Two numbers crossed my desk this week. Only one of them is real.
The first is the headline: US data centers will quadruple their water consumption by 2028, and scarcity is coming for the communities that host them. The second number never made the headline. A well-run liquid-cooled facility consumes roughly 0.1 to 0.3 liters of water per kilowatt-hour. An evaporative-cooled one sitting in the Arizona desert burns through 1.5 to 3.0. That is a 10x spread, and it lives entirely inside the cooling-technology decision. The "quadruple" story assumes that decision stays frozen. It won't.
I have watched this movie before. Between 2019 and 2022, the identical script was run against Bitcoin mining. Headlines promised miners would drink the grid dry and boil the lakes. The reality was messier, more technical, and far less cinematic. Mining migrated, cooling matured, and the industry that was supposed to collapse simply relocated to where power was cheap and stranded. Now AI has inherited the same infrastructure, the same water, and the same narrative — only with a bigger balance sheet and a friendlier press.
The story, and what it leaves out
The piece that triggered this is thin by design. It carries four information points, three of which are unsourced data or author opinion. There is no named research house, no methodology, no audit trail. That is not a reason to dismiss the subject. Water is a genuine constraint on the AI buildout. It is a reason to be precise about what is actually being claimed, because a scare number that travels faster than its evidence does real damage to real capital allocation.
Start with the metric everyone throws around: WUE, water usage effectiveness, measured in liters per kilowatt-hour. It is the water twin of PUE, power usage effectiveness. A data center that runs WUE of 0.2 is a rounding error on the local water ledger. One that runs WUE of 2.5 is an industrial user. Same racks, same chips, same workload. The difference is plumbing.
There are three cooling routes in production today. Evaporative cooling trades water for electricity: it uses the latent heat of evaporation to reject heat, which is cheap in capex and brutal in water consumption, especially in dry, hot regions like Arizona, Texas, and Nevada. Air cooling with mechanical chillers reverses the trade — low water, high power, which pushes the water footprint upstream into the power plant. Liquid cooling, whether direct-to-chip cold plates or full immersion, drives water consumption toward zero and can push PUE down to 1.05-1.1.
The headline attributes the quadrupling to data center growth. The engineering attributes it to cooling-route selection. Those are not the same claim. One is a forecast. The other is a decision, and decisions can change.
The crypto R&D lab nobody credits
Here is where the crypto industry's institutional memory becomes an asset rather than a liability.
Between 2018 and 2023, Bitcoin miners were the involuntary R&D lab for high-density thermal management. They were the first commercial operators to run immersion cooling at scale, the first to experiment with heat reuse for district heating and greenhouses, and the first to sign stranded-energy PPAs with hydro and flare-gas operators. When the AI buildout arrived, it did not start from zero. It inherited a cooling supply chain that crypto had already stress-tested, and it inherited the operators who knew how to run it.
The numbers on rack density make the switch unavoidable rather than optional. Traditional enterprise racks ran 5-10 kW. AI training clusters now run 40-100+ kW per rack, with some Nvidia GB200-class deployments pushing higher. That is physically beyond the capacity of air cooling. You cannot move that much heat with fans alone, no matter how clever the airflow design. The industry is being pushed into liquid cooling by thermodynamics, not by ESG marketing.
So the honest question is not "will data centers use more water." It is "how fast does the cooling mix shift from evaporative to liquid, and what does that do to the water curve."
Run the arithmetic. If the installed base grows 3x but the liquid-cooling share climbs from single digits to 40%+, the water intensity per unit of compute can fall even as absolute compute explodes. The "quadruple" projection is a linear extrapolation of a variable that is not linear. It is the same modeling error that produced a decade of wrong Bitcoin energy forecasts — and, I will note, the same error that made those forecasts so easy to sell.
Water is a power story in disguise
The water-energy-carbon nexus is the part the story buries.
Data centers are power-dense facilities. Their water use quadrupling implicitly assumes their electricity use quadruples. LBNL puts US data center electricity at roughly 176 TWh in 2023, about 4.4% of national consumption, rising to 325-580 TWh by 2028 — 6.7% to 12%. That is the real curve, and it is electricity, not water.
Electricity carries water with it. Thermal power plants — coal, gas, nuclear — withdraw 1.2 to 2.5 liters per kilowatt-hour for cooling. Wind and solar consume almost nothing. This means that a large share of a data center's true water footprint is indirect, incurred at the generator, not at the campus. The water bill is partly the power grid's bill, and it is settled by the generation mix, not by the data center's plumbing.
The piece never mentions the supply side. That is its largest blind spot. If the new load is met with gas and coal, indirect water rises. If it is met with renewables and nuclear PPAs, indirect water falls. Microsoft, Google, and Amazon have spent the last three years signing nuclear and renewable contracts precisely because they understand this coupling. The data center is not the villain in the water story. The generation mix is a co-author.
Now separate withdrawal from consumption, because this is where the shock number gets its leverage. Withdrawal is water taken from a source. Consumption is water that evaporates and never returns. A closed-loop facility can withdraw a large volume and consume a tiny fraction, because the water recirculates. Evaporative cooling is different — it consumes what it withdraws, because the whole mechanism is evaporation. A "quadruple" claim that does not specify withdrawal versus consumption is not a measurement. It is a mood.
And the mood is amplified by a scale illusion. Data centers are a small slice of national water withdrawal — typically well under half a percent. When you quadruple a small base, you get a headline that sounds apocalyptic and an absolute number that still sits below agriculture, which accounts for roughly 37% of US freshwater withdrawals, and thermoelectric cooling, which takes about 40%. The growth rate is real. The framing is theater.
The scarcity that actually binds is not physical. It is institutional. In stressed basins — the Colorado River, Arizona groundwater — the binding constraint is the water permit, not the water molecule. You can find water. You cannot always find a permit to withdraw it. That is a regulatory bottleneck dressed as a hydrological one, and the story's vague reference to "legislative actions" is gesturing at exactly this without naming it.
Which raises the price question. American industrial water is priced far below full-cost recovery almost everywhere. Data centers spend well under 1% of operating cost on water. That makes them price-insensitive and permit-sensitive. It also explains why the pressure arrives as legislation and permitting rather than as a price signal. When a market cannot clear on price, it clears on law. The story mentions legislation. It never explains why legislation is the only channel left.
Regulation is moving on two tracks, and neither is a subsidy. The United States has no federal water framework for data centers; it runs on state water-rights doctrine and local permits, which is why Arizona, Virginia, and Texas look nothing alike. The European Union goes the other way, with the Energy Efficiency Directive forcing data centers to report PUE, WUE, and renewable share, tied into the Water Framework Directive. Singapore showed the hard edge of the toolkit by pausing new builds outright for three years. The common thread is disclosure plus permit tightening plus, at the extreme, a building moratorium. None of that is priced into a growth story that assumes unlimited siting.
There is a second, quieter mechanism working in the opposite direction. Carbon policy raises the cost of thermal generation, which pushes data centers toward renewables, which lowers their indirect water footprint. Stronger carbon constraints and lower water footprints are not in conflict here. They are aligned. The decarbonization of the grid is also a decongestion of the water map. The story's water-crisis frame misses the relief valve that carbon policy quietly opens.
There is a third mechanism, and it is the one that keeps me up at night. Green electricity certificates cover power attributes. They cover nothing about water. A data center can run 100% renewable and still be a high-water facility, because the green certificate does not touch the cooling loop. The green halo can hide a water bruise. Carbon is visible and hedgeable. Water is invisible and unhedged. That asymmetry is the real ESG risk in this story, and it is nowhere in the headline.
Backup power is evolving along the same axis. Diesel generators are being displaced by lithium UPS systems and, at the frontier, hydrogen fuel cells and long-duration storage — Microsoft and Equinix have both tested the latter. This is the point where carbon neutrality and reliability stop competing and start reinforcing each other, because the same equipment that cuts on-site emissions also cuts on-site water and improves uptime. It is a rare case where the ESG story and the engineering story point the same direction.
The externality ledger
Follow the money and the water story becomes a profit-distribution story.
On the data center water chain, value and cost are distributed with almost perfect inversion. The compute layer — Nvidia, the hyperscaler margins, the AI application layer — captures the overwhelming share of the value. The infrastructure layer — power, water, cooling, land — absorbs almost all of the externality. The chip gets the multiple. The municipal water utility gets the bill.
This is the classic structure of an AI boom: returns privatized, costs socialized. The cooling vendors are the exception, because they sit on the monetizable side of the constraint. The utilities are the trap, because rate regulation caps their recovery while their expansion obligations explode, producing an under-investment loop that makes the bottleneck worse. The "quadruple" water cost is eventually paid by local ratepayers and utility shareholders, while the AI value chain keeps its margin. The story never asks who pays. That is the most important question in it.
Then there is the geography. The narrative is American, but the resource endowment is global. Nordic and Canadian sites trade on cold air and hydro. The Gulf states are building AI capacity on cheap power and desalinated seawater — an explicit "energy for water" arbitrage. Singapore ran a moratorium on new data centers from 2019 to 2022 while it audited its own resource limits. The scarcity is real in Phoenix. It is a non-issue in Iceland. A single national projection flattens a global rebalancing into a local panic.
The demand curve nobody is modeling
I have spent the last year arguing that the next wave of demand will not come from humans clicking. It will come from autonomous agents.
If AI agents become the primary liquidity providers in DeFi — and I think they will, well before 2026 — they will run continuously, settle on-chain, and consume compute around the clock. That is a fundamentally different load profile from a human trading desk, which idles overnight. Continuous autonomous demand does not just raise average utilization. It raises the floor, which raises the baseline water and power draw per unit of deployed hardware. The resource question is not a 2028 question. It is a permanent structural feature of an always-on machine economy.
This is why the cooling-route decision matters more than any single forecast. In a world of continuous load, thermal design is not a cost center. It is the difference between a facility that scales and one that gets permitted out of existence.
The contrarian read
Here is the counterintuitive part, and it is where most analysts get the direction wrong.
The water scare is not a warning about data centers. It is an advertisement for the cooling upgrade cycle. The tighter the constraint, the faster the switch from evaporative to liquid, and the faster the water intensity curve bends down. The companies that sell cold plates, immersion tanks, and heat-reuse systems are the direct beneficiaries of the very anxiety the story manufactures. Vertiv, Schneider, Delta, and a cluster of Asian thermal specialists are not victims of the water narrative. They are its monetization. Liquid cooling is projected to compound at 20-30% annually through 2030, and this headline is part of the marketing.
The same logic applies to the crypto-native side. The miners who survived 2022's liquidity vacuum did so by owning their energy and their thermal design. That vertical integration is now the template for hyperscalers, who are building their own plants, signing their own nuclear PPAs, and in some cases designing their own cooling. When a resource becomes a bottleneck, control of the resource becomes the moat. The competition stops being about chips and starts being about access.
And the timing is being reported in the wrong order. Grid interconnection queues in PJM, ERCOT, and Dominion territory already run four to seven years. Some utilities have begun deferring supply to large data centers outright. Electricity is the binding constraint before water ever is. The "2028" in the headline is not a water cliff. It is the year the AI buildout's resource constraints — power first, then water, then land, then permits — all come due at once. Water is the lagging indicator wearing the leading headline.
What to watch
Watch the cooling mix, not the water headline. The number that matters is not how many liters data centers will drink by 2028. It is what share of new racks ship liquid-cooled, and how fast the stranded-power map gets repriced. The same forensic lens that let crypto miners survive the last resource panic applies here: follow the physics, ignore the mood, and price the externality before the regulator does. The panic is loud. The engineering is quiet. The engineering wins.