In 2024, Google’s data centers consumed 8.1 billion gallons of water — nearly double their consumption three years earlier. A single Google facility in Council Bluffs, Iowa drank 1 billion gallons in 2024 alone, more than any other site the company operates. Across Loudoun County, Virginia — the densest data center cluster in the world — facilities used 1.6 billion gallons in 2023, roughly 10% of all water consumed in the county, after growing 250% in just four years. Direct US data center water use sits around 17 billion gallons annually today; the Lawrence Berkeley National Laboratory’s 2024 federal report projects that figure will reach 38 to 73 billion gallons by 2028, driven primarily by AI training and inference workloads.
The single most important fact for drinking water consumers: roughly 57% of data center direct water use comes from potable water supplies — the same lakes, rivers, and aquifers that supply municipal drinking water. In a growing number of US metropolitan areas, AI infrastructure is now buying treated drinking water and evaporating it for cooling, while the residents next door face escalating drought restrictions, rate increases, and source-water stress.
At a Glance: The Numbers That Matter
| Metric | Value | Source / Year |
|---|---|---|
| US data center direct water use | ~17 billion gallons | 2023 (LBNL) |
| US data center projected use by 2028 | 38–73 billion gallons | LBNL 2024 federal report |
| Hyperscale-only projected use by 2028 | Up to 33 billion gallons | Industry forecast |
| North American total (2025) | ~1 trillion liters (~264 billion gallons) | Mordor Intelligence |
| Avg ~100 MW hyperscale facility | ~528,000 gallons/day | Industry average |
| Largest single facilities | Up to 5 million gallons/day | Equivalent to towns of 10K–50K |
| Google global data centers (2024) | 8.1 billion gallons | Google 2025 Environmental Report |
| Google Council Bluffs, IA (2024) | 1 billion gallons (single site) | Google 2025 Environmental Report |
| Loudoun County, VA (2023) | 1.6 billion gallons (~10% of county) | Loudoun Water |
| Maricopa County, AZ (2025) | 905 million gallons (0.12% of county) | County data |
| Share from potable supply | ~57% direct use | Industry analysis |
| Share from “blue water” (lakes/rivers/aquifers) | 80–90% | Industry assessment |
| 30 minutes of generative AI use | ~0.16 gallons of water | Global Water Intelligence, Jan 2026 |
How Data Centers Actually Use Water
A modern hyperscale data center generates 30 to 100 megawatts of waste heat that has to be removed continuously to keep servers within their operating temperature range. Two cooling architectures dominate, and they have radically different water profiles.
| Cooling Method | Water Use | Power Use | Where It’s Used |
|---|---|---|---|
| Evaporative (cooling towers) | High — 1–4 gal/min per MW of cooling | Lower | Hot/dry climates, large hyperscale |
| Adiabatic / direct-to-chip liquid | Moderate, recirculating | Moderate | Newer AI-optimized builds |
| Air-cooled chillers | Low direct, higher indirect | High | Cooler climates, older facilities |
| Immersion cooling | Very low | Lower | Specialty / experimental |
| Free cooling (outside air) | Near-zero | Very low | Pacific NW, Nordic countries |
Evaporative cooling — the dominant approach for new hyperscale builds in warm and dry climates — uses cooling towers that evaporate water to remove heat. Each megawatt of cooling load typically evaporates 1 to 4 gallons of water per minute. A 30-megawatt facility running evaporative cooling year-round can directly consume 300 million to 1 billion gallons annually — the household water use of a town of 5,000 to 15,000 people. The water that returns to the watershed returns as atmospheric vapor; the local utility loses it from its supply portfolio.
Air-cooled facilities consume less water directly but require substantially more electricity, which carries an indirect water footprint at the power plant. Thermoelectric power generation is itself the largest industrial water user in the US — natural gas combined-cycle plants, coal plants, and nuclear plants all withdraw and consume substantial volumes for steam cycles and condenser cooling. Shifting from evaporative to air-cooled is not a clean win on water; it is a partial trade-off depending on the local generation mix.
Direct-to-chip liquid cooling is becoming standard for the highest-density AI server racks. Liquid cooling moves more heat per unit volume than air, and high-end systems recirculate coolant in closed loops, reducing on-site water use. The catch: the heat still has to go somewhere, and most large facilities reject it through evaporative cooling towers anyway. Liquid cooling reduces server-side water use without necessarily reducing facility-side water use.
The Semiconductor Layer
The picture is incomplete without the chip fabs that supply the GPUs running inside data centers. Semiconductor manufacturing is more water-intensive per unit area than data centers — and the same regions building data centers are building fabs.
| Facility Type | Water Use | Notes |
|---|---|---|
| Average chip fab | ~10 million gal/day of ultrapure water | TSMC, Intel, Samsung, Micron typical |
| Single fab high end | 20–38 million liters/day (5–10 million gallons) | Rivals daily use of a small city |
| Ultrapure water yield | ~1,500 gal feedstock → 1,000 gal UPW | Concentrate stream wastes 1/3 of input |
| TSMC Arizona (Phoenix) | Multiple fabs phased through 2030 | New construction |
| Intel Ohio (One) | Up to 5M gal/day at full buildout | Active construction |
| Samsung Texas (Taylor) | Multiple billion-dollar phases | Full ramp by late 2020s |
| Micron New York (Clay) | Largest US fab project announced | $100B over 20 years |
Producing 1,000 gallons of ultrapure water requires approximately 1,500 gallons of municipal feedstock, because the purification process itself rejects a substantial concentrate stream. The CHIPS Act-funded fab construction wave commits the US to substantially higher manufacturing-side water demand through the 2030s. Drinking water utilities serving these regions are now negotiating supply commitments to industrial users that span 30+ years.
Where the Pressure Is Concentrated
| Region | Documented Water Use | Pressure Point |
|---|---|---|
| Loudoun County, VA | 1.6B gal (2023); 250% growth since 2019 | Largest US data center cluster (“Data Center Alley”); ~200 operational facilities |
| Maricopa County, AZ (Phoenix metro) | 905M gal (2025) | Active Colorado River shortage; Tier 1 cuts ongoing |
| Mesa, AZ | Google permitted up to 4M gal/day | Industrial = ~6% of city potable in 2024 |
| The Dalles, OR | Google water use disputed in court | Columbia River water rights litigation |
| Council Bluffs, IA | Google: 1B gal at one site (2024) | Single-facility US record holder |
| Atlanta metro, GA | Multiple new hyperscale builds | Lake Lanier source; recurring drought |
| Central Texas (Austin/San Antonio) | Multiple new builds | Edwards Aquifer drought stress |
| Eastern Oregon / Washington | Significant cluster | Columbia River basin |
| Newton County, GA | Large new Google complex | Rural water utility expanding to serve |
| Chesapeake Bay region | Growing cluster | Bay restoration goals stress with industrial growth |
The siting pattern is not coincidental. Data centers are sited for cool climates (which reduce cooling load), cheap electricity, fiber connectivity, and tax incentives. Several of the cheapest-electricity, most-incentive-friendly regions in the US are also among the most water-stressed — particularly the Southwest and the southern Great Plains. Local economic-development agencies often offer water as a deal sweetener during siting negotiations, with terms that are typically not subject to public review.
Loudoun County: The Stress Test Case
Loudoun County, Virginia hosts roughly 200 operational data centers — more than any other county in the world. Loudoun Water reported data center potable water use grew more than 250% between 2019 and 2023, reaching 1.6 billion gallons in 2023 and approaching 10% of all county water consumption. Loudoun Water’s expansion plans now include reclaimed-water infrastructure designed specifically to serve data center cooling demand without further drawing from potable supply, but as of 2026, most facilities continue to use treated drinking water. Each new generation of large language models from frontier AI labs has expanded water demand at existing facilities; the underlying infrastructure was sized for traditional cloud workloads, not AI training.
Phoenix Metro: Drought-Plus-Growth
Maricopa County hosts a rapidly growing cluster against the backdrop of an active Colorado River shortage declaration. Three Phoenix-metro cities have moved to constrain industrial water demand:
- Mesa, Avondale, and the city of Phoenix have passed ordinances capping industrial facility water use and requiring developers to purchase supplemental water supplies if they want to exceed those caps.
- Chandler passed an ordinance back in 2015 restricting new water-intensive businesses unless they aligned with the city’s economic development plan, with data centers specifically capped at 115 gallons per day per 1,000 square feet. The ordinance was an early national example of a municipality refusing to subsidize water for low-employment industrial uses.
Maricopa County data centers consumed about 905 million gallons in 2025 — roughly 0.12% of total county water use. The number sounds small, but it understates two structural problems. First, growth is steep: project pipelines for 2026–2030 imply order-of-magnitude increases. Second, water under drought-period prioritization is not fungible with golf course irrigation (which uses 3.8% of county water). Industrial-customer reservation provisions typically protect supply during cuts.
The Dalles, Oregon: When Disclosure Meets Court
Google operates a major data center cluster in The Dalles, Oregon, drawing from Columbia River basin water. The city of The Dalles initially refused to disclose Google’s water use to The Oregonian newspaper on the grounds that the data was a trade secret. After multi-year litigation, the data was disclosed in 2022 and showed Google’s use had grown substantially. The case prompted Oregon’s 2024 legislation requiring data centers above a defined size threshold to disclose water use publicly — a transparency reform whose absence in most states is itself instructive.
Operator Disclosure Scorecard
| Operator | 2024 Disclosure | Site-Level Data | Notable Programs |
|---|---|---|---|
| Google (Alphabet) | 8.1B gal globally; per-facility data for owned sites | Yes — including Council Bluffs (1B gal) | >25% of campuses use reclaimed/non-potable water |
| Microsoft | Total reported, not by site | No | Committed “water positive by 2030” target |
| Amazon (AWS) | Reports water-use efficiency (0.15 L/kWh, 5× industry avg); no total | No | Claims 53% progress to “water positive”; 20+ sites cooled with purified wastewater |
| Meta | Limited public detail | Partial | Reports global aggregate |
| Equinix | Reports aggregate | No | Sustainability goals published |
| Digital Realty | Reports aggregate | No | Sustainability reporting |
In April 2026, more than a dozen institutional investors filed shareholder pressure on Amazon, Microsoft, and Alphabet asking the companies to disclose site-specific water and power consumption ahead of annual investor meetings. The campaign is part of a broader push to put water-use data on the same disclosure footing as carbon emissions. As of mid-2026, only Google publishes site-level water data among the three largest cloud providers.
Water Usage Effectiveness — and Why It Matters
The data center industry has converged on Water Usage Effectiveness (WUE) as the standard efficiency metric — the ratio of water consumed (in liters) to IT energy delivered (in kWh). Lower is better.
| WUE Range | Interpretation | Example |
|---|---|---|
| <0.2 L/kWh | Best-in-class, often air-cooled or reclaimed water | AWS reports 0.15 (2024) |
| 0.2–0.6 L/kWh | Modern, well-designed evaporative facility | Many newer hyperscale builds |
| 0.6–1.5 L/kWh | Industry average | Mixed older portfolio |
| >1.5 L/kWh | Older, hot-climate, evaporative-heavy | Older hyperscale and colocation |
Industry average WUE in 2024 was approximately 0.5–0.7 L/kWh, with substantial spread across portfolios. WUE is a useful internal metric but has limits as a public-policy measure: it normalizes water use against IT load, so a more energy-intensive facility can post a “good” WUE while still consuming more total water than a smaller facility with worse WUE. For drinking water utility planning, total absolute water demand matters more than the efficiency ratio.
What This Means for Your Tap
The connection to residential drinking water runs through four concrete mechanisms.
1. Drought-period prioritization. Many municipal utilities operate under drought response plans that impose mandatory or voluntary residential restrictions during shortages. Industrial water-supply contracts negotiated with data centers and fabs typically include reservation provisions that protect industrial supply ahead of residential restrictions. The result, in practice: residents face lawn-watering bans and tiered rate increases while the data center down the road continues evaporating cooling water at full demand. Phoenix, San Antonio, and several Atlanta-metro utilities have faced public scrutiny over precisely this trade-off in the past two years.
2. Aquifer drawdown and source-water salinization. When industrial demand draws down aquifers that supply both industrial and residential users, the result over time is rising chloride and sodium concentrations as deeper, saltier groundwater enters the active production zone. This is the same physical mechanism driving saltwater intrusion in coastal aquifers, applied inland to drought-stressed inland systems. Higher source-water salinity raises treatment costs and can shift utility blending strategies in ways that affect downstream contaminants like disinfection byproducts.
3. Rate impacts. Residential customers ultimately fund the infrastructure that makes additional industrial demand possible. New treatment capacity, source-water diversification, and conveyance infrastructure are typically funded through utility-wide rate increases rather than industrial-customer-specific surcharges, particularly where local economic-development agreements include preferential industrial tariffs.
4. Source-water competition during peak heat. Cooling water demand peaks during summer heat waves — exactly when residential demand also peaks and source water levels are lowest. The temporal overlap is a structural risk that capacity planning models have historically not captured well, because data centers are a relatively new category of high-volume continuous water user.
What’s Actually Being Done About It
| Jurisdiction | Year | Action |
|---|---|---|
| Chandler, AZ | 2015 | Ordinance restricting water-intensive businesses; data centers capped at 115 gal/day per 1,000 sq ft |
| The Dalles, OR (court order) | 2022 | Google water use disclosed after multi-year litigation |
| Mesa, AZ | 2023–2024 | Industrial water cap; supplemental supply purchase requirement for excess |
| Avondale & Phoenix, AZ | 2023–2024 | Similar industrial water cap ordinances |
| Oregon | 2024 | Mandatory water use disclosure for large data centers |
| Virginia (proposed) | 2025 ongoing | Disclosure legislation debated, not yet passed |
| Investor pressure | April 2026 | 12+ shareholders file disclosure resolutions at AMZN/MSFT/GOOGL |
| EPA federal action | None to date | No federal water-efficiency standards for data centers |
Federal and state policy responses are uneven. Oregon passed legislation in 2024 requiring data centers above a defined size threshold to disclose water use publicly. Virginia’s General Assembly has debated, but not passed, similar disclosure requirements for the nation’s largest data center cluster. The EPA has not yet proposed federal water-efficiency standards specific to data center cooling, despite establishing detailed efficiency standards for residential plumbing fixtures over the past three decades.
Industry self-regulation is more visible than federal action. The Uptime Institute publishes Water Usage Effectiveness benchmarks. Best-in-class facilities now report WUE values around 0.1–0.5 L/kWh, while older or evaporatively cooled facilities frequently report 1.5+ L/kWh. The voluntary disclosure remains the exception rather than the norm.
Municipal pushback has emerged in pockets. Hyperscale projects have been blocked or substantially scaled back by local water utility input in Mesa, Chandler, and Goodyear, Arizona, and several Atlanta-metro counties in the past three years. The pattern in those cases: utilities ran capacity analyses that revealed the proposed industrial demand would compromise drought reliability for residential customers within the existing service territory.
What to Watch in 2026–2028
| Signal | What It Means |
|---|---|
| State-level disclosure laws (post-Oregon) | Whether transparency reaches Virginia, Texas, Arizona, Georgia |
| EPA WaterSense for data centers | Could mark first federal efficiency standard |
| Chip fab buildout pace (TSMC, Intel, Micron) | Each fab adds ~10M gal/day demand |
| AWS / Microsoft / Google site-level disclosure | Investor-pressure resolutions test whether voluntary norms hold |
| AI workload growth rate | If LLM training scales as projected, water demand outruns efficiency gains |
| Reclaimed-water infrastructure investment | Could decouple new builds from potable supply |
| Drought conditions in AZ, TX, GA | Test cases for industrial-vs-residential prioritization |
What You Can Do
-
Find out what’s near you. State commerce departments and county economic-development agencies typically publish active and proposed industrial siting projects. Northern Virginia, Phoenix, Atlanta, Central Texas, the Pacific Northwest, and central Iowa currently host the densest concentrations.
-
Ask your utility about industrial water commitments. Public utilities are subject to open-records laws. Industrial-customer contracts that include reservation, drought-priority, or preferential-rate provisions are typically discoverable through public records requests. Watch for the term “take-or-pay” in industrial contracts — these guarantee payment for water reserved even if not consumed, and they typically include drought-priority provisions.
-
Engage at the planning stage. Industrial siting decisions are typically subject to county or municipal planning commission review. Water capacity analyses tied to these proposals are public information when conducted. The Mesa, Chandler, and Goodyear cases all involved meaningful municipal pushback at this stage.
-
Track the disclosure resolutions. The 2026 shareholder resolutions at Amazon, Microsoft, and Alphabet will be voted on at annual meetings in spring–summer 2026. Even non-passing resolutions establish disclosure expectations that frequently shape voluntary reporting in subsequent years.
-
Address residential exposure separately. Data center growth does not directly contaminate your drinking water. It does contribute to longer-term source-water stress, particularly in already-stressed regions. The residential interventions for that stress are the same ones for any source-water variability — see our guide to the best reverse osmosis systems for filtration that handles salinity changes, and our guide to reading your CCR for tracking shifts in your utility’s source water and treatment.
Frequently Asked Questions
How much water does ChatGPT use?
Per a January 2026 Global Water Intelligence analysis, 30 minutes of generative AI use consumes slightly more than 0.16 gallons of water once direct cooling and indirect electricity-generation water are accounted for. A widely cited 2023 academic study estimated ChatGPT consumes roughly 500 mL (about 17 ounces) per 5–50 prompts depending on data center efficiency and location. Per-query estimates vary by an order of magnitude based on facility WUE and electricity mix.
Are all data centers equally water-intensive?
No. Cooling architecture and climate matter enormously. An air-cooled facility in Oregon can consume well under 0.2 L/kWh of IT load. An evaporatively cooled facility in Phoenix can exceed 1.5 L/kWh. The same operator may run very different WUE values across its portfolio depending on local conditions.
Why is most water use not disclosed?
Industrial water-use data is typically treated as commercially sensitive by both utilities and operators. State-level open records laws vary in their treatment of industrial customer data. Of the three largest cloud operators (Amazon, Microsoft, Google), only Google publishes facility-level water use as of 2024–2025 — and only for sites it owns and operates directly.
Does my drinking water utility serve data centers?
Possibly. Most large municipal utilities serve a mix of residential, commercial, and industrial customers without restricting category. If a hyperscale data center is sited in your service territory, the utility almost certainly has a service contract with it. Loudoun Water’s residential-vs-data-center water mix is a published example; most utilities do not publish equivalent breakdowns.
Can data centers use reclaimed water instead?
Yes, and a growing number do. Google reports that more than 25% of its data center campuses use reclaimed wastewater or other non-potable sources. Amazon Web Services reports 20+ facilities cooling with purified wastewater. Reclaimed-water infrastructure is more expensive to build and requires utility coordination, but it removes new builds from competing with residential drinking-water supply. Loudoun Water and several Phoenix-metro utilities are actively expanding reclaimed-water infrastructure for this purpose.
How WaterVerge Tracks This
WaterVerge does not currently track facility-level industrial water consumption — that data is not centrally compiled by the EPA and not consistently disclosed by operators. Our city pages emphasize EPA-regulated drinking water contaminant data: violations, lead, PFAS, nitrate, disinfection byproducts, and other regulated parameters from SDWIS and the Envirofacts LCR API.
As state-level disclosure regimes (following Oregon’s 2024 lead) generate consistent industrial water data, we will integrate it into city- and state-level coverage where that data affects source-water stress and treatment outcomes. For now, the most direct way to assess local data center water pressure is to combine state economic-development project lists, county planning commission records, and utility-published industrial customer disclosures.
Search your city to see your local utility’s regulated contaminant profile, and watch the source-water and treatment-changes sections of your annual Consumer Confidence Report for shifts that may signal upstream demand pressure. For ongoing source-water variability, the most effective household intervention remains a certified reverse osmosis system at the kitchen tap.