AI Data Centers May Use Far More Water Than Cooling Figures Suggest

Google News surfaced a stark conflict: AI data centers in Texas and New Mexico may carry a water footprint roughly 10 times larger than corporate cooling figures suggest.

That comparison does not necessarily mean operators hid ten gallons for every gallon reported. It usually reflects a broader accounting boundary. Corporate disclosures often emphasize water consumed directly at a facility, while independent estimates also count water consumed while generating its electricity.

That distinction matters in two states attracting enormous computing projects despite recurring drought, stressed aquifers, and incomplete facility-level reporting. It also pits Big Tech’s low-water cooling promises against the full resource demands of an electricity-intensive AI buildout.

The available evidence does not support treating every proposed data center as equally thirsty. Cooling designs, climates, power sources, operating loads, and local water systems vary widely. A closed-loop facility can consume little water on-site while still creating substantial demand at gas, nuclear, or hydroelectric power facilities.

The real story is therefore not a single universal multiplier. It is an accounting problem with local consequences. Residents and regulators cannot judge those consequences when companies disclose only selected parts of the system.

What the 10-Times Water Claim Actually Measures

The apparent gap comes from counting the water behind electricity, not simply from finding unreported pipes inside data centers.

A data center has two main water footprints. Direct consumption occurs at the facility, primarily through cooling. Indirect consumption occurs when power plants generate the electricity needed by servers, cooling equipment, pumps, and supporting infrastructure.

Water withdrawal and water consumption are also different measurements. Withdrawal covers water taken from a river, reservoir, or aquifer, including water later returned. Consumption covers the portion removed from the immediate local supply, often through evaporation.

Those distinctions can produce dramatically different totals. The national energy report from Lawrence Berkeley National Laboratory estimated that US data centers directly consumed about 17.4 billion gallons during 2023.

The report associated their electricity use with another 211 billion gallons of indirect consumption. That second figure is about 12 times the direct estimate.

It offers a credible basis for a headline describing a footprint more than 10 times larger than commonly highlighted operating figures. However, it is a national model, not a meter reading for one Texas or New Mexico campus.

The estimate also includes water consumption associated with hydroelectric reservoirs. Some analysts argue that attributing reservoir evaporation to individual electricity users can inflate comparisons with methods that exclude hydropower.

An independent policy analysis from the Information Technology and Innovation Foundation accepted the importance of indirect consumption but questioned parts of the national methodology. It noted that electricity-related water demand changes sharply with the regional generation mix.

That debate does not make indirect water disappear. It shows why a single nationwide multiplier cannot describe every facility accurately.

A server campus supplied mostly by wind and solar power will have a different indirect footprint from one dependent on water-cooled thermal generation. The same cooling design can therefore create different total impacts in different markets.

Texas illustrates this problem clearly. Its grid draws from wind, solar, natural gas, coal, and nuclear generation. Each source has a different water profile, and the mix changes by location and time.

New Mexico’s proposed facilities bring another complication. A developer may promise a closed-loop system with limited routine refilling, yet the campus can require new generation that consumes water elsewhere.

The Google News headline is strongest when read as a warning about incomplete boundaries. It becomes weaker if interpreted as proof that every company has secretly consumed exactly 10 times its disclosed amount.

The most defensible conclusion is narrower. Direct cooling disclosures do not describe the complete water footprint of AI computing, especially when electricity demand is excluded.

Google News Exposes a Texas Reporting Blind Spot

Texas is planning around an industry whose water demand remains poorly documented at the facility level.

The state’s data problem is no longer theoretical. Texas officials surveyed data centers and cryptocurrency mining facilities about their electricity and water requirements. Fewer than one-third of the 341 facilities contacted supplied information.

The Public Utility Commission received responses from 28 companies covering 92 facilities in different development stages. That count included both data centers and cryptocurrency operations, so it does not provide a clean AI-only measurement.

The response to a separate Texas Water Development Board survey was even lower. Only 17% of the state’s data centers answered its 2025 annual water-use survey, according to legislative hearing coverage.

That survey is required by state law. Nonparticipation can be punished as a Class C misdemeanor, yet officials still lacked answers from most operators.

The number of facilities contacted by the water board rose from 22 in 2023 to 341 in 2025. Planning systems built for a smaller industry are now chasing a rapidly expanding infrastructure sector.

Texas has surveyed data centers since 2020. At that time, the board contacted one facility and received one response. The subsequent decline has left planners estimating missing consumption from past use and other available records.

A Texas Water Development Board briefing disclosed another limitation. The most recent facility information available when planners developed the 2026 regional plans came from only one or two facilities in 2020.

That does not mean every gallon used by a data center vanished from state accounting. Water purchased from a municipal utility can appear within broader commercial demand. Power-generation requirements can also appear elsewhere in regional projections.

However, those categories do not reveal which campuses created the demand, which cooling systems they use, or how their requirements change during extreme heat. They also cannot show whether projected growth matches the actual development pipeline.

The timing is especially difficult. AI servers concentrate more computing power within each rack, producing heat that must be removed continuously. Operators are simultaneously proposing campuses whose electrical loads can rival those of cities.

Berkeley Lab found that US data centers consumed 176 terawatt-hours of electricity in 2023. It projected demand between 325 and 580 terawatt-hours by 2028, depending on growth and efficiency.

Texas must plan water and electricity infrastructure years before those upper scenarios become visible in utility bills. Weak reporting transfers the risk of forecasting errors from developers to public agencies and local communities.

This is where the Google News framing becomes useful. The argument is not merely that Big Tech uses water. It is that public planning cannot reconcile corporate claims with local impacts when disclosures use inconsistent boundaries.

Several operators did appear before Texas lawmakers, including Amazon, Google, and Vantage Data Centers. That participation provides more accountability than silence, but testimony is not a standardized facility-level dataset.

Lawmakers from both parties have called for clearer disclosure. Republican state representative Cody Harris said resource transparency should not be optional and identified water sources as a necessary reporting field.

The missing information also prevents fair comparisons. A low-water operator cannot demonstrate its advantage when regulators lack uniform measurements. A high-water facility can disappear inside municipal or industrial totals.

Better reporting would separate direct withdrawal, direct consumption, recycled water, potable water, and estimated electricity-related consumption. It would also identify peak demand during hot or dry conditions.

Without those figures, the debate will continue to rely on competing models and selective examples. Neither corporate averages nor worst-case national multipliers can replace local measurements.

Closed-Loop Cooling Solves Only Part of the Problem

Low-water cooling can reduce direct demand, but it often shifts the tradeoff toward electricity use and off-site water consumption.

Traditional evaporative cooling removes heat by evaporating water. This method can be energy efficient, particularly in suitable climates, but the evaporated water is consumed rather than returned immediately to the watershed.

Dry cooling releases heat through air and radiators. It can sharply reduce direct water consumption, although fans and related equipment generally require more electricity.

Direct-to-chip liquid cooling sends a liquid coolant near high-temperature processors. The coolant can circulate inside a closed loop for extended periods, but another system must ultimately release the collected heat.

These designs explain why developers can truthfully describe a campus as using very little operational water on-site. They do not establish that the complete system has a negligible water footprint.

Oracle offered a concrete example at the Stargate campus in Abilene, Texas. The company described an initial municipal fill for a closed-loop cooling system, followed by limited annual additions for each building.

According to Stargate project reporting, the initial fill was one million gallons. Oracle expected each of eight buildings to require another 12,000 gallons annually.

Those numbers suggest a major reduction from continuous evaporative cooling. They remain company projections, and operating data will be needed to verify performance under full AI workloads and Texas heat.

The project also illustrates the boundary problem. A closed-loop cooling system can avoid ongoing municipal demand while its electricity suppliers consume water during power generation.

The final result depends on how much electricity the campus uses, which plants respond to that load, and how those plants obtain cooling water. Annual renewable-energy purchases do not always reveal the generation serving a facility each hour.

Google has also emphasized location-specific cooling decisions. The company says it plans to prioritize air cooling in water-stressed watersheds and seek reclaimed wastewater where appropriate.

Its broader target is to replenish more freshwater than its offices and data centers consume by 2030. Google reported that stewardship projects replenished about 7.7 billion gallons in 2025, equal to roughly 78% of its freshwater consumption.

The company’s water project portfolio included 165 projects at the end of 2025. These programs can improve wetlands, irrigation systems, aquifer recharge, and municipal infrastructure.

Replenishment still requires careful interpretation. Restoring water in one watershed does not erase a shortage beside a facility in another. The timing and reliability of each project also matter.

A company-wide target can therefore coexist with local pressure. Both statements may be accurate because they answer different questions.

Direct water efficiency asks how much freshwater the facility consumes per unit of computing or electricity. Replenishment asks how much water a company claims to restore through projects. Local risk asks whether a particular watershed can support demand when it occurs.

Those metrics should complement each other, not substitute for one another. Publishing only the most favorable metric leaves communities unable to evaluate the tradeoff.

Closed-loop systems also face operational uncertainties. Equipment performance can change during exceptional heat, maintenance, equipment failures, or expansion. Backup cooling arrangements may use different amounts of water.

Developers should disclose both routine and peak requirements. Municipal systems must be sized for difficult days, not merely annual averages.

That obligation becomes more important as AI chips grow hotter and rack density rises. A design that performs well for conventional cloud servers may face different thermal demands from tightly packed accelerators.

None of this invalidates low-water engineering. Closed-loop and dry designs are among the clearest tools available for reducing direct consumption.

The mistake is presenting them as complete proof of a water-free AI facility. They solve one part of a connected water and energy system.

New Mexico Turns the Accounting Debate Into a Local Test

In New Mexico, small routine cooling claims must be weighed against desert hydrology, power generation, and the scale of proposed campuses.

Project Jupiter in Doña Ana County has become a focal point. The proposed AI infrastructure is associated with Oracle and OpenAI’s broader Stargate buildout near Santa Teresa.

Developers have described its water requirements as limited because the cooling system would circulate liquid rather than continuously evaporate freshwater. Reports have cited an initial fill measured in millions of gallons.

An initial fill and routine annual consumption describe only two parts of the project. Construction, electricity generation, backup systems, landscaping, dust control, and future expansion can add other demands.

The water source matters just as much as the volume. New Mexico communities depend on a mixture of groundwater, surface water, and legally constrained river systems. A modest withdrawal can carry greater consequences in a stressed basin.

Local officials must also consider when water is consumed. Demand during an abundant season has a different impact from the same demand during drought or extreme heat.

The economic case deserves equal scrutiny. Data centers can create construction work, tax revenue, infrastructure investment, and demand for local services. Their permanent staffing levels may remain modest compared with their land, electricity, and capital requirements.

That imbalance does not automatically make a project undesirable. It means public incentives and resource commitments should be compared with verified, durable local benefits.

The wider industry often frames dry cooling as the answer for desert development. It can be part of the answer, but dry cooling usually requires more electricity than evaporation-based heat rejection.

If added generation relies on water-consuming thermal plants, some of the saved cooling water moves to another site. If the project instead adds wind, solar, storage, or other low-water generation, the indirect footprint can fall significantly.

Transmission constraints complicate that choice. A corporate agreement for renewable energy does not necessarily mean the local grid can deliver carbon-free, low-water electricity during every operating hour.

Developers can address this uncertainty through hourly energy data, generation contracts, and transparent estimates of marginal power demand. Marginal demand refers to the generators that increase output when the facility consumes another unit of electricity.

Average grid figures are easier to publish but less useful for causal analysis. A campus operating continuously may depend on different resources at night than during a sunny afternoon.

New Mexico should therefore avoid reducing its decision to whether a cooling loop needs frequent refilling. The relevant question is whether the complete project fits within regional water and energy limits.

That review should include the full planned buildout, not only its first phase. Large campuses often expand through separate buildings and permits, which can obscure cumulative demand.

Communities also need enforceable operating limits. A developer’s current cooling plan can change after permits, ownership transfers, or equipment upgrades.

Permits can establish maximum direct withdrawals, approved sources, drought procedures, and reporting intervals. Utilities can require separate meters rather than burying demand inside broad commercial categories.

Facility-level publication would allow residents to compare promised and actual performance. It would also protect developers from exaggerated claims when measured consumption remains low.

The debate has already moved beyond engineering. Political backlash against data centers has spread through rural areas, including traditionally conservative parts of Texas.

Water, electricity rates, land use, noise, and property rights now occupy the same political argument. A project that cannot answer basic resource questions risks becoming a symbol for every unresolved concern.

New Mexico’s test is therefore institutional as much as technical. Officials must determine whether existing permitting and disclosure systems can evaluate AI infrastructure at unprecedented scale.

Big Tech’s Water-Positive Promises Need Local Receipts

Corporate replenishment goals provide useful direction, but they cannot replace comparable site-level operating data.

Google says water use should not be a black box. That principle sets a reasonable standard for Google and every other major infrastructure operator.

The company publishes freshwater consumption for many data center locations. However, disclosures across the industry remain inconsistent, making company-to-company comparisons difficult.

Some reports cover owned facilities but exclude leased colocation space. Others combine offices, warehouses, and data centers. Companies can also define recycled water, replenishment, withdrawal, and consumption differently.

Indirect water receives even less consistent treatment. Corporate sustainability reports often discuss electricity emissions without attaching a water figure to power generation.

This omission can make two facilities appear similar when their actual footprints differ greatly. One may use an evaporative cooling tower supplied by reclaimed water. Another may use dry cooling but depend on water-intensive electricity.

Neither direct consumption nor indirect consumption alone provides the whole answer. Regulators need both, alongside the source and local scarcity context.

Corporate averages can also conceal outliers. Google’s 2025 data showed that its stewardship projects replenished roughly 78% of total freshwater consumption. That figure does not show whether replenishment occurred beside the facilities generating the greatest demand.

The same problem applies to a 120% replenishment target. Exceeding company-wide consumption is valuable only when the projects deliver measurable water benefits. It does not grant every facility unlimited local demand.

A credible facility report should identify annual and peak withdrawal, actual consumption, potable supply, reclaimed supply, and cooling technology. It should show planned capacity and the percentage operating.

The report should separately estimate electricity-related consumption using a disclosed method. Because that estimate depends on contested assumptions, operators should present a range rather than false precision.

Companies could also publish the power sources included in that calculation. Excluding hydropower, including reservoir evaporation, or using annual renewable contracts can materially alter the result.

Independent assurance would make these disclosures more credible. Auditors should verify meter data, calculation methods, and whether replenishment projects produced benefits during the reporting year.

Data also needs a consistent unit. Water usage effectiveness, or WUE, measures water consumed per kilowatt-hour of computing equipment energy. It is helpful for efficiency comparisons but does not replace total consumption.

A campus can improve WUE while increasing its total water demand because it adds more servers. Both intensity and absolute volume must remain visible.

The same principle applies to electricity efficiency. A more efficient accelerator can still raise total energy use when companies deploy it at far greater scale.

Big Tech’s strongest response is therefore not a national advertising campaign. It is a set of local receipts that lets communities connect engineering promises with measured outcomes.

This approach also benefits enterprise AI buyers. Companies purchasing cloud capacity increasingly face their own environmental reporting obligations and customer scrutiny.

Buyers cannot evaluate an AI service’s footprint when providers offer incompatible data. Standardized facility and regional information would support better procurement decisions.

Knowledge workers following the story through Google News should also distinguish company claims from measured results. A planned cooling design is evidence of intent, not evidence of full-load performance.

The transparency gap invites two opposite errors. Industry advocates may describe nearly water-free campuses while ignoring electricity. Critics may apply a national 12-times multiplier to every facility regardless of its grid.

Local, standardized data is the path between those extremes.

Three Signals Will Show Whether the Water Gap Is Closing

The next stage of this story will be decided by mandatory disclosure, measured cooling performance, and the power sources serving new AI loads.

The first signal is whether Texas converts legislative frustration into enforceable reporting rules. Voluntary surveys have not produced a dependable dataset, even when participation was legally required.

A meaningful rule would cover facilities above a clear electricity or water threshold. It would require direct withdrawal, direct consumption, water sources, cooling technology, peak demand, and operating capacity.

The rule should also distinguish existing campuses from proposed expansions. Otherwise, officials may continue planning with historical figures while developers add much larger AI buildings.

If Texas adopts standardized, public facility reporting, the claim that the industry’s footprint is unknowable will weaken. If another planning cycle passes with broad categories and sparse responses, skepticism will strengthen.

The second signal is measured performance from closed-loop campuses. Oracle’s Abilene figures and similar developer forecasts create testable commitments.

Annual disclosures should show whether routine refilling remains near projected levels after facilities reach full utilization. They should also identify water used during unusual heat, maintenance, or emergency operation.

If actual consumption stays low across several hot summers, developers will have strong evidence that direct water demand can be controlled. If usage rises materially with load, the engineering narrative will require revision.

New Mexico should apply the same test to Project Jupiter. The public needs operating data for each phase, not only a single initial-fill estimate announced before full operation.

The third signal is the generation mix behind new data center demand. Berkeley Lab’s work shows why electricity-related consumption can exceed direct cooling demand by more than tenfold nationally.

That relationship will not remain fixed. More wind, solar, and storage can reduce electricity’s water intensity, while additional thermal generation can preserve or increase indirect demand.

The scale of future electricity use raises the stakes. BloombergNEF projected 118 gigawatts of installed US data center capacity by 2030 and 194 gigawatts by 2035.

Those forecasts remain uncertain because many announced projects will face grid, financing, equipment, and permitting constraints. Still, even partial development would reshape regional electricity planning.

Watch what utilities actually build, not only what technology companies contract. New gas generation, nuclear expansion, renewable capacity, storage, and transmission will produce different water outcomes.

Hourly matching will become an important indicator. Annual clean-energy contracts can offset consumption on paper while the facility relies on water-consuming generation during other hours.

These three signals can confirm or weaken the central judgment behind the Google News story. Better disclosure would narrow uncertainty. Verified closed-loop performance would reduce direct demand. Low-water electricity would shrink the hidden portion.

The opposite pattern would deepen the conflict. Missing reports, higher-than-promised cooling demand, and water-intensive generation would show that corporate figures still describe only part of the system.

Readers should resist both easy conclusions while those signals develop. AI data centers are not universally capable of draining a region, and low-water cooling does not make them resource-free.

The practical question is whether each watershed, utility, and community can support a specific project under realistic operating conditions. That answer requires measurements rather than slogans.

Follow the next Texas reporting rules, the first full-load results from desert campuses, and the generation built to serve them. Those records will reveal whether the 10-times gap shrinks through engineering or survives through accounting.