outside air cooling. A facility with direct liquid cooling consumes dramatically less water than one with evaporative cooling. These variations make broad generalizations about AI's water footprint difficult, but the aggregate effect is substantial.

The Scale of Consumption

The numbers that have emerged from corporate disclosures and academic research paint a striking picture. Microsoft's global water consumption increased by 34% between 2021 and 2022, to 6.4 million cubic meters. It attributed much of this increase to AI training workloads. Google's water consumption increased 20% over the same period, to 5.6 billion gallons. These are global figures for their entire operations, but AI data centers are a growing fraction of the total.

A research paper published in 2023 by researchers from the University of California, Riverside and the University of Texas, Arlington, estimated that ChatGPT needs to 'drink' 500 milliliters of water for every conversation of 20 to 50 questions. At hundreds of millions of daily conversations, this adds up to hundreds of millions of liters per day, for a single application. The figures were estimates based on disclosed data center efficiency metrics and are subject to uncertainty, but the order of magnitude is striking.

Training runs consume even more concentrated bursts of water. A large model training run lasting weeks on a cluster of thousands of GPUs requires continuous cooling throughout. The water consumption of a single large training run is estimated to be in the millions of liters. As each successive generation of frontier models is larger than the last, training water consumption grows proportionally.

The broader data center industry consumed an estimated 600 billion liters of water globally in 2022, according to research from Virginia Tech. AI workloads are growing faster than the rest of the data center industry. If current trends continue, AI-related water consumption will be a significant fraction of total data center water use within the decade.

These figures are almost certainly underestimates. Corporate water disclosures cover only directly operated data centers. Co-location facilities, where data center operators lease space in third-party buildings, typically do not report their water usage in the operator's sustainability reports. A substantial fraction of cloud computing infrastructure, including portions of all major providers, runs on co-location. The actual water footprint of AI is likely higher than publicly disclosed figures suggest.

Microsoft, Google, and the Numbers They Buried

Microsoft's 2022 Environmental Sustainability Report disclosed the 34% increase in water consumption and, in a footnote, attributed much of it to AI training for GPT-3 and related models at a single facility in Iowa. The facility consumed approximately 700,000 liters of water on a day when the local temperature was high, and the cooling systems were running at full capacity. This localized consumption was noted and moved on from quickly in a report running to many pages.

Google's water consumption data is disclosed annually in its Environmental Report. The company has set a goal of replenishing 120% of the water it consumes in its operations by 2030, through investments in water restoration projects. Critics note that replenishment certificates, like renewable energy certificates, may not address the local water impacts of specific facilities. Replenishing water in a river basin in Oregon does not help a community in Arizona whose local water table is being depleted by a nearby data center.

Amazon has been less forthcoming. AWS's water consumption is reported at a company-wide level and is not broken down by facility or service type. The company has made commitments to water positivity but has not published the granular data that would allow independent verification. When researchers have attempted to estimate AWS's water consumption from disclosed efficiency metrics and facility sizes, the resulting figures are substantially larger than any publicly stated numbers.

The disclosure gap is significant for accountability purposes. Without consistent, granular, publicly available data on water consumption by facility and by workload type, it is impossible for regulators, communities, investors, or the public to hold companies accountable. The companies that are most transparent about their water consumption tend to be those who have the best performance to report. Opacity is often a signal that the underlying numbers are not favorable.

Some independent monitoring has become possible through satellite imagery and analysis of utility filings. Researchers have used this data to estimate data center water consumption where direct disclosure is absent. The estimates are imprecise but consistent: AI data centers are among the most water-intensive industrial facilities being built today. The pace of new data center construction, particularly in regions with limited water resources, is creating local impacts that are only beginning to be understood.

Building in Drought Regions

The American Southwest is experiencing a decades-long megadrought, the worst in 1,200 years according to tree ring data. Lake Mead, which provides water to 25 million people across Nevada, Arizona, California, and Mexico, reached its lowest level on record in 2022. And yet this is precisely where several large AI data center campuses have been built or are being planned.

Phoenix, Arizona, has become a major data center hub. The combination of cheap electricity, low taxes, and a business-friendly regulatory environment has attracted major investments from Meta, Microsoft, Google, and Amazon. These facilities require water for cooling. Phoenix is in a region where water scarcity is an increasingly urgent problem. The Arizona Department of Water Resources issued a formal finding in 2023 that a major Phoenix suburb lacks sufficient groundwater for planned residential development. The same constraints that limit residential water availability affect data center water availability.

Mesa, Arizona, has been a particular flashpoint. Meta's large data center campus in Mesa, combined with other tech company facilities, has led local officials and community groups to raise concerns about the impact on the Gila River aquifer, one of the primary groundwater sources for the region. The capacity for indefinite extraction from this aquifer has been questioned by hydrologists.

Texas has seen similar dynamics. Data centers have expanded rapidly across the state, partly attracted by its deregulated electricity market and low costs. Parts of Texas are also experiencing drought conditions and groundwater depletion. The Eagle Ford and other aquifers that provide water to communities in central and western Texas are under pressure from multiple sources, including agriculture, population growth, and industrial users including data centers.

The fundamental tension is between local water scarcity and the economic benefits that data centers bring. Companies make genuine financial contributions to local economies through taxes, employment, and infrastructure investment. Local governments are often eager to attract these investments. But the water impacts are diffuse, long-term, and often felt by communities that are not the primary beneficiaries of the economic activity. This mismatch creates political and governance challenges that have not been adequately resolved.

Community Impact: When a Town Loses Its Water

The impact of data center water consumption is most visceral in smaller communities where a single facility represents a large share of local water use. West Des Moines, Iowa, became a case study in 2022 when it was revealed that a Microsoft data center had consumed an unexpectedly large amount of water during a heatwave while training an early GPT model. The data center's water consumption on the hottest days approached the consumption of the entire surrounding community.

Urlingford, Ireland, a small town of approximately 1,500 people, found itself adjacent to a planned large-scale data center. Residents raised concerns about water impacts, noise from cooling equipment, and the incongruity of a facility consuming enormous resources in a community that would see limited direct benefit. Ireland has become Europe's primary location for hyperscale data centers, concentrated in a country with a relatively small population and limited water infrastructure.

In the Netherlands, data center operators in the region around Amsterdam have faced resistance from local governments concerned about water and energy impacts. The province of North Holland imposed a temporary moratorium on new large data centers in 2022. This was partly a water issue: the Netherlands has sophisticated water management infrastructure, but it is not unlimited, and the rapid growth of data centers in the region was straining it.

Indigenous communities in the United States have raised concerns about data center impacts on sacred sites and traditional water sources. In New Mexico, data center projects have been proposed near sites significant to Pueblo communities. The consultation processes required by federal law have been contested. The power dynamics between large technology companies with extensive legal resources and communities with limited resources to engage in regulatory proceedings are unequal.

The emotional dimension of these conflicts is often overlooked in corporate sustainability reports and policy discussions. Water is not just an input to be managed. For many communities, it is a source of cultural identity, economic security, and physical health. When a data center depletes a local aquifer or draws water that would otherwise be available for agriculture or residential use, the impacts are felt by specific people in specific places. These are not abstract externalities. They are consequences.

Industry Promises vs Actual Progress

Every major technology company operating large data centers has made water stewardship commitments. Microsoft has pledged to be water positive by 2030, meaning it will replenish more water than it consumes. Google has made a similar commitment. These are ambitious goals and deserve credit as genuine corporate commitments.

The mechanism for achieving water positivity is primarily through investment in water restoration projects: restoring wetlands, improving agricultural irrigation efficiency to reduce water consumption, and supporting water recycling infrastructure in communities near data centers. These projects are real and valuable. They do not, however, directly address the local water impacts of specific facilities. A wetland restoration project in Oregon does not replenish the groundwater drawn by a data center in Texas.

Progress toward these goals has been uneven. Microsoft disclosed in its 2023 sustainability report that its water consumption intensity, liters per megawatt-hour of energy used, had improved, but its absolute water consumption had continued to grow because of the expansion of its data center footprint. This is the efficiency-growth paradox: doing more with less per unit of activity, while doing so much more total activity that absolute consumption rises.

The replenishment accounting methodology is also contested. What counts as water replenishment? Is investing in a water recycling plant that reduces consumption elsewhere equivalent to replenishing the specific water drawn from a specific aquifer? Water, unlike carbon, does not mix globally. Water drawn from the Colorado River basin does not become available again because water was added to the Mississippi River basin. The local nature of water resources means that aggregate replenishment metrics may be misleading indicators of local impact.

Independent audit of water consumption and replenishment claims is limited. Unlike financial reporting, which is subject to rigorous audit requirements, environmental sustainability reporting is largely voluntary and self-certified. The lack of standardized methodology and independent verification means that water commitment claims vary widely in meaning and reliability across different companies. This makes comparison and accountability difficult.

Technical Alternatives to Water Cooling

The most promising technical alternative to evaporative cooling is liquid cooling, which circulates water or other coolants directly through the chips rather than through cooling towers that evaporate water into the atmosphere. Liquid cooling can reduce water consumption by 90% or more compared to evaporative cooling. It is also more efficient at removing heat, allowing chips to run hotter without performance degradation.

Immersion cooling takes liquid cooling further. Servers are submerged in tanks of non-conductive dielectric fluid, which absorbs heat directly from the chips without any airflow. The warm fluid is then pumped to a heat exchanger, which transfers the heat to a cooling loop. The secondary cooling loop can be closed, using water that is recirculated and not consumed. Immersion cooling can achieve near-zero water consumption for cooling and is being adopted in some specialized high-density deployments.

Air-side economizers use outside air directly for cooling when ambient temperature and humidity are low enough. In cold climates, data centers can be cooled by outside air for most of the year, requiring supplemental mechanical cooling only during hot weather. Iceland has become a destination for data center investment partly because its climate makes air-side cooling viable year-round. This approach works well in cool climates but is limited in hot ones.

Waste heat recovery is an underutilized opportunity. Data centers generate enormous amounts of heat that is typically rejected into the environment. This heat can instead be captured and used for space heating, industrial processes, or even electricity generation. Several European data centers are using waste heat to warm district heating networks, replacing natural gas heating in nearby neighborhoods. This makes the data center's thermal output productive rather than wasteful.

The economics of these alternatives are improving rapidly. The cost of liquid cooling systems has fallen as the market has grown. The operational savings from reduced water bills and reduced electricity costs for cooling can offset the higher upfront capital costs over the lifetime of a facility. As water becomes scarcer and more expensive in key data center markets, the economic case for water-efficient cooling will strengthen. Regulation requiring water efficiency would accelerate this transition.

What Can Be Done

The most important near-term action is improved transparency. Mandatory, standardized, facility-level disclosure of water consumption, including water consumed at co-location facilities and by subcontractors, would allow regulators, communities, and investors to make informed decisions. Voluntary reporting with inconsistent methodologies is insufficient. The SEC's climate disclosure rules, which took effect in phases in 2024 and 2025, cover some environmental metrics for public companies. Extending these requirements to cover water consumption in a rigorous way would be a significant step.

Siting requirements for new data centers should incorporate local water availability. Building a large evaporative-cooling data center in an area experiencing drought should require a rigorous hydrological assessment and community input. Some jurisdictions are moving in this direction. Consistent national standards in the United States would prevent companies from simply choosing the state with the weakest requirements.

Efficiency standards for data center water use, analogous to energy efficiency standards for buildings and appliances, could accelerate the adoption of water-efficient cooling technology. Water use effectiveness (WUE), measured as liters of water consumed per kilowatt-hour of energy used, is a standard metric. Setting performance targets and requiring operators to meet them within defined time periods would create clear incentives.

Pricing water appropriately is fundamental. In many parts of the United States and elsewhere, water is significantly underpriced relative to its true scarcity value. Artificially cheap water reduces the incentive for data center operators to invest in water efficiency. Market reforms that allow water prices to reflect scarcity, combined with social protections for essential uses like residential water, would create better incentives across all water users.

Individual users and organizations can play a role too. Choosing AI service providers with strong water stewardship practices, demanding transparency from providers about their environmental impacts, and advocating for stronger regulation are all meaningful actions. The AI water crisis is not inevitable. It is a consequence of choices about where to build, how to cool, and how to account for the true cost of AI computation. Different choices, made by companies, regulators, and users, can lead to different outcomes.