US Data Centres’ Water Demand Is Becoming a Local Infrastructure Test

US data centres directly consumed an estimated 17.4 billion gallons of water in 2023, but the national total says little about whether a particular town can support a new campus. For local utilities, the critical issues are the source, timing and quality of a facility’s water demand, alongside the cost of expanding supply and wastewater infrastructure. AI is intensifying those pressures, but cloud growth and colocation are also reshaping the sector’s water needs.

That distinction gets lost in arguments about AI infrastructure. National totals provide context; local water availability is the constraint that decides whether a project works. A facility can be a rounding error in US water consumption and still be a very large new customer for a small utility.

What does the 17-billion-gallon estimate actually measure?

First, a correction: this isn’t a measurement made by the Congressional Research Service (CRS). The CRS’s July 2026 water FAQ cites a Lawrence Berkeley National Laboratory (LBNL) model. LBNL estimated that US data centers directly consumed 66 billion litres in 2023—around 17.4 billion US gallons. Its definition is on-site water consumption: water removed from its immediate source through evaporation or another irreversible process, chiefly in thermal-regulation systems.

That differs from a withdrawal. A cooling system may withdraw or receive water, recirculate it several times, discharge some mineral-rich blowdown (water removed to control mineral build-up), and evaporate the rest. For a community concerned about water available to other users, consumption is usually the sharper measure: water that doesn’t promptly return to the original local system in usable form.

The figure is a modelled national estimate, not the sum of utility invoices from every server room in America. LBNL combines equipment-shipment data, estimates of data-center locations and types, cooling-system market data, climate simulations and assumptions about operating practice. The approach is broader than relying solely on a handful of hyperscaler sustainability reports, but it remains an estimate, in part because facility-level information isn’t routinely public.

The trend is clear. LBNL puts direct consumption at 21.2 billion litres in 2014, rising to 66 billion litres by 2023. What matters most is where that demand sits: hyperscale sites—very large data centers—and colocation sites, where operators house customers’ equipment, accounted for 84% of the modelled 2023 direct total. Those are the large facilities most likely to become a planning issue for an individual locality.

LBNL’s 2024 US Data Center Energy Usage Report is the primary source for these estimates. The CRS data-centers-and-water briefing is useful for its account of the policy and data gaps, without suggesting they have already been solved.

Is the indirect water footprint more important than cooling water?

At national level, it may be. LBNL estimated that producing the electricity used by US data centers in 2023 consumed nearly 800 billion litres of water—about 211 billion gallons—based on the regional electricity mix serving their locations. That’s roughly twelve times the direct on-site estimate.

That isn’t a reason to ignore cooling-water demand. It is a reason not to congratulate a site for a low water-usage effectiveness (WUE) score without asking what happened to its electricity use. WUE measures on-site water consumed per kilowatt-hour used by IT equipment. It is a useful operating metric, but it doesn’t capture water used at power stations.

The trade-off is real. Dry coolers and air-cooled chillers can drastically reduce water use at the facility boundary. In hot conditions, they can require more electricity than an evaporative system. A water-cooled chiller may offer better energy performance and capacity at large scale, but cooling towers deliberately evaporate water to reject heat. Shifting the burden from the water bill to the grid isn’t a complete environmental improvement, especially where the power plants increasing output to meet additional demand have a significant water footprint.

The answer changes with the grid, climate and time of day. LBNL’s indirect estimate uses local balancing-authority grid mixes because facility-level electricity contracts and behind-the-meter generation—generation serving a site without passing through the utility meter—aren’t transparent enough to model nationally. That’s a sensible limitation to state plainly. A corporate renewable-energy claim doesn’t automatically show the water impact of the electricity physically meeting a site’s load in a particular hour.

Is AI the reason data centers are becoming thirstier?

AI is a major reason, but “AI did it” is too convenient. GPU-heavy training and inference clusters, using graphics processing units (GPUs), pack far more power into racks than conventional enterprise servers. More heat in less floor space changes the cooling problem and is accelerating adoption of direct-to-chip liquid cooling and other high-density designs.

However, LBNL attributes the rise in data-center energy use through 2023 to rapid AI-server growth and continued growth in conventional server demand. Its model also treats hyperscale and colocation expansion as central structural changes. Cryptocurrency mining is a separate, specialised form of compute infrastructure with its own opaque energy profile; it shouldn’t be casually folded into a water total for conventional data centers.

The more defensible conclusion is that AI has made the next build cycle larger, denser and harder to cool, while cloud services and colocation had already concentrated computing in very large facilities. The latest LBNL 2025 update raises the electricity outlook further, estimating that data centers could account for 11.8% of US electricity use by 2030 in its central scenario. Water forecasts are less certain because cooling choices, climate and the future generation mix matter so much.

Does liquid cooling solve the water problem?

No. It first solves a different problem: moving heat away from increasingly dense processors efficiently. A direct liquid-cooling loop transfers heat from server components into a recirculating coolant loop rather than dumping it into room air first. That can reduce fan energy, permit warmer supply temperatures and improve heat transfer.

But the heat still has to leave the building. If the final heat-rejection stage is a cooling tower, the site can still consume substantial water. If it uses dry heat rejection, on-site water consumption can be low, with a possible electricity and capital-cost penalty. Immersion cooling, rear-door heat exchangers and direct-to-chip cold plates should be assessed by their actual thermal chain, not the appealing word “liquid”.

There is no universal best system. Smaller or older facilities may find a wholesale cooling redesign uneconomic. In a cool climate, air-side economisation—using suitably cool outside air—can avoid mechanical chilling for long periods. In a hot, dry area, evaporation can be exceptionally energy-efficient but politically awkward when it relies on drinking water. In a hot, humid area, its usefulness changes again.

The US Department of Energy’s guidance makes the point well: raising temperature setpoints within equipment limits, avoiding needlessly tight humidity control, improving cooling-tower chemistry and increasing cycles of concentration—allowing water to circulate longer before discharge—can cut water use. Yet reverse-osmosis treatment, a membrane-filtration process, and some other recovery measures add energy use, maintenance and cost. The engineering question is which constraint matters most at a particular site—not which headline metric looks best.

Can reclaimed water make growth acceptable?

It can make a meaningful difference, especially where cooling would otherwise use potable supply. It isn’t magic water. Reclaimed or recycled water still needs treatment appropriate to its use, storage and pipework, reliable supply arrangements, and a plan for concentrated brine or blowdown. It can also shift costs onto a utility that must build treatment and distribution capacity before the data center arrives.

The Environmental Protection Agency describes an example in Quincy, Washington, where Microsoft and the city developed a closed-loop reuse system that treats cooling water and returns it to the data center, reducing pressure on potable groundwater. The scheme still needs make-up water because evaporation is unavoidable, and it took about a decade to plan, build and test. That is an important qualification for reuse claims: it requires infrastructure rather than simply a change in reporting.

The Environmental Protection Agency is now explicitly supporting recycled-water use for industrial and data-center cooling, including work on regulatory barriers. That is worthwhile. But communities should insist on a clear distinction between potable water avoided, water recycled on site and water ultimately consumed through evaporation. They are related measures, not interchangeable ones.

Why is a modest national share still a local problem?

CRS notes that direct data-center use is estimated to be around 2% of total US water consumption. That statistic is often used to end the discussion. It shouldn’t be. Water is local: an aquifer, river basin, municipal treatment works and distribution network don’t become less constrained because national agriculture, thermoelectric generation or industry uses more water overall.

Demand also peaks when supply is under greatest stress. Sites using evaporative cooling need more water in the hottest hours and months. A local utility needs to assess peak daily and seasonal demand, not simply divide an annual allocation by 365. It should also examine drought restrictions, source resilience, pipeline and treatment upgrades, wastewater capacity, and whether other users face higher rates or reduced headroom.

Most US data centers are believed to receive water from municipal systems rather than draw it directly themselves. That can make their use difficult for outsiders to identify: a utility may have the data, but it can be aggregated with other customers or covered by commercially sensitive service agreements. Self-supplied facilities may be no clearer, depending on state rules.

Would federal reporting rules fix the accountability gap?

They would improve it, but they wouldn’t decide whether a particular project fits a particular watershed. Water rights, water-service agreements and land-use decisions remain primarily matters for states, local utilities and planning bodies.

There is still a clear federal information problem. CRS says the federal government has not conducted a systematic assessment of data-center water use. USGS water data has historically not separated data centers as a category, while facilities on public supply can disappear into broad commercial or industrial figures. Researchers are left modelling a fast-changing industry instead of checking a consistent public record.

Several bills introduced in the 119th Congress would require some form of data-center water and energy disclosure. One proposal, the Data Center Water and Energy Transparency Act of 2026, would require covered operators to report monthly water use and source, average PUE and WUE, and forward projections for new or expanded sites. As of August 2026, these are proposals, not settled rules.

The sensible policy baseline isn’t exotic: require facility-level, monthly reporting above a meaningful demand threshold; distinguish withdrawals from consumption; identify potable and non-potable sources; publish seasonal peaks and expansion forecasts; and disclose who funds water and wastewater upgrades. Add location-aware electricity and water metrics where practical. That wouldn’t settle every development dispute, but it would replace a great deal of heat with evidence.

The US data-center water debate isn’t really about whether 17 billion gallons sounds large or small. It is about whether infrastructure planning has caught up with the physical consequences of placing ever-denser computing loads in particular places. At present, the construction pipeline is moving faster than the public data. That’s the imbalance worth fixing.