The industry-average server rack draws about seven and a half kilowatts. An NVIDIA GB200 NVL72 rack draws one hundred and twenty to one hundred and forty. It weighs roughly 1.36 metric tons, it cannot be air cooled, and NVIDIA does not offer an air-cooled version, because there is not one to offer. The magnets, specialty alloys, and minor metals inside it are a supply story of their own; the heat coming off it is this one.
That is the whole of the cooling problem stated in three numbers, and everything else follows from it. Air has a specific heat capacity of about one kilojoule per kilogram per degree Celsius and a density of roughly 1.2 kilograms per cubic meter. Water carries about four times the heat per kilogram and is roughly eight hundred times denser. Moving one hundred and forty kilowatts of heat with air requires a volume of air, at a velocity, that a data hall cannot physically supply and a fan array cannot economically push. The transition to liquid was not a sustainability decision. It was a thermodynamics decision that arrived with a sustainability press release attached.
Which brings up the thing that most coverage of data center cooling gets structurally wrong. The public argument is about water, treated as a scandal to be eliminated. The engineering reality is that water and electricity are two currencies for the same purchase, the heat has to leave the building through one of them, and every configuration on the market is a decision about which meter to run it through.
Why air cooling stopped working
Rack density is the variable that broke everything, and the curve is steeper than the public conversation registers.
A conventional enterprise rack in 2015 ran five to ten kilowatts. The Uptime Institute’s 2025 industry average sits around seven and a half, which tells you most of the world’s installed base has not moved. An eight-GPU H100 server draws roughly ten kilowatts on its own, putting an H100 rack in the thirty-five to forty-five kilowatt range. The GB200 NVL72 puts seventy-two Blackwell GPUs in one rack at one hundred and twenty kilowatts nominal, with observed full-load figures in the low one-thirties. NVIDIA’s Rubin-generation roadmap points at two hundred and fifty to nine hundred kilowatts per rack, and the industry is developing eight-hundred-volt DC distribution architectures for megawatt racks targeted around 2027. Whether those figures materialize is a separate question from whether they are being planned around, and infrastructure gets committed on the roadmap rather than on the shipped product, which is a structural feature of this buildout worth watching and which the long history of infrastructure projects sized against projected demand suggests is where the expensive mistakes live.
Air cooling has a practical ceiling somewhere in the twenty to fifty kilowatt range depending on how much you are willing to spend on containment, fan power, and floor space. Past that, the required airflow produces noise, vibration, and pressure differentials that become their own engineering problems, and the fan energy starts consuming a meaningful fraction of the power you were trying to deliver to the chips.
The individual processors tell the same story. The chips themselves depend on a supply chain with its own constraints, from the hafnium in the gate stack to the noble gases the lithography consumes, and the thermal envelope is fixed at design time long before a building exists. An H100 has a thermal design power around seven hundred watts. A B200 runs one thousand to twelve hundred. The GB300 generation pushes higher. At those densities a cold plate sitting directly on the package is not an optimization, it is the only mechanism with enough thermal conductance to keep the die below its throttling threshold.
So the first thing to understand about data center cooling in the AI era is that the choice was made by physics rather than by procurement. The semiconductor roadmap that produced these parts did not consult the facilities engineers, and the buildings had to follow.
The stranding problem that creates is worth naming early, because it recurs at every layer of this subject. A data hall built to twenty kilowatts per rack cannot host a Blackwell rack, and not for one reason but for four: the electrical distribution is undersized, the floor loading will not carry 1.36 metric tons in a rack footprint, there is no coolant distribution infrastructure, and the heat rejection plant is sized for a load that no longer exists. Retrofitting is frequently more expensive than building new, which is why so much of the current buildout is greenfield and why data center cooling design has become a constraint on where capacity can go at all.
What liquid cooling actually is
The term covers several distinct technologies with different costs, and conflating them makes the tradeoffs invisible.
Rear-door heat exchangers are the gentlest step. A radiator mounted on the back of the rack captures hot exhaust air and transfers the heat to a liquid loop, which means the servers remain air cooled internally and only the room-level heat rejection changes. It handles up to roughly forty or fifty kilowatts per rack and it is not sufficient at Blackwell density.
Direct-to-chip liquid cooling is the current standard for AI racks. Cold plates sit directly on the GPU and CPU packages, coolant circulates through microchannels inside them, and the heat moves into the liquid without ever entering the room air. It captures the large majority of rack heat, typically seventy to ninety percent, with the remainder still handled by air for memory, power supplies, and networking, which means a direct-to-chip facility still needs an air-handling system and still has a data hall with fans in it. Hybrid rather than replacement is the accurate description. This is what the GB200 NVL72 requires.
Immersion cooling submerges entire servers in a dielectric fluid. Single-phase immersion circulates the fluid to a heat exchanger. Two-phase immersion uses a fluid that boils at low temperature, exploiting latent heat for much higher heat flux, and it ran into a supply problem rather than a technical one: the fluorochemical fluids involved fall within the PFAS category, and 3M announced exit from PFAS manufacturing by the end of 2025, which removed a principal supplier from a market that had been building around those chemistries. That is a supply-chain constraint determining a thermal architecture, and it belongs alongside the other cases where a specialty chemical or material quietly gates a technology roadmap.
The plumbing that makes any of this work is the coolant distribution unit, which sits between two loops and is the piece of the architecture that the entire water argument turns on.
Every one of these approaches introduces failure modes that air cooling did not have. Liquid near energized electronics means leak detection becomes a safety system rather than a convenience. Coolant chemistry has to be maintained against corrosion, biological growth, and galvanic incompatibility between the metals in the loop. Filtration matters because a microchannel cold plate is a small aperture and fouling it destroys the thermal path. The cold plates themselves are typically copper, which puts them in the same industrial metals demand stream as everything else in the electrification buildout, and the quantities at fleet scale are not trivial. And the whole assembly introduces a serviceability problem, since a technician can no longer simply pull a hot-swap component without addressing the plumbing attached to it.
The two loops of data center cooling, and where the water goes
This is the distinction most coverage misses, and missing it makes the numbers unintelligible.
A liquid-cooled facility runs at least two separate loops. The technical cooling system is the closed loop that carries treated coolant from the coolant distribution unit through the cold plates and back. That loop is sealed. It is filled once, monitored for leaks, and it does not consume water in any ongoing sense. When a vendor says the cooling is closed-loop, this is usually the loop being described.
The facility water system is the second loop, and it is where the heat actually leaves. The coolant distribution unit transfers heat from the technical loop into the facility loop through a heat exchanger, and the facility loop has to dump that heat somewhere. The options are a cooling tower, which evaporates water, or a dry cooler or chiller, which uses mechanical refrigeration and fans, which uses electricity.
That is the entire argument, structurally. Direct-to-chip cooling moves heat out of the chip more effectively than air, and it does not by itself determine whether the building consumes water. A direct-to-chip facility rejecting heat through an evaporative tower still evaporates water. A direct-to-chip facility rejecting heat through dry coolers consumes almost none and spends more power doing it.
Which means the widely repeated claim that liquid cooling saves water is true only when the heat rejection stage changes as well, and the reason it usually does change is that direct-to-chip lets the loop run hot enough that dry rejection becomes viable. That is the actual mechanism, and it is more interesting than the marketing version.
There is a third loop in many facilities worth mentioning, since it accounts for a share of consumption that gets attributed to cooling. Humidification systems add moisture to data hall air to keep static discharge within tolerance, and that water is consumed outright. Modern designs have widened acceptable humidity ranges considerably and reduced this substantially, and it remains a line item in any honest data center cooling water accounting.
The physics of evaporating water, and why it is so cheap
Evaporative cooling looks wasteful and is thermodynamically excellent, which is why the industry adopted it and why abandoning it costs something real.
Vaporizing water absorbs roughly 2,260 kilojoules per kilogram at atmospheric pressure. That is the latent heat of vaporization, and it is an enormous amount of energy to move per unit mass with no compressor involved. A cooling tower works by exposing warm water to an airstream, allowing a fraction to evaporate, and returning the rest at a lower temperature. The energy that leaves is carried by the vapor.
The critical advantage is that evaporation can cool water below the ambient air temperature, down toward the wet-bulb temperature, which in a dry climate can be twenty degrees Celsius below the dry-bulb reading. A dry cooler cannot do that. It can only approach ambient, which means on a hot day a dry system either fails to reject enough heat or has to fall back on mechanical refrigeration, and mechanical refrigeration is where the power goes. That relationship is why an evaporative facility in a humid climate performs worse than the same facility in a dry one: high wet-bulb temperature collapses the advantage, which is why Gulf Coast siting carries a cooling penalty that Phoenix does not.
That relationship is why the industry built in Phoenix and Mesa in the first place. Arid climates have low wet-bulb temperatures, which makes evaporative cooling maximally effective, which makes the facility maximally power-efficient. The same aridity that makes evaporative cooling attractive is what makes water contentious there, and the siting logic and the political problem have the same cause.
Water consumption for an evaporative facility runs on the order of 1.8 to 1.9 liters per kilowatt-hour at industry average, with the best conventional designs reaching 0.3 to 0.7. At a hundred megawatts of IT load running continuously, the arithmetic gets large quickly.
Cooling towers also have a second water loss that gets omitted from casual descriptions. Evaporation concentrates dissolved minerals in the remaining water, and past a threshold the concentrated water has to be discharged and replaced, which the industry calls blowdown. Cycles of concentration, meaning how many times the water is recirculated before discharge, is the operational lever, and pushing it higher requires chemical treatment that then creates a discharge quality problem under the Clean Water Act. Water efficiency and water quality compliance are in tension, and the engineering that improves one frequently complicates the other.
WUE, PUE, and the trade nobody states plainly
The industry has two metrics and they move in opposite directions, which is the fact that most reporting omits.
Power usage effectiveness is total facility energy divided by IT equipment energy. A PUE of 1.0 would mean every watt entering the building reaches a chip. Evaporative designs achieve low PUE because evaporation does the cooling work without a compressor.
Water usage effectiveness is annual water consumption divided by IT energy, in liters per kilowatt-hour. An air-cooled facility with no evaporation has a WUE near zero and a PUE that is considerably worse.
Set them side by side and the substitution is visible. Improving WUE by eliminating evaporation degrades PUE by requiring mechanical cooling. Improving PUE by using evaporation degrades WUE by consuming water. These are not independent sustainability metrics that a well-run facility optimizes simultaneously. They are two readings on one tradeoff, and a facility reporting an excellent number on one of them without reporting the other is showing you half a ledger.
The aggregate direction is worth knowing. Lawrence Berkeley National Laboratory put the aggregate site WUE for the United States fleet at 0.36 liters per kilowatt-hour through 2023 and projected a rise to between 0.45 and 0.48 by 2028, even as individual facilities using the newest designs approach zero. Fleet-average water intensity is going up while best-in-class goes down, because the fleet is growing faster than the new designs are deploying.
The measurement conventions deserve a caution of their own. WUE as normally reported covers cooling and humidification at the site, which means it excludes the water embedded in electricity generation, the water used in chip fabrication, which is substantial and considerably more contaminated, and the water in construction. A metric that draws its boundary at the fence line will always make the fence line look good, and every party publishing data center cooling figures chose that boundary knowing it. The same boundary trick appears wherever an industry reports its own footprint, from by-product metals whose environmental cost sits with the primary producer to recycling rates that count manufacturing scrap and omit end-of-life.
The zero-water design, and what it actually costs
In August 2024 Microsoft announced a data center design consuming no water for cooling during operation, and the announcement is worth reading closely because it contains its own rebuttal.
The design uses chip-level closed-loop cooling with the coolant filled once at construction and recirculated continuously. Microsoft put the avoided consumption at more than 125 million liters per year per facility, with pilots at Phoenix, Arizona and Mount Pleasant, Wisconsin. Reported fleet WUE improved from 0.49 liters per kilowatt-hour in 2021 to 0.30 in 2024 and 0.27 in 2025.
Then the sentence that matters. In the company’s own announcement of the zero-water evaporation design, the stated result is a nominal increase in annual energy usage compared with the evaporative designs across the global fleet. The infrastructure engineering lead’s framing in trade coverage of the announcement was that moving from evaporative to mechanical cooling is expected to increase PUE, mitigated by chip-level cooling permitting warmer coolant temperatures and therefore high-efficiency economizing at elevated water temperatures.
That is the whole trade, stated by the party with the least incentive to state it. Water goes to zero. Power goes up. The mitigation is real and it is a mitigation rather than an elimination.
The warmer-water point is the genuinely clever part and deserves credit. Air cooling requires supply air in the range of eighteen to twenty-seven degrees Celsius, which forces cold coolant. A cold plate sitting on a die that tolerates much higher junction temperatures can run coolant at thirty to forty-five degrees, and warm coolant can reject heat to ambient air across far more hours of the year without a compressor running. Direct-to-chip does not merely move heat better. It raises the temperature at which the heat is available, which is what makes dry rejection affordable. That is the single most important sentence in the engineering of modern data center cooling, and it is almost never the one that gets quoted.
The water nobody counts
Here is the accounting problem that makes most of the public numbers incomparable, and it is the single most useful thing to carry out of this subject.
Thermoelectric power generation consumes water. Coal, gas, and nuclear plants reject waste heat through cooling systems that evaporate water at rates well above what a data center consumes directly per unit of electricity delivered. A facility that eliminates onsite evaporation and draws grid power from a thermal fleet has not eliminated the water. It has moved the water upstream, to a plant that reports it in a different document, under a different regulator, in a different county.
Which produces the uncomfortable version of the zero-water claim. A closed-loop facility with a slightly worse PUE draws more electricity, and if that electricity comes from thermoelectric generation, the additional offsite water consumed by generating it can exceed the onsite water saved. The scope boundary is doing the work, not the engineering.
That is not an argument against closed-loop design, and it is an argument for insisting that any water figure specify its boundary. Onsite consumption, onsite withdrawal, and total water footprint including generation are three different numbers, and coverage moves between them without notice.
It also explains why the cooling question cannot be separated from the power procurement question. A facility on hydro or wind has genuinely low water intensity in its power supply. A facility on combined-cycle gas does not. The cleanest way to make a data center water-efficient is to change what is generating its electricity, which is a decision made several layers away from the cooling plant. That is why geothermal generation and hydro attract data center siting for reasons that have nothing to do with carbon accounting, and why the nuclear restarts now being negotiated carry their own water footprint that rarely appears in the announcement. Grid-scale storage changes the arithmetic again, since a facility firming intermittent supply with flow batteries or lithium packs is buying a different water profile than one firming with a gas peaker.
Where the water comes from
Source matters as much as volume, and the public argument frequently conflates them.
Potable municipal supply is the contested case, since it puts a data center in direct competition with residential and agricultural users through the same infrastructure. Reclaimed and non-potable water is the mitigation the industry has leaned on hardest, with Google reporting reclaimed water at facilities in Singapore, Georgia, and elsewhere accounting for roughly twenty-two percent of data center cooling volume in 2023. Reclaimed water carries its own engineering cost, since treated effluent has higher dissolved solids and biological load than municipal supply, which means more aggressive chemical treatment, lower achievable cycles of concentration, and more frequent heat exchanger maintenance. It is the right answer environmentally and it is not free, and the treatment chemistry required is its own small specialty-chemical supply question at scale.
Groundwater is where the lecture title comes from and where the durable damage lives. An aquifer recharges on a timescale set by geology and precipitation, and pumping above the recharge rate is mining rather than use. The distinction that matters is between a facility drawing from a river with a permitted allocation and a facility drawing from a confined aquifer whose drawdown is measured in feet per year and whose recovery, if it happens, happens over decades. The same extraction arithmetic that governs any finite subsurface resource applies, and the fact that the resource is water rather than ore does not change the mathematics. The recharge arithmetic is the part that separates a permit dispute from a permanent loss. Land subsidence is the irreversible version: pumping a confined aquifer compacts the sediment, the pore space collapses permanently, and the storage capacity is destroyed rather than merely drawn down. Parts of the Central Valley and the Houston area have subsided by meters, and no amount of subsequent recharge restores the volume.
The siting overlap is the aggravating factor. The climates that favor evaporative cooling are arid. The regions with cheap land, cheap power, and permissive permitting frequently overlap with those climates. As of mid-2026, a substantial majority of the continental United States has been under drought conditions, and a meaningful share of new AI capacity is being built inside those regions.
The politics of a permit
The third lens is where the physics arrives at a county board agenda, and the pattern has become consistent enough to predict.
In August 2025 the Tucson city council unanimously rejected a data center project on water grounds. In South Carolina, residents near an overdrawn aquifer have pushed for limits on groundwater withdrawal by data centers. Comparable disputes have run in Arizona, New Mexico, Georgia, and internationally in Chile, Uruguay, and Spain. The pattern is consistent enough to be predictive: a facility announced with a jobs number and no water number, a permit process that surfaces the water number late, and a community that discovers the magnitude after the tax abatement has been negotiated.
The structural feature that makes these fights bitter is asymmetric information. Project water figures are frequently confidential during negotiation, sometimes under non-disclosure agreements signed by local officials, and a community is asked to approve a withdrawal whose magnitude it cannot verify. The economic development case arrives with employment figures, and the employment figures for a hyperscale facility are small relative to the capital, because a data center is an unusually capital-intensive and labor-light industrial installation. A multi-billion-dollar campus may employ a few dozen to a few hundred people in steady state, which is a ratio no traditional manufacturing plant of comparable capital would produce, and which changes what a county is actually being offered. The critical minerals buildout runs the same negotiation with the same asymmetry, and the counties on the receiving end have generally learned the same lessons at the same speed.
Water rights add a legal layer that varies enormously by jurisdiction. Prior appropriation in the western United States assigns seniority by date of first use, which means a new entrant gets a junior right that is curtailed first in a shortage, unless it purchases a senior right from an existing holder, which is generally agricultural. That transaction is legal, orderly, and locally understood as a data center buying a farm’s water. Whether that is a good trade depends on what you think the water was producing, and reasonable people in the same county reach opposite answers, which is why these fights do not resolve on technical grounds.
None of this is unique to data centers. It is the standard pattern for any large industrial facility that arrives in a rural county with an extraction requirement, and the towns that were built around a single extractive employer are the long-run version of the same negotiation. The difference is duration. A mine works an orebody for decades. A data center’s silicon has a useful life closer to five years, and what happens to a building sited on a water permit when the hardware inside it becomes obsolete is a question nobody has had to answer yet.
What the numbers do and do not say
An audit, because this subject has produced more circulated figures with unstated units than almost any other in current technology coverage.
Withdrawal and consumption are different quantities and get used interchangeably. Withdrawal is water removed from a source; consumption is water not returned to it. A once-through cooling system withdraws enormously and consumes little. An evaporative tower withdraws less and consumes most of what it takes. Any comparison between facilities using different figures is meaningless.
The claim that an AI query consumes about half a liter of water is the most-shared figure in the field and it rests on a specific 2023 analysis with specific assumptions about model, hardware, region, and whether offsite generation water is included. It is a defensible estimate under its own assumptions and it is not a constant of nature, and it circulates without any of them.
Training GPT-3 evaporated 700,000 liters is from the same body of work and carries the same caveats, and it describes one training run on one generation of hardware in specific locations, on a model that has since been superseded by architectures with different efficiency characteristics.
A data center uses as much water as a town of X people is a comparison that requires knowing whether the town figure is withdrawal or consumption, and whether the data center figure includes generation.
Liquid cooling eliminates water use conflates the technical loop with the facility loop, as above.
Zero-water means zero water is true at the fence line and not true at the system boundary.
And the reverse error deserves equal treatment. Data centers are draining the aquifers overstates a share that, nationally, remains small against agriculture, which accounts for the overwhelming majority of United States freshwater consumption. The legitimate concern is local and concentrated rather than aggregate, which is a different argument and a stronger one. A facility consuming a rounding error of national freshwater can still be the largest single withdrawal in its county, and county is the unit at which the permit gets issued and the aquifer gets drawn down.
What data center cooling is actually telling us
Assemble the pieces and the cooling problem stops looking like an environmental controversy and starts looking like a thermodynamics ledger with a political interface.
The heat is not optional. A hundred megawatts of IT load produces a hundred megawatts of heat, essentially all of the electrical input, because computation does not store energy. Essentially every joule delivered to a chip leaves as heat, which makes a data center a resistive heater that happens to produce tokens on the way through, and which is why the materials and thermal engineering constraints that govern any high-flux industrial process apply here without modification. That heat leaves through evaporation, which spends water, or through mechanical refrigeration, which spends power, or through ambient rejection when the temperature differential permits it, which is free and unavailable on a hot afternoon. There is no fourth option, and every design on the market is a weighting of the first three. Data center cooling is therefore not a problem that gets solved. It is a ratio that gets set, and the setting is made by climate, by power price, by water price, and by which of the two the local politics will tolerate.
The engineering is genuinely improving and the improvement is real. Warmer coolant temperatures from direct-to-chip cooling extend the hours per year when free rejection works. Closed-loop designs eliminate onsite evaporation. Heat reuse into district heating networks captures value that would otherwise be rejected, which works where the district heating network exists and is a European result more than an American one so far. The physics constrains it: rejected data center heat arrives at thirty to forty-five degrees Celsius, which is useful for space heating and useless for industrial process heat, so the addressable demand is space heating in cold climates near the facility, which rules out the higher-value industrial process heat applications entirely. It is a genuine efficiency gain in Denmark and Finland and mostly unavailable in Arizona, and it is the closest this industry comes to the closed-loop recovery that works elsewhere only when the waste stream is concentrated and arrives at a known address.
Air-side economization is the other lever and it is climate-bound in the same way. A facility in a cold dry climate can reject heat to outside air directly for a large fraction of the year, which is why the Nordics and the upper Midwest attract capacity, and why siting decisions are as much a thermal calculation as a real estate one.
What is not improving is the aggregate, because capacity is growing faster than efficiency. Fleet WUE is projected upward even as best-in-class approaches zero, and that gap is the whole story of this buildout in miniature: the newest facility is always the most efficient one, the fleet average is set by everything already built, and the fleet is doubling. Efficiency per unit is a real achievement and it is not a trend line that intersects the aggregate, which is the same relationship that shows up in every materials system where demand growth outruns intensity improvement.
Which leaves the question the ten-lecture briefing on how AI data centers work exists to work through. The cooling stack determines how much power the building needs on top of the compute. The power determines what has to be built to supply it. And the thing being supplied is a load that did not exist five years ago, arriving at a grid that spent twenty years planning for flat demand.
The useful discipline, and the one that survives contact with every figure in this subject, is to demand three things of any water number: the boundary it was measured at, whether it is withdrawal or consumption, and what the corresponding power number did. A water figure without its power figure is not a measurement of environmental impact. It is a measurement of which meter somebody chose to read.
A cooling tower in Arizona and a combined-cycle plant two hundred miles away are consuming water for the same reason: something has to absorb the heat, and evaporation is the cheapest absorber available. Move the tower indoors and the absorber moves to the power plant. The bill does not disappear. It changes address.
Leave a Reply