Someone on the internet is wrong, and it is apparently my job to fix it.
In this case, the someone is nearly everyone who has written about AI and water consumption over the past two years. The claim, which has achieved the status of received wisdom, is that AI data centers are drinking us dry. A widely-cited study from UC Riverside produced the headline figure that generating a 100-word email with ChatGPT consumes roughly 519 milliliters of water, about one standard water bottle. The Washington Post ran with it. Fortune ran with it. Dozens of breathless articles followed, describing data centers as “vampires” with a “voracious, inhuman appetite” for freshwater. The Lincoln Institute of Land Policy compared them to Dracula, which is a choice.
The Bloomberg analysis that followed was more careful, noting that data centers using evaporative cooling evaporate roughly 80 percent of the water they draw and discharge the remaining 20 percent back to wastewater treatment. This framing was intended to be alarming. It is alarming, if you do not think carefully about what evaporation actually means in a hydrological system. If you do think carefully about it, the picture becomes considerably more complicated, and the policy implications change substantially.
The concern is not entirely wrong. There are real and specific places where data center water use is a genuine problem, and those situations deserve serious attention. But the framing that has dominated the coverage, AI is consuming our water supply, gets the underlying physics wrong in a way that leads to the wrong policy conclusions. Let me explain what is actually happening.
What Data Centers Do With Water
Data centers generate enormous amounts of heat. Servers running at full capacity produce heat that, left unchecked, would destroy the equipment. Cooling is not optional. The question is how you do it.
The dominant method for large-scale data center cooling is evaporative cooling, sometimes called swamp cooling. Warm air is drawn through wet pads or cooling towers. Water evaporates, carrying heat away from the facility. The cooled air goes back over the servers. This is the same basic physics as sweating. It works, it is cheap, and it has been the standard approach for industrial heat management for over a century. Steam-generating power plants, steel mills, and petrochemical refineries have used evaporative cooling at scale for decades. We did not describe any of them as “drinking” water.
Here is the critical point that almost every alarming headline gets wrong: most of the water does not disappear. It undergoes a phase change.
Evaporation converts liquid water to water vapor. That vapor enters the local atmosphere, participates in the regional water cycle, and eventually returns as precipitation somewhere in the watershed. Globally, the water cycle is a closed system, water is neither created nor destroyed. At the local and regional scale, however, the story is considerably more complicated, and that is where the genuine environmental questions live.
The UC Riverside Number Is Being Misread
Before getting into the hydrology, it is worth correcting the viral number itself, because it is wrong by a considerable margin.
The UC Riverside study that produced the 519-milliliter-per-email figure is legitimate research. Shaolei Ren and his colleagues were doing serious work on a real question. But the number that went viral is not what the paper actually says. The study estimated 500 milliliters per 20 to 50 queries, not per query, and not per 100-word email. That works out to somewhere between 10 and 25 milliliters per query. Not a water bottle. About a tablespoon or two. And that figure was for GPT-3, running on Microsoft Azure data centers in 2022, using evaporative cooling systems that have since been significantly upgraded.
Sam Altman stated in February 2026 that a typical ChatGPT query uses approximately 0.3 milliliters of water. Independent analysts working through the same methodology with current hardware have landed around 5 milliliters as a realistic figure. Even at the high end, we are talking about a teaspoon, not a water bottle. The viral claim is off by somewhere between 100 and 1,700 times, depending on which version you encountered.
None of that means data centers have a negligible water footprint in aggregate. They do not. But the individual-level framing, your email is costing the planet a bottle of water, is wrong on the numbers and wrong on the framing simultaneously.
Consumptive Use and Why It Actually Matters
Here is where the hydrology gets important, and where most of the coverage goes wrong in a more fundamental way than just the numbers.
Water resource management distinguishes between water withdrawal and consumptive use. Withdrawal is the total volume of water removed from a source. Consumptive use is the portion of that withdrawal that does not return to the local watershed. The difference matters enormously for understanding actual environmental impact.
Irrigation agriculture is highly consumptive. Water applied to crops is taken up by plants, transpired through leaves, and exported as agricultural products or lost to evapotranspiration. Relatively little returns to the local aquifer or stream. That is why irrigated agriculture accounts for roughly 70 percent of global freshwater withdrawals and an even higher share of consumptive use.
Thermoelectric power plants withdraw enormous quantities of water for cooling, the US Energy Information Administration estimates that thermoelectric generation accounted for roughly 40 percent of all US freshwater withdrawals before the shift to air-cooled facilities, but most of that water is returned to the source, warmer than it left. The consumptive use fraction is much smaller than the withdrawal figure suggests.
Data centers using evaporative cooling sit somewhere between these cases. The 80 percent that evaporates is, at the local scale, consumptive, it leaves the immediate watershed as water vapor rather than returning to the stream or aquifer from which it came. The 20 percent discharged to wastewater treatment stays local and re-enters the regional system after treatment. So evaporative cooling data centers have a real consumptive use fraction, higher than a once-through cooled power plant, lower than irrigated agriculture.
Whether that consumptive use is a problem depends entirely on where the data center is.
The Watershed Question
No water system is perfectly closed, but some are effectively closed in ways that matter enormously for resource management.
The Great Lakes basin is the clearest example in North America. The basin contains roughly 21 percent of the world’s surface freshwater. Water withdrawn from Lake Michigan for industrial cooling, evaporated, and carried by prevailing winds still largely falls as precipitation within the basin. The Great Lakes Compact of 2008 manages diversions and consumptive use precisely because the basin functions as an effectively closed system, water that stays within it cycles back through the lakes, rivers, and groundwater. A data center in Chicago or Milwaukee drawing from Lake Michigan for cooling is operating within a water budget that is, for practical purposes, self-replenishing at regional scale.
Contrast this with a data center in the Phoenix metropolitan area drawing from the Salt River Project or from Central Arizona Project deliveries of Colorado River water. The Colorado River basin is not an effectively closed system for practical purposes, it is a system under severe structural deficit, where more water is allocated than flows. Lake Mead and Lake Powell have fallen from roughly 90 percent capacity in 2000 to around 30 percent today. Water evaporated from a Phoenix data center does not return to the Colorado system in any meaningful timeframe. It joins atmospheric circulation patterns that carry it east and north, eventually falling as precipitation over the Great Plains or the Gulf Coast. For the Colorado basin, that water is gone.
Bloomberg’s analysis found that roughly two-thirds of data centers built since 2022 are located in areas already experiencing water stress. That is not a coincidence. Data centers follow cheap land, cheap power, and favorable regulatory environments. Water stress has historically been a secondary consideration because water, unlike electricity, has been priced far below its actual scarcity value in most American markets. That pricing failure is the root cause of the problem, not AI per se.
The Thermal Problem Nobody Is Talking About
There is a water-related environmental concern with data centers that deserves more attention than it receives, and it is not the one dominating the coverage. It is not about consumption at all. It is about thermodynamics.
When evaporative cooling handles 80 percent of a data center’s heat rejection, the remaining 20 percent has to go somewhere. In facilities that discharge to municipal wastewater systems, the thermal load is largely attenuated through the treatment process, residence time in tanks and lagoons, biological treatment, and volume dilution all work to moderate temperature before the treated effluent reaches a surface water body. That pathway, while worth monitoring, is not where the acute ecological concern lives.
The concern lives in direct discharge situations. Large data centers, like large power plants, sometimes have sufficient scale and appropriate siting to hold their own National Pollutant Discharge Elimination System permits and discharge cooling water directly to a river, bay, or estuary. When that happens, the thermal load goes straight to the receiving ecosystem without attenuation. Elevated water temperatures reduce dissolved oxygen, alter the metabolic rates of cold-blooded aquatic organisms, disrupt spawning cycles, and in extreme cases can cause direct thermal mortality in fish populations. This is not a theoretical concern. It is a documented phenomenon with a half-century of regulatory history behind it.
We have been managing exactly this problem at nuclear and thermoelectric power plants since the 1970s, and the regulatory framework for doing so is well developed, well tested, and sitting right there waiting to be applied to data centers.
Calvert Cliffs Nuclear Power Plant, operating on the western shore of the Chesapeake Bay in Lusby, Maryland, draws cooling water directly from the Bay and discharges it directly back to the Bay. The water is clean, it is not radioactive, it has not contacted the reactor core, it has simply passed through heat exchangers that transfer thermal energy from the reactor cooling system to the Bay water. But it comes out warmer than it went in, and the Chesapeake Bay is one of the most ecologically sensitive and publicly monitored estuaries in North America. This has been managed successfully for fifty years through a combination of NRC licensing conditions and EPA permits under Clean Water Act Section 316. Anyone who wants to understand what successful thermal discharge regulation looks like can drive to Lusby and look at it.
Peach Bottom Atomic Power Station on the Susquehanna River in York County, Pennsylvania operates under the same framework. The Susquehanna is the largest tributary to the Chesapeake Bay by volume, and its thermal regime matters to the Bay’s ecology downstream. Peach Bottom has been drawing from and discharging to the Susquehanna since 1974 under permits that specify allowable thermal increments, monitoring requirements, and conditions under which discharge must be curtailed to protect aquatic life. The framework works. The science supports it. The river and the Bay continue to function.
Section 316 of the Clean Water Act provides two distinct regulatory tools. Section 316(a) governs thermal discharge standards, allowing site-specific alternatives to technology-based effluent limits where the discharger can demonstrate through biological studies that the proposed discharge will not impair the designated uses of the receiving water body. Section 316(b) governs intake structures, requiring that the location, design, construction, and capacity of cooling water intake structures reflect the best technology available for minimizing adverse environmental impact, specifically protecting aquatic organisms from impingement on intake screens and entrainment through the cooling system.
Both provisions are directly applicable to large data centers with direct surface water discharge. Neither is being consistently applied to them. This is not because the law does not reach data centers, it does, through the general NPDES permit framework, but because the regulatory attention and the permit conditions that power plants have operated under for decades have not followed data centers into their new role as major industrial water users.
The irony of the current coverage is almost elegant. Millions of words have been written about whether AI is consuming our water supply through evaporation, which is the portion of data center water use that participates in the water cycle and returns to the watershed, however redistributed. Almost nothing has been written about the thermal load in the discharge fraction, which is the portion that actually poses a documented ecological risk in direct discharge situations and which already has a mature regulatory framework that is simply not being applied. We are arguing about the wrong number and ignoring the right one.
This is not an argument for halting data center construction. It is an argument for applying the same environmental analysis to data centers that we have applied for fifty years to nuclear and thermoelectric power plants. The science is settled. The regulatory tools exist. The Chesapeake Bay and the Susquehanna River have been successfully managed under exactly this framework within living memory. There is no reason data centers should be exempt from it.
The Right Policy Questions
The coverage of AI and water has been asking the wrong question. “How much water does AI use?” is less important than “where is that water coming from, and what happens to it?”
A data center in a water-stressed basin using evaporative cooling and drawing from public water supplies is a genuine environmental problem. It is competing with municipal and agricultural users for a shrinking resource, and the consumptive fraction of its withdrawal is leaving the local system permanently at the scale of decades.
A data center in a water-abundant, effectively closed watershed using appropriate cooling technologies is a much more manageable situation. The water it uses cycles back through the regional system. The thermal load on its discharge can be managed through existing regulatory frameworks.
A data center anywhere that is using direct liquid cooling, circulating coolant through chips rather than evaporating water into the atmosphere, has a dramatically smaller water footprint than evaporative cooling regardless of location. Microsoft has committed to zero water cooling for its next-generation data centers. This technology exists and is being deployed. The question is whether market forces and regulatory requirements will accelerate its adoption or allow cheap evaporative cooling to continue dominating in places where it should not.
Three policy tools would actually help. Water impact assessments should be required as part of data center siting approval in water-stressed basins, with the same rigor applied to consumptive use analysis that irrigated agriculture projects face. Thermal discharge permits under Clean Water Act Section 316 should apply to data center discharge streams, not just power plant cooling water. Water pricing that reflects actual scarcity value, rather than the subsidized rates that currently prevail in many western markets, would change siting economics without requiring prescriptive technology mandates.
None of these require treating AI as uniquely dangerous. They require treating data centers as the industrial water users they are, subject to the same frameworks that govern other large-scale industrial water use. The regulatory tools already exist. The question is whether we apply them.
The Actual State of Affairs
AI data centers are not drinking us dry. The viral framing is wrong on the numbers, wrong on the physics, and leads to policy conclusions that are too broad to be useful.
The genuine concern is narrower and more actionable: evaporative cooling is the wrong technology in water-stressed basins, siting decisions are being made without adequate attention to local hydrological budgets, thermal discharge from data center cooling water is underregulated, and water pricing fails to signal scarcity in the markets where scarcity is most acute.
These are solvable problems with known solutions. They do not require stopping the buildout of AI infrastructure. They require applying the same environmental analysis to data centers that we apply to power plants, steel mills, and irrigated agriculture, industries that withdrew and returned far more water than AI data centers do today, and learned, over decades, to do it better.
The water cycle is not broken. In some places, we are asking it to do more than local conditions support. That is a real problem. It is also a much more tractable one than “AI is thirsty” suggests.