The Water Nobody Counts

Arizona’s Salt River Valley has hosted data centers since the 1990s. The desert climate appealed to operators — stable geology, minimal flood risk, historically cheap power, and a state government friendly to technology investment. The region now hosts a concentration of hyperscale data center campuses that rivals Northern Virginia and the Bay Area. It also sits in one of the most water-stressed metropolitan areas in North America, drawing on the Colorado River at a moment when the river’s flow is at its lowest recorded levels since measurement began in the late 19th century.

The average hyperscale data center in the Phoenix metro area uses between 3 and 5 million gallons of water per day for cooling. The region hosts dozens of such facilities. This is not a marginal or speculative concern — it is a documented, quantifiable consumption pattern that sits in direct conflict with the municipal water supply obligations of a region that has been negotiating water allocation cuts with six other states and the federal government for the better part of a decade.

The AI boom accelerated the data center buildout in exactly these kinds of regions. Land is cheap where water is scarce. Power purchase agreements are often favorable in deregulated energy markets in the American Southwest. And until very recently, water consumption was not meaningfully included in the regulatory calculus for data center permitting.

The global numbers are what you’d expect once you start looking. Data center water consumption worldwide reached approximately 1.9 trillion liters in 2028, according to estimates compiled from utility water use reports, facility disclosures under the EU’s Data Center Environmental Regulation, and hydrological monitoring data. This is water withdrawn from rivers, aquifers, and municipal water systems — most of it for evaporative cooling, which means most of it is consumed rather than returned to the watershed.

For context: 1.9 trillion liters is roughly equivalent to the annual household water consumption of the entire city of Los Angeles, approximately 3.9 million people. The global data center sector is consuming water at urban-city scale, invisibly, in service of compute.

The water footprint of AI specifically is harder to disaggregate, because data centers serve many workloads beyond AI. But based on the share of data center energy attributable to AI (roughly 35% at major hyperscalers in 2028, a figure that comes from facility-level energy audits where available) and the rough proportionality of water and energy consumption in evaporative cooling systems, AI-related water consumption was probably in the range of 600 to 800 billion liters in 2028. More water than several mid-sized countries.

Two cooling technologies dominate at scale: air cooling and evaporative (wet) cooling. Air cooling is simpler and uses essentially no water — it blows ambient air through the facility. Evaporative cooling is substantially more efficient at rejecting heat in hot climates, but it consumes water by design. The water absorbs heat, evaporates, and is gone.

Immersion cooling — submerging server hardware in a thermally conductive liquid — has been growing in deployment since 2025, particularly for high-density AI accelerator racks that generate heat concentrations that air cooling cannot handle. Immersion cooling can reduce or eliminate water consumption entirely, but it requires different facility design, different hardware handling procedures, and higher upfront capital costs. It is not yet dominant.

The industry’s water efficiency metric is Power Usage Effectiveness for water, or WUE — liters of water per kWh of IT load. The best facilities achieve WUE around 0.2–0.5 L/kWh, using various heat recapture and dry-cooling strategies. The worst are above 2.0 L/kWh. The average across the industry sits around 1.0–1.2 L/kWh. A 100 MW data center at average efficiency consumes about 876 million liters of water per year. These numbers make the aggregate figures less surprising once you work through the arithmetic.

The geography of data center water use is where the policy failure becomes most visible. Water-stressed regions — defined by the World Resources Institute as areas where more than 40% of available freshwater is withdrawn annually — host a disproportionate share of global data center capacity. This is partly historical accident (data centers were sited for cheap land and power without water stress being a primary consideration) and partly active economic optimization (the same factors that make land cheap in dry places — low density, limited competing economic activity — also make it attractive for large industrial facilities).

The concentration is stark. The American Southwest, Northern Africa’s coastal zones, parts of the Middle East where data center investment has surged under sovereign wealth fund initiatives, and portions of South Asia all have large data center concentrations in areas of extreme water stress. Chile, which has become a significant data center hub for South American connectivity, hosts major hyperscaler facilities in regions where Indigenous communities are simultaneously fighting copper and lithium mining operations for water rights.

This is not primarily a technology story. It’s a land use and resource rights story, where the technology investment cycle has moved faster than regulatory frameworks designed for a different era.

What the AI companies say about this is illuminating. Most major hyperscalers now include water stewardship sections in their annual sustainability reports. Microsoft has committed to being “water positive” by 2030 — returning more water to watersheds than it consumes. Google has made similar commitments. These are presented as meaningful environmental commitments.

The methodology for “water positive” claims involves purchasing water rights that are then retired or redirected to environmental flows, funding watershed restoration projects, and investing in local water infrastructure. These are real activities. The question is whether they are equivalent to simply not withdrawing water in the first place — and whether local watershed restoration in Oregon offsets water consumption in Phoenix in any meaningful sense for Phoenix’s water system.

Locational materiality matters enormously for water in ways it doesn’t for carbon. A ton of CO₂ emitted anywhere affects the global atmosphere. A billion liters of water consumed in Phoenix’s watershed does not draw down the water table in Oregon, regardless of what water restoration projects are funded there. The portfolio approach to water stewardship that parallels the renewable energy certificate approach to carbon is vulnerable to the same fundamental criticism: it counts credits that don’t match the geography of the actual impact.

Some data center operators are making genuine progress on this. Facilities designed from the ground up in the last three years often use air cooling or liquid-to-liquid systems that require no evaporative water loss. Several operators have moved capacity expansion away from water-stressed regions explicitly because of water risk — both environmental and reputational. The Dutch government began regulating new data center water consumption as part of planning approval in 2026, and the effect on facility design has been visible.

But the existing installed base is not going anywhere. The data centers built in Phoenix and Las Vegas and the Mojave in 2024 and 2025 and 2026 will be running for twenty years, consuming water that the Colorado River does not reliably have, in quantities that the region’s municipal water planners are not adequately accounting for in their projections.

The water crisis in AI infrastructure is less visible than the energy crisis because water doesn’t have the same set of associated statistics and IEA reports. It needs them. The AI industry’s environmental conversation has been disproportionately structured around the language of carbon and electricity, where metrics and accountability mechanisms exist, while the water dimension remains in a pre-regulatory state where voluntary commitments and portfolio accounting dominate.

Six years of AI energy data finally lets us have an honest conversation about electricity. The equivalent conversation about water is about three years behind. The aquifers are not waiting.