The aquifer wasn’t in the pro forma when it came to some data center deals
Water permitting is behaving like a site veto. The deals underwritten in 2024 and 2025 haven’t priced it that way.
The aquifer wasn’t in the pro forma. That’s the category error sitting inside several hundred data centre construction deals underwritten in 2024 and 2025. The unpriced risk isn’t interest rate exposure. It isn’t hyperscaler renewal risk. It’s water: the absence of secured water permits on the critical path at financial close.
The market has treated water as an operating cost — a line item for facilities managers to optimise after the land deal closes and the EPC contract is signed. But water is increasingly behaving like a permitting prerequisite. It can kill a project before a shovel goes in the ground, the same way a grid interconnection denial can. Over 40 per cent of planned and existing US data centres are in areas classified as high or extremely high water scarcity. Two-thirds of new hyperscale campuses built since 2022 are in high water-stress counties, according to Bloomberg News. Developers chose those sites for cheap power and permissive land use.
The consequence is financial, not primarily environmental. A lender who has underwritten a 36-month construction schedule against a Texas site without secured water permits has financed a permitting risk that isn’t in the deal price. A developer who closes land and signs an EPC contract before community water opposition materialises has absorbed a delay risk that the hyperscaler anchor tenant — who retains the option to relocate or renegotiate — doesn’t share. The disclosure vacuum that makes this possible is structural: Texas has no mandatory water use reporting for data centres, Virginia’s governor vetoed a bill that would have required it, and utilities are often bound by NDAs from releasing consumption data.
Risk accumulates invisibly until it surfaces as a permit denial, a community veto or a project delay. At that point, the balance sheet that can move has already moved. What follows is not an argument that the AI buildout stalls on water — at the national level it won’t. It’s an argument that a specific class of assets, financed against specific assumptions about permitting timelines, is carrying water risk that isn’t priced in. Understanding which deals those are — and who holds the residual when the assumption breaks — is the question worth taking into the next diligence call.
Water as a site veto: what the permitting map actually shows
The operating assumption in most data centre site selection has been that power and land are the hard constraints. Water is cheap — often less than US$5 per thousand gallons — and treated as abundant. That assumption is getting stress-tested in real time.
Over 40 per cent of planned and existing US data centres are in areas classified as high or extremely high water scarcity, according to analyses cited by Nixon Peabody. The regulatory consequence is starting to materialise. Arizona has effectively closed new groundwater certificates to hyperscale projects. Virginia has introduced no-net-increase water clauses for new permits. Google’s Lancaster, Ohio facility won approval only after committing to 100 per cent non-potable or reclaimed water. Microsoft’s Mount Pleasant, Wisconsin campus faced a multi-year delay triggered by aquifer studies. These aren’t outliers. They’re the leading edge of a permitting regime shift that is two or three legislative cycles from becoming standard practice in most major data centre markets.
The water rights framework makes this worse, not better. In the western US, prior appropriation law means late arrivals — and hyperscale data centres are almost always late arrivals — face the real possibility of being last in line during drought conditions. In the east, riparian rights offer more flexibility but less certainty in high-stress counties. A developer building a 500MW campus in a Texas county where the aquifer is already over-allocated isn’t acquiring a water right. They’re acquiring a litigation risk. Permitting timelines for water access are now, in some markets, rivalling or exceeding electrical interconnection queues, according to legal advisors tracking this space. That’s a structurally important fact that isn’t showing up in most development pro formas.
The disclosure problem compounds everything. There is no uniform method for reporting or tracking data centre water usage at the state or national level. Texas sent surveys to 70 data centres in 2024 asking how much water they used; many didn’t respond. Utilities are sometimes bound by non-disclosure agreements from releasing consumption data. Virginia’s governor vetoed a bill that would have required water use reporting. California and New Jersey passed mandatory reporting requirements through their legislatures, then watched their Democratic governors veto them. The result: state water authorities can’t forecast demand, communities can’t assess impact, and lenders can’t price the risk — because the data doesn’t exist in any systematic form.
The Memphis case: what happens when the goodwill runs out
xAI’s Colossus facility in Memphis is the clearest available case study for what water risk looks like at full scale — not as a regulatory abstraction but as a live permitting, community and financial dispute.
Colossus came together in 122 days in 2024. xAI bought the former Electrolux building in south Memphis, a predominantly Black neighbourhood already carrying a disproportionate environmental burden, and stood up what it described as the world’s largest AI supercomputer, running initially on temporary methane gas turbines that bypassed standard Clean Air Act review. Peak water demand for the facility is expected at 5 million gallons per day. The Memphis Sand Aquifer — already being drawn down faster than it’s replenished — was the fallback source. Community groups, including Protect Our Aquifer and Memphis Community Against Pollution, were not consulted before the deal was announced.
The water goodwill proposition xAI offered was an US$80 million Colossus Water Recycling Plant, designed to pull 13 million gallons per day of treated effluent from the adjacent T.E. Maxson Wastewater Treatment Facility. The plant broke ground in October 2025, received Clean Air Act permits in January 2026, and obtained new construction permits in March. Then, in April, construction stopped. As of this writing, xAI has given no public explanation. Based on public reporting, there does not appear to be a binding public obligation requiring completion. The city’s ability to compel it appears limited absent a formal development agreement. The community has no remedy except continued litigation. E&E News reported this week that critics have speculated the halt is tied to cost discipline ahead of xAI’s upcoming SpaceX IPO — a causal link xAI has not confirmed.
The operational expansion is continuing regardless. Colossus 2, in south Memphis, began operations in January 2026. A third facility is planned in Mississippi. Both are too far from the recycling plant to use its output, and both will draw millions of gallons per day from drinking water infrastructure. What started as a single site with a credible water mitigation plan has become a three-site footprint with a stalled mitigation commitment and an expanding potable water draw.
The Texas problem: the state can’t see what it can’t measure
Texas is the highest-stakes version of this story. The state has more than 400 data centre facilities, with 70 more planned, and is now set to surpass Virginia in total new data centre construction. The Texas Water Development Board already identifies a 4.8 million acre-foot annual shortage under existing demand. Data centres are not in that model. Texas does not require operators to disclose projected water use or report actual consumption.
HARC’s January 2026 white paper puts numbers on the gap. Texas data centres already consume an estimated 25 billion gallons annually when both direct cooling and indirect power-generation use are included. HARC’s range for 2030 — 29 to 161 billion gallons per year — is not analytical sloppiness. It’s the data blindness made visible: the underlying consumption figures don’t exist in any systematic form, and HARC’s central conclusion is that Texas’s water planning framework is structurally blind to an industry growing faster than the five-year planning cycle can track. At the top of the range, that’s 2.7 per cent of the state’s total water use from a single sector that currently has no reporting obligation.
The Abilene disclosure shows how hard it is to reconcile nameplate gigawatts with actual water intensity. The Stargate campus — 1.2GW anchored by OpenAI, Oracle and SoftBank — is using a closed-loop water system that city officials describe as drawing 8 million gallons to fill the initial cooling system, against Abilene’s average daily use of around 22 million gallons. But the same official could not say how much or how often additional make-up water, maintenance drains and replenishment would be required thereafter. This isn’t a cover-up story. It’s a data availability story: even the best-disclosed campus in the country can’t yet fully reconcile engineering intensity, initial fill volumes and ongoing operational draw — because the industry hasn’t been asked to report these numbers in any standardised way.
The water disclosure gap isn’t just a public policy problem. It’s a capital allocation problem. Lenders underwriting construction debt against future revenue streams from these assets can’t model a risk they can’t see. The Texas GOP adopted a December 2025 resolution calling for the data centre industry to follow the same water management and recycling protocols required of oil and gas — a politically revealing comparison. ExxonMobil moved its Permian Basin operations from 64 per cent recycled water in 2022 to 87 per cent in 2024, under regulatory pressure and market discipline. The data centre industry, absent comparable requirements, has no equivalent forcing function.
Technology is real, but it’s solving yesterday’s build, not today’s
The strongest bull case goes like this: the industry is already solving it. Direct-to-chip cooling and immersion cooling systems can achieve Water Usage Effectiveness ratios below 0.2 litres per kilowatt-hour, against an industry average of 1.8 to 1.9. Google uses non-potable or reclaimed water wherever local stress conditions indicate scarcity. Meta pledged water-positive operations in Arizona by 2030. Amazon Web Services made a similar pledge for its entire network in 2022 and has delivered on reclaimed water commitments at several sites. And Nvidia’s Blackwell GPU architecture, which runs hotter than its predecessors, is accelerating the shift to liquid cooling simply because air cooling can’t handle the rack densities. On this view, the technology transition solves the problem before the regulatory pressure bites hard enough to matter.
The problem with this argument isn’t the technology. It’s the vintage of the assets it doesn’t cover. Liquid cooling adds an estimated 15 to 20 per cent to upfront capex — a range cited by developers and EPC contractors in public commentary — and it doesn’t pencil for projects underwritten at speed in 2024 and 2025. Many of those assets appear to retain evaporative-cooling or hybrid-cooling water profiles, and will sit in high water-stress counties on municipal water draws for their entire 20-year operating life, through regulatory cycles the bull case assumes will be benign.
The geography problem doesn’t respond to technology commitments either. A water-positive pledge in Maricopa County, Arizona — where Phoenix draws 40 per cent of its water from a Colorado River system that has dropped from roughly 90 per cent capacity in 2000 to around 30 per cent today — is a voluntary commitment made against a structurally declining watershed. The basin trajectory is set by agriculture, population growth and climate. The basin math doesn’t care what Meta pledged in Mesa.
The standard industry rebuttal — agriculture uses 70 per cent of all freshwater withdrawals nationally, so data centres at 0.3 per cent of public supply are a rounding error — is true at the national level and irrelevant at the local level. The Meta facility in Newton County, Georgia uses 500,000 gallons per day: 10 per cent of the entire county’s water budget. The 25 data centres in Prince William County, Virginia average 18,000 gallons per day each. Local systems don’t have national averages as a relief valve. They have what’s in the aquifer, what’s in the pipe and what the community will tolerate. Those are the three constraints that generate permit denials, and none of them scale with the national adoption curve.
The developer holds the stranded position
The risk allocation structure here follows the same pattern as the power infrastructure buildout and the Entergy/Meta Hyperion deal: the parties with the strongest balance sheets retain optionality, and the weaker hands absorb the residual risk the hyperscalers have declined to own.
Hyperscalers — Meta, Microsoft, Google, Amazon — control site selection and campus architecture. They can build to liquid-cooled specifications, commit to reclaimed water and absorb the incremental opex. When community opposition creates a project delay, they have the balance sheet to absorb it. When a permit gets rejected in Arizona, they can shift capital to Iowa or Ohio. The optionality travels with the credit.
The parties that can’t move are the municipal water authorities, the communities and, increasingly, the project-level lenders and developers who have underwritten construction schedules against sites that didn’t model water as a permitting constraint. A developer who closes on land and signs an EPC contract in a Texas county before water permits are secured has taken on a permitting risk that isn’t in the capital markets price. A lender who has underwritten a 36-month construction schedule against that deal has financed that risk without knowing it.
The disclosure vacuum is the mechanism that makes this structural. In a world where water permits were required before financial close, or where state-level reporting requirements gave lenders visibility into aggregate aquifer drawdown, the risk would be visible and priceable. In the current environment — Texas has no mandatory disclosure, Virginia governors are vetoing reporting bills, utilities are bound by NDAs — the risk accumulates invisibly until it surfaces as a community veto, a permit denial or a project delay. At that point, the hyperscaler with the option to move has already moved. The lender and the developer hold the stranded position.
The sequencing question for capital
The Crusoe/Abilene deal shows what it looks like when water is in the model from the start. The 1.2GW Stargate campus selected Abilene in part because the city controls its own water and wastewater infrastructure and had the capacity and political will to accommodate a closed-loop system. City officials put numbers on the record before the deal closed. Those numbers may not yet be sufficient to fully reconcile initial fill, make-up water and operational draw — but the city at least produced a disclosure. That’s the diligence standard. Most of the deals getting done right now don’t meet it.
Mandatory water-use reporting looks increasingly likely across major US data centre markets, though timing will vary by state. California, New Jersey, Minnesota and Virginia are all either in active deliberation or have already passed bills that governors chose to veto for now. The next governor may not. When mandatory disclosure arrives, the gap between what was permitted and what was drawn will reprice assets in both directions. Facilities built to reclaimed-water or closed-loop specs will carry a compliance premium. Facilities fast-tracked in 2024 and 2025 on evaporative cooling in water-stressed counties will not.
Before getting to that question, there’s a more basic one: does the deal model include water at all? The eight variables I’d want on the table before underwriting any data centre construction debt in a water-stressed market are: confirmed water source and permit status at close; distinction between initial fill volume and ongoing make-up water requirement; wastewater reuse or recycled water obligation and its legal enforceability; drought curtailment priority relative to municipal and agricultural users under applicable water rights law; municipal system capacity headroom above the facility’s projected peak draw; community challenge history and any active litigation; and whether any water-related disclosure covenant exists in the offtake or lease agreement. Most current deal packages address one or two of these. Treating the rest as post-close operating detail is how a permitting risk gets underwritten as a construction schedule.
My read on this: the deals most exposed aren’t the ones with the worst water profiles — those will get caught in diligence eventually. The most exposed are the ones where water wasn’t modelled at all, priced as though a municipal hookup and a facilities manager were all the coverage needed. That’s the gap between where most 2024–2025 deal packages are and where the permitting map is heading.
The xAI water recycling plant sitting half-built in south Memphis is what that gap looks like when cost pressure arrives and the legal obligation doesn’t.