The Water Question: AI's Quiet Reckoning With the Hydrological Constraint

The number Amazon finally put on paper

This week Amazon disclosed that its data centers consumed roughly 2.5 billion gallons of water in 2025. The number landed in a sustainability report, was picked up by Bloomberg, dissected on Hacker News, and reframed by Data Center Dynamics and The Register within a day. The interesting thing about the disclosure is not the number itself — it is the fact that we now have one at all, and the contortions required to publish it.

Amazon framed the figure as “small.” Critics framed it as enormous. Both can be partially right, which is exactly the analytical problem with water in the AI buildout. Unlike power, which is measured in watts and dollars and reported quarterly in 10-Ks, water is measured locally, priced politically, and accounted for inconsistently. A gallon withdrawn from a recharging aquifer in coastal Oregon is not the same asset as a gallon evaporated in the Arizona desert, and 2.5 billion of either is meaningless without knowing what watershed they came out of.

That is the structural question worth chewing on this weekend. The AI buildout has spent two years training the industry — and the press, and regulators, and capital markets — to think about scaling in terms of one binding constraint: power. Power gets the headlines, the FERC filings, the gigawatt commitments, the nuclear PPAs, the Oracle capex jitters that rattled markets this week after the company guided to a roughly $70 billion buildout. But beneath the power story, a second constraint is quietly hardening into the binding one for an increasing share of new projects. Water is becoming the resource that decides whether a site advances, whether a community resists, and whether the cooling architecture the industry just standardized on actually survives the decade.

This essay is about that second constraint — its physics, its accounting, its politics, and the architectural choices the industry is making right now that will determine whether water becomes a tractable cost line or a hard ceiling on where AI compute can be built.

How we got here: the brief, accidental career of evaporative cooling

For roughly the last fifteen years, hyperscale data centers solved their thermal problem by exploiting one of the cheapest physical effects on the planet: the latent heat of vaporization of water. Evaporative cooling — whether through direct cooling towers or adiabatic assist on dry coolers — works because evaporating a kilogram of water absorbs roughly 2,260 kilojoules. That is, in energy terms, an absurd bargain compared with the electrical work required to chill the same air mechanically.

The bargain is also why the industry quietly became one of the most water-intensive forms of light industry in regions that, on paper, host “knowledge work.” Power Usage Effectiveness (PUE) — the metric the industry trained itself on — measured electricity, not water. Water Usage Effectiveness (WUE), introduced by The Green Grid in 2011, was always the orphan metric: nobody had to report it, nobody priced into capex models, nobody underwrote a project on it. Hyperscalers reported WUE figures when it suited them and aggregated globally so that a Singaporean facility evaporating municipal drinking water and an Irish facility on free cooling collapsed into a single, flattering ratio.

That worked while the industry was small relative to its host watersheds. It is not working anymore. The 2.5-billion-gallon Amazon disclosure is what happens when a single firm’s annual consumption begins to register against the surface-water budgets of the counties it sits in. And the disclosure is itself a tell: Amazon almost certainly would not have published the number absent the wave of municipal opposition that has made water the single most effective lever for blocking projects.

We saw the lever pulled this week. Sentinel DC walked away from a paper-mill conversion project in Maine, after the state legislature passed and the governor vetoed a moratorium that was, at its core, an argument about water and grid. Sentinel’s withdrawal even after the veto is the more telling fact: developers are now reading not just the law but the political weather, and water is the headwind.

What the buildout is doing to the hydrological balance sheet

To make the constraint legible, separate two numbers that are often conflated.

The first is withdrawal — total water pulled from a source. Most of this water, in a once-through or partially recirculating system, eventually returns to the watershed, warmer and possibly chemically altered, but liquid.

The second is consumption — water removed from the local hydrological cycle, typically by evaporation. Consumed water leaves the watershed entirely; it returns somewhere, but as precipitation that may fall hundreds or thousands of kilometers away, on a timescale measured in days to weeks.

Hyperscale evaporative cooling is overwhelmingly a consumption story. When Amazon discloses 2.5 billion gallons, the relevant follow-up question is what fraction was consumed, where, and from what kind of source. The headline number obscures the gradient that actually matters: a gallon evaporated from a stressed aquifer in the high desert is a withdrawal from a stock that took millennia to fill. A gallon evaporated from a Pacific Northwest river fed by snowmelt is, on annual timescales, closer to a withdrawal from a flow.

The buildout is concentrating in places where the gradient is bad. This week’s source feed includes a 640-acre, 1.35-million-square-foot data center approved in Iron County, Utah — high desert, low precipitation, an aquifer system already stressed by agriculture. Jinko Power’s reported 1 GW solar-powered AI campus in western China sits in similar terrain. The economic logic that drives projects to these locations is clean: land is cheap, solar is abundant, and large parcels are permittable. The hydrological logic is the opposite. The same dryness that makes the air a good heat sink for evaporative cooling makes the water that goes into that cooling the least replaceable water on the system.

This is the underappreciated structural twist. Evaporative cooling becomes thermodynamically more effective as ambient humidity drops — that is its appeal in arid regions. But the regions where it works best are the regions where the water it consumes is most scarce. The technology and the geography are in physical disagreement with each other, and the disagreement is getting sharper as the parcels get larger.

The power-water entanglement nobody fully prices

The conversation about water and the conversation about power are usually held in separate rooms, by separate teams, with separate spreadsheets. They should not be. The two constraints are coupled in ways that make optimizing for one without modeling the other actively misleading.

Consider three coupling channels.

Thermal generation consumes water. Most U.S. thermal generation — coal, gas, nuclear — uses water for cooling, and a non-trivial share is consumptive. The U.S. Geological Survey’s water-use surveys have, for two decades, identified thermoelectric power as the single largest category of freshwater withdrawal in the country. When a data center procures grid power, it inherits a water footprint upstream of its meter that is often larger than its on-site cooling water use. The Lawrence Berkeley National Laboratory and academic literature have pegged the embodied water intensity of grid electricity at roughly one to two liters per kilowatt-hour in the U.S. average, varying enormously by generation mix.

This means a hyperscaler’s water disclosure that only includes on-site water is a fractional disclosure. Amazon’s 2.5 billion gallons is the visible portion of the iceberg. The water consumed at the gas plants and nuclear stations supplying its grid power is real, and is paid for — through power prices — by Amazon. It just does not show up on the sustainability report.

Direct-to-chip liquid cooling moves water around but does not abolish the need for it. The industry is in the middle of a forced transition, well-covered in this week’s feed. Vertiv closed its acquisition of ThermoKey to expand cooling tech. An opinion piece in Data Center Dynamics argued for direct-to-chip liquid cooling as the inevitable response to AI rack densities. The transition is genuine: 100+ kilowatt racks are no longer cooled effectively by air, and direct-to-chip cold plates carry heat off the silicon at temperatures and densities air cannot match.

But what happens to the heat? It is moved to a coolant distribution unit and then dumped to a facility loop — and from there, in most current designs, to an evaporative cooling tower. Direct-to-chip is a tech change on the chip side; on the facility side, the heat-rejection path is overwhelmingly still evaporative. Liquid cooling reduces the air-side parasitic load and improves PUE, but it does not, by itself, materially reduce water consumption. Dry-cooler heat rejection exists, but at a power penalty: more compressor work, higher condenser temperatures, lower IT capacity per megawatt of grid connection. In a power-constrained world, that penalty translates directly into stranded capex.

So the cooling architecture the industry is converging on — chip-level liquid loops feeding facility loops feeding evaporative rejection — is a compromise that preserves the water-for-power trade the industry has always made. It does not resolve it.

Renewable buildout shifts the embodied-water mix but does not eliminate it. Solar and wind are dramatically less water-intensive at the generator. But the grid services around them — thermal peakers, hydropower (which has its own evaporative losses from reservoirs), and increasingly battery storage manufactured with water-intensive processes — keep the embodied water in the system. A “solar-powered AI data center” headline, like the Jinko Power one this week, is meaningfully better than a coal-powered one on water. It is not water-free.

The accounting problem the industry has not yet faced

When Oracle announced this week that it would spend on the order of $70 billion in AI data center capex and Wall Street wobbled, the conversation immediately turned to depreciation cycles, GPU obsolescence, and customer concentration. It did not turn to water risk. That is an artifact of how the disclosure regime currently works, and it is one of the more interesting structural gaps in the industry’s risk surface.

Power risk is priced. Lenders ask about grid interconnection. PPAs are scrutinized. FERC queues are monitored. Power Purchase Agreements show up in 10-Ks. Tapestry’s deployment for PJM’s interconnection process, reported this week, is itself a market response to the fact that power risk has become so material it has its own AI tooling layer.

Water risk is not similarly priced, for three structural reasons.

It is local. A national water footprint is almost meaningless; what matters is the watershed-level balance, which is rarely measured at a granularity that maps cleanly to a project parcel. There is no equivalent of FERC for water — water rights are state-administered in the U.S., with eight or nine different doctrinal regimes, and the data is fragmented across utilities, irrigation districts, and state engineers’ offices. A national hyperscaler operating in twenty states is dealing with twenty water-rights regimes, and lenders typically do not require it to disclose the aggregate exposure.

It is cheap until it is forbidden. Municipal water rates in the U.S. typically run $3 to $10 per thousand gallons. At those rates, even 2.5 billion gallons annually costs Amazon something in the low tens of millions of dollars — a rounding error on a hundred-billion-dollar capex program. Water pricing does not transmit scarcity until the moment scarcity transmits politically and the permit is denied. The transition from “negligible cost” to “project killed” can happen in a single municipal vote, with no intermediate signal that a risk model could have priced.

Disclosure standards are immature. SEC climate disclosure has a CO2 framework. CDP Water has a reporting framework, but it is voluntary, and the granularity is poor. The result is that an investor reading a hyperscaler 10-K can build a defensible model of power exposure and almost no model of water exposure. The Amazon disclosure this week is, in this light, a small step toward closing the gap — but it is a single number, in a sustainability report, without watershed-level breakdown.

The structural implication is that water is currently mispriced as a risk factor at the project level and, by extension, mispriced across the capex stack. When that mispricing corrects — through permit denials, watershed-level moratoria, or eventually through forced disclosure — the correction will not be smooth. It will arrive as a step-function repricing of which sites are buildable and which are stranded.

Three scenarios for the next five years

Scenario analysis is more useful here than point forecasting, because the binding-constraint question genuinely has multiple plausible answers. Steelman each.

Scenario A: Water becomes a managed cost, not a constraint. In this world, the industry’s existing transitions — direct-to-chip liquid cooling, higher facility supply-water temperatures, increased use of reclaimed and non-potable water, geographic dispersion to wetter sites — are sufficient. Hyperscalers double down on locations like the Pacific Northwest, Nordic countries, and the U.S. Southeast where water is abundant and average WUE figures drop industry-wide. WUE becomes a standard disclosure metric, modeled by the same teams that model PUE, and the constraint never binds because the industry routes around it.

This is the steelman case for incumbent hyperscalers. It is supported by the fact that the cooling-tech vendors (Vertiv, ThermoKey, and others in this week’s feed) are genuinely shipping equipment that can support warmer facility loops and partial dry cooling. It is also supported by the falling marginal water intensity of the better-engineered new builds, which run materially below the industry average WUE.

The weakness in this scenario is that it assumes the binding constraint stays where it has been — in cost — rather than migrating to political legitimacy. The Maine moratorium was vetoed and a developer still walked. That suggests the constraint is already migrating.

Scenario B: Water becomes the binding political constraint, ahead of power. In this world, by 2028 the rate-limiting step on new AI capacity in North America and Europe is not megawatts but watershed permits. Communities that cannot easily oppose a substation can oppose a million-gallon-a-day withdrawal. The moratorium model that almost passed in Maine becomes a template that travels to other states. Hyperscalers respond by concentrating new builds in coastal regions with desalination, in genuinely water-rich jurisdictions, and in air-cooled designs that take a power and density penalty.

The implication for capex is significant. A meaningful fraction of land already optioned for data center development in the U.S. Southwest and Mountain West becomes unbuildable on water grounds. Capex is not destroyed but is geographically reshuffled, and the reshuffling raises costs through higher land prices in water-rich regions and grid-interconnection delays in regions not previously prioritized for capacity. This is the scenario where the Iron County, Utah approval that just cleared looks, in retrospect, like a late approval rather than an early one.

Scenario C: Cooling architecture is forced past its current local optimum. Both A and B assume the basic facility-side cooling stack — chip-level liquid, facility loop, evaporative tower — persists. In this scenario, water constraint forces the industry past that local optimum. Dry coolers become standard despite their power penalty. Two-phase immersion gets a second wind. And, at the speculative edge, the SpaceX orbital data center announcement this week — putting compute on Starlink craft, with planned 2027 launches — becomes less a stunt and more a marker that “compute somewhere else entirely” is on the menu. The orbital case has its own physics problem (radiative heat rejection from a vacuum is not magic), but the very fact that a publicly listed company is making it as part of its IPO narrative tells you the industry is beginning to entertain heterodox heat-rejection architectures because the conventional one is hitting walls.

The weakness in this scenario is timing. Dry cooling and immersion are mature enough to deploy; orbital is not. The question is whether the water constraint binds inside the timescale on which the alternatives can scale, or whether it binds first and forces premature commitments.

First principles: what is actually scarce

Strip the analysis back. What is the underlying driver?

A data center is, thermodynamically, a device for converting electrical work into low-grade heat in a small physical envelope. Every joule of electricity that enters the building must leave as heat. The only question is what working fluid carries that heat out of the envelope, at what temperature, and to where.

Evaporative cooling carries heat out of the building in water vapor and into the atmosphere. It is, in the physicist’s sense, the cheapest way to do this because water’s latent heat of vaporization is large and air is free. Mechanical chilling carries heat out via a refrigeration cycle, using electricity to lift low-grade heat to a temperature where it can be dumped to ambient. Dry cooling does the same without the phase change, paying a larger electrical penalty in exchange for not consuming water.

The industry’s current architecture is, fundamentally, a choice to spend water rather than power. That choice made sense when water was effectively free and power was the binding constraint. It stops making sense the moment water becomes contested — either through political opposition, through priced scarcity, or through embodied water in the power supply becoming a disclosed liability.

What is genuinely scarce, in first-principles terms, is heat-rejection capacity to the environment, at the location where the compute needs to sit, in a form the local community will tolerate. Power is one input to that capacity. Water is another. Land is another. Political legitimacy — the most underpriced input of all — is the fourth.

The buildout has been spending all four, in a mix optimized for an earlier era’s binding constraint. The mix is being forcibly rebalanced. The interesting analytical work over the next two years is not predicting which constraint binds first; it is mapping which sites, which architectures, and which firms are positioned for which mix.

Synthesis: what to actually watch

Several signals are worth tracking quarterly. None are headline-grade individually. Together they form the indicator set that will tell you whether water is migrating from sustainability-report footnote to binding constraint.

Watershed-level moratoria and their political travel. The Maine moratorium failed on a governor’s veto, but the legislature passed it. Watch how many other states introduce comparable bills in 2026 and 2027 legislative sessions, and watch which ones survive executive review. The political technology of a water-based moratorium is portable. It will travel.

The fraction of new hyperscaler capacity sited in water-rich regions. If Scenario A is winning, average new-build WUE will fall and capacity will visibly migrate to the Pacific Northwest, Nordics, and Southeast. If Scenario B is winning, capacity will continue to be announced in water-stressed regions and then quietly delayed, withdrawn, or rerouted. The Sentinel withdrawal in Maine and the Jinko announcement in western China are the two endpoints of this distribution to watch.

Disclosure granularity. If hyperscaler sustainability reports begin breaking out water consumption by watershed or by basin, that is the leading indicator that the underwriting community has begun pricing the risk. If they stay at the aggregate-gallon level, the constraint is still being managed as a PR problem rather than a financial one.

Dry-cooling and air-cooled retrofits in announced builds. The cooling vendors will publish the order books. A meaningful shift from evaporative to dry cooling in new orders — even at a power penalty — is the signal that the industry’s internal capex models have started repricing water risk above power risk.

The treatment of embodied water in power procurement. If hyperscalers begin specifying water-intensity ceilings in PPAs the way they currently specify carbon intensity, the embodied-water dimension has entered the procurement function. That would be the most consequential development, because it would close the loop between the power and water constraints and force the entire upstream generation stack to compete on water as well as carbon.

Capex disclosure. Oracle spooked investors this week with its $70 billion buildout commitment. If a future capex announcement of that scale draws investor questions specifically about water risk at named sites, the market has begun pricing the constraint. So far it has not.

The longer arc

The AI buildout is, in the long arc, a story about converting capital into compute through a chain of physical bottlenecks that are revealed one at a time. The chain has so far been: chips, then packaging (the CoWoS/HBM bottleneck that is still binding), then power, then grid interconnection. Each constraint, when it became binding, was at first dismissed by the industry as a near-term cost issue and then, belatedly, recognized as a structural one. Each forced an architectural response: more fabs, more advanced packaging capacity, gigawatt-scale PPAs, nuclear restarts, behind-the-meter generation.

Water is next in the chain. It will not announce itself with a single headline event. It will arrive as a creeping pattern: a permit denied here, a moratorium passed there, a sustainability report that finally publishes the absolute number, a developer that walks from a site without explaining why. The Amazon disclosure this week is the absolute-number signal. The Sentinel withdrawal is the developer-walking signal. They are the two halves of the same emerging picture.

The firms that recognize this and reprice their site portfolios — geographically, architecturally, and politically — will retain optionality. The firms that continue to underwrite on power and land alone will discover that their cheapest acres are also their least defensible. And the analysts who insist on treating water as a sustainability metric rather than a capex risk factor will be the ones surprised when the constraint binds.

The structural question is not whether water will become a binding constraint on the AI buildout. The disclosures and withdrawals this week suggest it already is, in specific watersheds, at specific developers’ specific sites. The structural question is how fast the industry’s accounting, underwriting, and architectural choices catch up to a fact that is already on the ground — and what gets stranded in the gap between the two.

Power gets the headlines. Water will get the projects.