Almost every public argument about data centers and water is about cooling towers. It is the image in the coverage, the number in the objections, and the thing companies answer when they announce a closed-loop system.

Cooling accounts for roughly a quarter of the measured footprint.

The International Energy Agency estimates that data centers consumed about 560 billion liters of water in 2023. Of that, roughly two-thirds, some 373 billion liters, was indirect: water consumed generating the electricity that powers the facility. About 140 billion liters, a quarter, was direct cooling. Around 8 percent, 47 billion liters, went into hardware manufacturing. The IEA projects the total could reach roughly 1.2 trillion liters by 2030, a figure reported by BNP Paribas Markets 360 in its analysis of water stress as a sector risk.

A facility can therefore run an efficient closed-loop system, report low direct withdrawals, and still rest on a substantial water footprint located somewhere else entirely, in the watershed of whatever power plant serves it. The visible share is the smallest one.

Four different things called "water use"

Precision here is not pedantry. The categories carry different consequences and different remedies.

Withdrawal is water taken from a source. Consumption is the portion not returned, typically lost to evaporation. Discharge is what returns, and its temperature, chemistry, and treatment status determine whether return means restoration. Indirect use is the water embedded in electricity supply, the largest share and the one least visible to any local planning process.

These are routinely collapsed into one figure. A facility may have low consumption but water-intensive power. Another may use reclaimed water but discharge heated effluent. A third may be sustainable today in a watershed whose future allocation nobody has modeled. One number cannot distinguish these, and a community given one number cannot evaluate what it is being asked to host.

Water does not scale the way electricity does

Electricity can be generated in many places and moved across a grid. Water is bound by rainfall, aquifers, watershed capacity, seasonal variation, and prior claims from households, agriculture, and ecosystems. Its availability is local in a way that power is not.

Timing compounds the constraint. Cooling demand peaks in the hottest months, precisely when supplies are most stressed.

Technical responses are real and insufficient alone. Direct-to-chip and immersion cooling reduce water intensity. On-site recycling and reclaimed water reduce competition with drinking supplies. None answers whether a particular watershed can absorb a particular facility, which is a planning question rather than an engineering one.

Where the official record complicates the story

The most authoritative public assessment available comes from Virginia's Joint Legislative Audit and Review Commission, the legislature's own audit body, and it does not support the loudest version of the water alarm.

JLARC found that current data-center water use in Virginia is sustainable, and that most facilities use about as much water as a large office building. That finding deserves to be taken at face value. It is not evidence that data centers are draining Virginia's water supply, because they are not.

What JLARC also found is where the argument actually sits. The commission noted that Virginia has limited oversight of how water is allocated among competing local uses. Sustainable in aggregate today is not the same as governed for tomorrow. A statewide finding of current sufficiency establishes nothing about whether individual watersheds can absorb concentrated growth, or about who decides between a data center and an agricultural user when they compete.

JLARC was blunter on land. It concluded that the industrial scale of these facilities makes them largely incompatible with residential uses, and that inadequate local planning had allowed roughly a third of Virginia's data centers to be sited near homes.

The issue is rarely any single facility. Pressures emerge cumulatively. A watershed that can comfortably support one development may be stressed by a cluster of developments approved through separate planning processes, each individually reasonable on its own terms. The governance challenge is therefore not only evaluating individual projects but understanding the aggregate demand they create together.

What local approvals cannot see

Two established bodies of scholarship explain why a planning system can function correctly at every individual decision and still produce an outcome no one selected.