AI Data Center Boom Breaks Water Safety in Drought Zones

America is building its AI infrastructure in the wrong places, and a new analysis makes the scale of that problem impossible to ignore. Of the 809 new data centers planned across the United States, 517, nearly two-thirds, are sited in areas currently experiencing drought, according to analysis published June 9, 2026. These are not marginal locations. They're in the counties where NOAA's National Integrated Drought Information System has classified land as drought-stricken for the past year, in states where 63% of the country's land area is already water-stressed. The AI industry is about to demand more water than it has ever demanded before, and it has chosen to do it where water is already running out.

What Actually Happened

Analysis published June 9, 2026 by multiple infrastructure research outlets, covered by Tom's Hardware and Environmental Link Network, mapped the geographic distribution of all 809 planned US data center projects against the current drought monitoring data from NOAA's National Integrated Drought Information System. The finding: 517 of those projects are sited in areas classified as drought-stricken over the past twelve months, a figure that translates to 63.9% of all planned US AI infrastructure. The geographic concentration reflects a familiar pattern, low land costs, favorable zoning, and existing power grid access in the Sunbelt, Southwest, and inland Southeast have made those regions default targets for data center development, even as those same regions face the country's most severe water stress.

The scale of water demand that this siting pattern implies is substantial. According to Barchart, AI data centers collectively consumed 264 billion gallons of water in 2025, equivalent to the annual water usage of approximately 1.8 million Americans. On a daily basis, that translates to roughly 550 million gallons per day, placing AI infrastructure's water footprint in the same category as major industrial sectors that have operated under strict water use permitting for decades. The demand is driven primarily by cooling requirements: modern GPU compute racks, particularly Nvidia's Hopper and upcoming Rubin-generation systems, generate thermal loads that require continuous water-based cooling to operate within safe temperature ranges. A single Rubin GPU cluster can require hundreds of kilowatts of cooling capacity running continuously.

The distribution of water demand across the AI supply chain is counterintuitive. Cooling for the data centers themselves accounts for only approximately 4% of the additional water AI will require by 2050, according to lifecycle analysis cited in EnkiAI's grid strain analysis. Power generation, specifically the thermoelectric generation that cools steam turbines at natural gas and nuclear plants powering data centers, accounts for roughly 54% of AI's total water footprint. Semiconductor fabrication adds approximately 42%. The upshot is that a data center's direct cooling water use is the visible part of an iceberg; the full water cost of an AI workload runs through the power grid and chip supply chain in ways that current siting decisions don't account for.

Why This Matters More Than People Think

The drought zone siting pattern is not primarily an environmental story, it's an infrastructure risk story with financial consequences for every company building or depending on AI compute in the affected regions. Water use permits in drought-stressed areas are becoming contested. Arizona, which hosts roughly 15% of total US data center capacity and is among the most drought-stressed states in the country, has already moved to raise data center power bills by 45% to protect residential grid access. Water-use conflicts follow a similar political logic: when a drought forces a county to choose between residential water supply and data center cooling, the political outcome is predictable. Several Western states have enacted or are considering legislation that would cap data center water use or require drought contingency planning as a condition of permitting.

The energy dimension compounds the water problem in ways the industry is only beginning to quantify. US data center electricity demand reached 42 gigawatts in 2026, up from 23 GW in 2023, a growth rate that outpaces the grid additions being built to serve it. PJM Interconnection, which manages the grid for over 65 million people across 13 states, now projects a 6 GW shortfall against reliability requirements by 2027. Capacity prices in PJM have risen 833% between the 2024-25 and 2025-26 delivery years as the gap between power demand and available supply tightens. Data centers sited in drought regions often depend on thermoelectric generation that is itself water-constrained, power plants that need water for cooling cannot run at full capacity during drought conditions, creating a feedback loop where the infrastructure stress that harms data center water supplies also limits the power generation that the data centers depend on.

However, skeptics argue the industry is responding faster than the drought-zone framing suggests. Microsoft CEO Satya Nadella recently highlighted the company's new closed-loop cooling systems that reduce annual data center water consumption to volumes comparable to a restaurant, an improvement of up to 80% over the millions of gallons that legacy open-loop cooling systems require. Google has committed to achieving water-positive operations across its data center fleet by 2030, which would mean returning more water to local watersheds than the facilities consume. The bear case for these commitments is that they apply to new builds and retrofits, not the existing installed base, and that the 809 planned projects in the new analysis include facilities from operators who have made no comparable commitments.

The Competitive Landscape

The data center siting problem is creating differentiated risk across the major hyperscalers in ways that are not yet priced into their AI infrastructure narratives. Microsoft's Azure has been most aggressive in diversifying data center geography toward Nordic countries with abundant hydroelectric power and cold ambient air that reduces cooling water requirements. Google's investment in clean energy procurement has focused on solar and wind in regions with lower water stress. Amazon Web Services has historically favored Virginia and the Pacific Northwest, though its expansion into drought-prone Texas and Arizona has followed similar economic logic to the industry-wide pattern the analysis documents. Meta and Oracle, both of whom signed the White House Ratepayer Protection Pledge in March 2026, have committed to directly funding grid infrastructure improvements, a commitment that addresses the electricity side of the problem without directly addressing water stress.

The Chinese AI infrastructure buildout offers a comparative reference point. China announced a $295 billion plan for domestic AI infrastructure in June 2026, with facilities planned across a range of climate zones including water-rich regions in the south and northwest. The geographic distribution is mandated by industrial policy rather than purely by land cost and zoning, which produces a different risk profile than the US market-driven siting pattern. Whether central planning produces better infrastructure siting outcomes than market signals is a contested empirical question, but the US pattern revealed by the drought zone analysis suggests that market signals, low land cost, existing grid access, favorable permitting, are systematically pointing data center development toward water-stressed locations without incorporating the full cost of that water stress into investment decisions.

The historical parallel is the nuclear power industry's early facility siting decisions, which concentrated plants near rivers and coastlines for cooling access without accounting for long-term climate trends that would affect cooling water availability decades later. Several nuclear plants built in the 1970s now face periodic shutdowns during summer heat waves when river temperatures rise above cooling permit thresholds. The AI data center industry is making analogous siting decisions at dramatically faster timescales, with infrastructure that will still be operational in 2045 when climate projections show the current drought zones experiencing conditions substantially more extreme than today.

Hidden Insight: The Water Price Signal That Isn't There Yet

The fundamental market failure driving the drought zone siting pattern is the absence of a water price signal that reflects actual scarcity. In the regions where most of the 809 planned data centers are located, water is priced as a utility service at rates that bear little relationship to its scarcity value. A data center that consumes millions of gallons per month in a drought-stressed Arizona county pays the same per-gallon rate as a residential customer, with no premium for drawing down an aquifer that is not replenishing. The infrastructure investment decision treats water as a fixed-cost input rather than a constrained resource with a shadow price that rises as the aquifer depletes. When the constraint eventually bites, through permit caps, rate increases, or physical supply limitations, the data centers already built in those locations face stranded asset risk that their financial models don't currently contemplate.

The semiconductor fabrication component of AI's water footprint adds a supply chain dimension that the data center siting debate largely ignores. TSMC's Arizona fabs, the facilities producing the AI chips that run in the data centers being built in drought zones, are themselves major water users in one of the most water-stressed regions of the United States. Fab-level water consumption for advanced nodes runs to millions of gallons per day; the ultra-pure water required for chip washing and chemical process control cannot be recycled at high rates without degrading purity below the specifications that advanced lithography requires. The combination of fab water demand and data center water demand in the same drought-stressed geography creates geographic concentration of AI supply chain vulnerability that has no equivalent in prior technology cycles.

What the drought zone data reveals, if read through an infrastructure investment rather than environmental lens, is an industry-wide repricing event in slow motion. Water use permitting, grid capacity reservation, and environmental impact requirements are all moving toward more restrictive regimes in the affected states, driven by both water stress and residential grid protection politics. The companies building data centers in drought zones are effectively front-running a permit regime that doesn't yet reflect full scarcity pricing, counting on the fact that existing facilities tend to receive grandfather status when new restrictions are enacted. That logic worked for data centers built in the 2010s. It may not work for 517 new projects announced in 2026, given how rapidly permitting sentiment has shifted in Western and Sunbelt states.

The most underappreciated pressure point is the insurance market. Catastrophic drought risk is a known and growing peril in the Sunbelt and Southwest, and commercial property insurers have begun incorporating drought probability into underwriting models for long-duration infrastructure. A data center with a 20-year operational life, sited in a region projected to face D3-D4 extreme drought conditions for 30-40% of years in that period, represents a materially different insurance risk than a facility built in a climate-stable region. As insurance pricing catches up to climate risk models, a process already underway in coastal flood zones, the economics of drought-zone data center siting will deteriorate faster than the current investment wave anticipates.

What to Watch Next

The 30-day signal is state-level permit activity in Arizona, Nevada, and Texas, the three states that host the largest share of planned drought-zone data center projects. Arizona's legislature is in late-stage deliberation on a commercial water use cap for facilities above 1 million gallons per month, a threshold that would affect most hyperscale data center projects. If that legislation passes before the next legislative recess, it would effectively freeze permitting for several of the 517 drought-zone projects in the analysis and force developers to negotiate mitigation plans that could add months to timelines and tens of millions of dollars to project costs.

At 90 days, the signal is whether the Biden-era EPA water use guidance, currently under review by the Trump administration's EPA, is modified in ways that ease or tighten commercial water use requirements for large industrial facilities. The Trump administration has generally favored infrastructure permitting acceleration, which would be expected to ease water permit requirements for data centers. However, the political dynamics in drought-stressed Sunbelt states that are part of the Republican coalition complicate that calculus: residential voters in Phoenix and Las Vegas are attuned to water availability in ways that create intraparty pressure toward conservation requirements regardless of the federal administration's general posture on industrial permitting.

The 180-day market signal is the emergence, or absence, of a drought risk premium in data center REIT pricing. Publicly traded data center REITs including Equinix, Digital Realty, and Iron Mountain have geographic concentration disclosures that allow investors to calculate the proportion of their assets in drought-stressed regions. If institutional investors begin pricing drought risk into REIT valuations the same way they have begun pricing coastal flood risk into office real estate portfolios, the capital market signal would reach data center developers through the cost of capital rather than through regulation. That repricing, once it begins, tends to accelerate faster than the physical infrastructure constraints it reflects.

The AI industry is building its future on land that is running out of water, at a speed that leaves no time to notice before the permits stop coming.

---