The expansion of AI and hyperscale data centers has brought the water–power–thermal nexus into the center of infrastructure planning. Cooling is no longer a secondary facilities concern; it is increasingly a determinant of site selection, grid interconnection, permitting, community acceptance, operating reliability and the pace and economics of compute deployment. This matters directly to NVIDIA because sustained accelerator demand produces higher rack densities, greater electricity intensity and more demanding thermal-management requirements across the infrastructure ecosystem, even though NVIDIA generally does not own or operate the facilities in which its products are deployed.
The evidence is concentrated in late July and August 2026, with a smaller set of forward-dated December 2026 monitoring and efficiency claims that should receive less weight in an August 11, 2026 assessment. The central conclusion is nevertheless well supported: water availability is becoming a material constraint in water-stressed regions. The claim that evaporative cooling consumes substantial freshwater resources is supported by three sources spanning May 17 to August 7 1,2,7, while desert-region water availability is likewise supported by three sources 4. A separate three-source claim warns that poorly managed extraction of water and critical minerals for AI infrastructure could damage ecosystems, social structures and political stability 67. Most other claims rely on a single source; individual project allegations and precise consumption figures should therefore be treated as directional rather than independently verified.
Cooling and the Water–Power Trade-off
Water withdrawal is not the same as water consumption
The engineering trade-off is straightforward in principle but often obscured in public discussion. Water-based systems absorb heat and return warmer water 37, while evaporation removes only a portion of circulating water 37. Water withdrawal, water circulation and net consumption are therefore not interchangeable measures: a substantial volume may be circulated or discharged rather than permanently consumed 12,37.
The more cautious and better-corroborated conclusion is that traditional cooling towers depend on evaporation and consequently consume water 65. Evaporative cooling may improve energy efficiency, but it can still impose pressure on local supplies 76. Claims that large data centers do not necessarily consume significant quantities of water 37,49 are not necessarily inconsistent with reports of very large water requirements. Actual demand depends on system design, climate, utilization, operating conditions and the definition of consumption 35,37,61.
Reported benchmarks illustrate the range without making the figures directly comparable. A typical 100MW hyperscale facility is estimated to use approximately 300,000 gallons, or 1.1 million liters, per day 35. Other reports describe individual facilities using several million gallons daily 39, while broader estimates place data-center cooling requirements in the millions of gallons per day 57,70. Fairfax County estimates of 90–100 gallons per day per 1,000 square feet provide another facility-level reference point 35. These figures may measure withdrawals, gross circulation, consumptive use, seasonal peaks or the full campus, and should not be treated as equivalent.
Lower water use can require more power
The fundamental economic difficulty is that reducing direct water use can increase electricity demand. Large-scale cooling requires either substantial water or energy-intensive alternatives, including air-cooled chillers 63. The choice is therefore often one of higher water consumption versus higher power requirements 63, with more power-intensive cooling potentially increasing the water embedded in electricity generation 36. One estimate assigns more than 60% of non-IT data-center power consumption to cooling 77, while other claims link cooling requirements to higher operating costs and sustainability concerns 51. During summer grid stress, both direct electricity costs and the indirect water costs associated with power generation may rise 36, creating a particularly unfavorable operating environment for evaporative systems 36.
Liquid and immersion cooling should consequently be understood as strategic responses to GPU-driven density rather than merely premium equipment categories. Liquid cooling can improve power usage effectiveness 74 and may reduce operating-energy consumption 74. Immersion cooling can reduce dependence on space-intensive precision-air containment 75, and liquid cooling becomes effectively mandatory for racks above 100kW in the cited framework 77. Yet adoption entails integration complexity 75, additional failure points and costs associated with thermal storage, UPS-backed pumps and coolant-distribution units 38. The absence of international standards for dielectric fluids, together with the phase-out of PFAS-related products, remains an industry constraint 75. Fluid substitution is itself an adoption risk 75, and high switching and qualification costs can slow deployment 75.
The proper system boundary extends beyond the facility wall. Lower on-site cooling-water consumption does not eliminate electricity, manufacturing, fuel-cell, battery or upstream-resource impacts 47. Conversely, exclusive reliance on air cooling may become inefficient or exceed practical fan-power and thermal-management limits as rack density rises 38,74. Cooling choices therefore affect long-term sustainability 63, operating costs 63 and competitive positioning across the data-center solutions market 74. Competition is likely to move away from individual air- or liquid-cooling products and toward integrated system performance, deployment compatibility, reliability and scalability 52.
Water Stress as a Siting and Valuation Constraint
Water availability is repeatedly identified as an operational dependency for high-performance computing 26, a site-selection and permitting variable 65, and a determinant of long-term reliability 65. Water stress can raise cooling costs and restrict siting 26, create permitting and community-relations challenges 26, and redirect investment toward regions with more dependable water and energy resources 10,26. Groundwater scarcity or restrictions can increase construction and operating costs 10, constrain technology spending 10, delay projects and complicate permitting 11, and limit site selection, expansion capacity and infrastructure development 11,36.
Virginia as an institutional example
Virginia provides the clearest illustration of how water constraints can enter the development process. The Virginia Department of Environmental Quality assessed future groundwater withdrawals from newer industries such as data centers 35 and concluded that these facilities have the potential to affect groundwater resources significantly 35. A modeled withdrawal scenario reached a similar conclusion 35. Claims indicate that groundwater constraints are already affecting expansion plans 11, while new developments seeking groundwater may need to rely on municipal or alternative supplies 36. Municipal water is likely to become more important as on-site wells become less viable 36.
Developers in the Coastal Plain may need to incorporate groundwater-management requirements into site planning 36. Reliance on municipal and alternative-water infrastructure can increase capacity, cost and schedule risks 36, and developers may ultimately need to move outside established Virginia clusters if local constraints cannot be resolved 36. Water availability has consequently been described as a potentially decisive constraint on Virginia’s data-center ecosystem 36.
Regional and international exposure
The risk is geographically broad. Water availability is a significant constraint in desert regions 4, can stress precarious supplies in Arizona 41, and is particularly acute in West Texas and the Texas Panhandle 69. Texas projects are being evaluated for both electricity and water use 24, face scrutiny over water supply 24, and may encounter conflicts over water resources 18. Wisconsin projects likewise face concerns about significant consumption 15, adequate access 15, and possible environmental or permitting consequences from withdrawals from the Wisconsin River 25. The proposed Jay facility’s projected 300,000 gallons per day is specifically identified as an operating risk 41.
International examples reinforce rather than resolve the concern. Approximately 72% of China’s computing capacity was reportedly located in areas facing severe water scarcity as of 2022 73, while roughly two-thirds of U.S. data centers built since 2022 were reportedly located in areas of high water stress 73. Water scarcity affects major Indian cities where data centers operate or are being developed 7, constrains approval and expansion in India 7, and is operationally important even where water costs are financially small 7. Indian data centers are estimated to consume close to 150 billion liters annually 63, although the methodology is not provided.
Taiwan’s semiconductor industry has experienced conflicts with agricultural users and communities over water 33, and semiconductor fabrication demand can create freshwater tensions with those same groups 34. The proposed Google–Adani project illustrates the potential conflict with household and rural drinking-water needs 64. Google’s proposed Visakhapatnam facility has drawn claims that it could worsen rationing and reduce long-term water availability 17. Microsoft’s Aragón project faces similar water-scarcity, availability and operating-cost risks 39, while the PNK Group project’s water requirements could contribute to cost inflation and local resource competition 25.
Competition among essential uses
The macroeconomic risk extends well beyond the price of water. Persistent groundwater depletion can affect households, public systems, industrial users, ecosystems and communities 36. Water may be contested among industrial, agricultural, ecological and residential users 26, with data centers potentially competing directly with drinking-water requirements 63, agriculture and households during hot weather 55, or agricultural and community uses more generally 31. Months without meaningful rainfall and glacier loss can intensify competition among power generation, cooling, agriculture and human consumption 26.
Climate-driven supply shocks can affect energy, transportation, agriculture and technology simultaneously 26. Prolonged drought, glacier loss and low river levels can generate cross-sector infrastructure failures 26, while a Danube drought is cited as a scenario capable of disrupting power generation and transportation together 26. Water constraints may therefore become a bottleneck across data centers, utilities, nuclear power, logistics, agriculture and drinking-water systems 26.
This correlation matters for valuation. Drought-sensitive utilities, nuclear operators, infrastructure owners and transport businesses may face interruptions and cash-flow volatility 26. Regional water stress can affect industrial output, agriculture, food prices, logistics, tourism and technology operating costs 26. Glacier loss and changing precipitation patterns increase uncertainty about the recurrence and persistence of shortages 26, while climate stress can affect energy infrastructure, manufacturing predictability, worker productivity and adaptation or insurance costs 48. For AI infrastructure, the risk set includes water availability, drought, cooling failure, power interruption, downtime, higher operating costs, capacity constraints and geographic concentration in stressed regions 26. Major shortages could sharply restrict new data-center development 69, making water availability a potential long-duration operating liability for developers and AI-infrastructure companies 36.
Regulation, Transparency and Social License
The policy environment is moving from voluntary sustainability discussion toward approval conditions and public accountability. Toronto and Mississauga are reassessing proposed developments because of electricity demand, water use, heat, infrastructure and environmental effects 61. More broadly, data-center development faces permitting, infrastructure, environmental, social-license and community risks 19. Protests and petitions can threaten data-center, cloud and AI-infrastructure businesses 13, while community opposition may encompass electricity demand, water use, environmental degradation, public costs, future expansion and loss of local control 55. Local resistance and grid constraints may limit deployment 41, and reported displacement of residents near U.S. developments adds another community-relations and permitting dimension 27.
Political scrutiny is increasingly concerned with who bears infrastructure costs. Regulators and communities may question whether ratepayers are financing infrastructure dedicated to data centers 60. Ohio-related claims allege that environmental and infrastructure costs could be shifted to households through higher utility bills 57,70, while central Ohio expansion has been accused of draining water tables, overloading power grids and consuming farmland 70. A project could also shift electricity and operating costs onto residents and students 16. These are claims from individual reports rather than corroborated findings, but they identify a recurring social-license risk: economic-development benefits may not offset local resource and infrastructure burdens 15. If data-center growth underperforms, public budgets may face shortfalls 55, potentially making local authorities more sensitive both to promised tax revenues and to the risk of stranded infrastructure.
Disclosure is a particular weakness. Regulatory loopholes in California and parts of the Midwest allegedly allow data centers to avoid reporting water consumption 57,70, while secrecy surrounding Texas development can impede assessment of project water use 9. Water-use reporting loopholes may weaken accountability 57, and unverifiable water-neutral claims create greenwashing risk 63. Governance proposals increasingly emphasize ownership disclosure and transparent reporting of electricity and water use 66. Sustainability rules require disclosure of energy use, water use, heat reuse and renewable power 62, while mandated audits of facility power and water usage are becoming part of the approval process 24.
The Texas interconnection audit reportedly examines whether projects have disclosed water usage 22. This requirement adds a regulatory and operational hurdle for high-load data-center, cloud and GPU projects seeking grid connection 20. The likely policy direction is tighter control of industrial withdrawals, environmental permitting, river use, thermal discharge, drought contingencies and allocation among drinking water, agriculture, energy and industry 26. Competition among jurisdictions for data-center investment is increasingly constrained by consumer-protection and environmental requirements 66. Sustainability rules could restrain construction if they are not accompanied by power and permitting reform 62. EU policy is moving toward periodic measurement and reporting of energy efficiency, emissions, water use and waste-heat recovery 5,42,62, with emerging requirements covering energy use, efficiency, heat utilization and sustainability reporting 5. These obligations create compliance costs while supporting demand for measurement, monitoring and infrastructure services.
Mitigation and Its Limits
The most prudent mitigation hierarchy consists of closed-loop systems, wastewater or reclaimed water, lower-water cooling, improved heat management, renewable or lower-carbon power, and better monitoring. Newer closed-loop designs and wastewater use may reduce environmental impacts relative to traditional evaporative cooling 35. Closed-loop systems can materially reduce water demand 65, dissipate heat while reducing water use 37, and increasingly recirculate water to reduce freshwater withdrawals 35. Wastewater reuse is identified as a principal sustainability issue 35, and some Virginia data centers already use wastewater rather than potable water or groundwater 35. Moving away from cooling towers could substantially reduce site-level water consumption 47. Fuel cells may also reduce water consumption 59, while broader efficiency, renewable-energy, intelligent-management, liquid-cooling and immersion-cooling approaches are operational priorities 53.
Heat recovery offers a second pathway. Large data centers must manage excess heat that would otherwise be dissipated 37, and failure to use or transfer waste heat sacrifices an efficiency benefit while imposing greater local cooling or energy burdens 37. Proposed integrated systems would reuse waste heat for vertical farming and localized food production 30, potentially reducing energy waste and improving food-security resilience 29. The concept is not without technical difficulty: integration between liquid and air cooling loops presents challenges 29. Cooling inefficiency, airflow recirculation and overcooling also affect the potential for heat recovery 5.
Coastal or floating concepts illustrate more radical alternatives. Seawater cooling and coastal siting could address resource-consumption concerns 14, while coastal desalination could support data-center water needs 14 and reduce freshwater use by combining water production with data-center infrastructure 14. Floating facilities are claimed to avoid potable-water use and discharge 68 and could reduce competition for land, potable water, thermal discharge and noise 68. Yet desalination is energy-intensive and may increase rather than eliminate power demand 14. Coastal projects also face brine-disposal, marine-ecosystem, seawater-intake, construction, noise and governance risks 14, together with coastal hazards, complex permitting and saltwater corrosion 14. These approaches may diversify the solution set but introduce new capital, execution and siting risks.
Monitoring and operational intelligence are important because efficiency gains depend on workload, weather and facility behavior. Standardized measurement can expose inefficient cooling, wasted energy, airflow problems and opportunities for waste-heat recovery 5. Cooling coordination, forecasting, carbon-aware management, liquid cooling, AI-controlled cooling, load distribution and continuous monitoring could reduce energy waste 42,43. Efficient cooling and workload orchestration can lower both operating costs and environmental impact 44, while free-cooling systems may reduce cooling-energy consumption by approximately 30–50% 44. Sustainability pressure and regulatory requirements are structural drivers of monitoring adoption 5, with planned indicators including PES, REF, ERF and WUE 5.
The operational caveat is data quality and model risk. Monitoring data may be missing, invalid, unsynchronized, heterogeneous or sampled at different frequencies 5. Thermal failures, recirculation, leakage, overheating and cooling inefficiency may cluster under changing workloads 5, alongside thermal bottlenecks, airflow anomalies, hot-air recirculation, cold-air bypass and nonuniform inlet temperatures 5. Overheating, overcooling, inefficient cooling, excess facility power use and concentrated server loads remain risks 5, while maximum rack density is associated with inefficient cooling 21. Ambient-temperature fluctuations and uncertain load demand affect server operating conditions 6. More complex forecasting models can create implementation and trust risks if inaccurate decisions impair reliability, thermal stability, energy costs or carbon management 43. Cybersecurity and data-integrity risks include false-data injection, compromised sensors, false triggers and manipulation of AI-driven controls 44.
The Interdependence of Water, Power and Infrastructure
Water constraints cannot be separated from grid availability. AI-campus expansion may increase electricity use, carbon emissions, water use and local infrastructure demands 8, while grid constraints are a structural driver of the thermal-management industry 75. Energy availability is a risk for a proposed 10GW campus 72, and energy-infrastructure constraints can limit project development 32. Texas facilities face grid-reliability concerns 23,28, with grid-reliability failure identified as a principal expansion risk 71. Power-grid outages and scarcity during heat events threaten businesses dependent on reliable electricity 3. Data centers that disconnect from the grid to protect their own servers can shift disturbances onto the grid, creating systemic reliability risk 47.
This makes co-optimization of power, cooling and water more important than treating them as separate procurement decisions. Energy-efficient design, renewable generation, high-efficiency power supplies, workload scheduling and AI optimization can reduce environmental impact 44. Battery lifetime, water consumption and carbon emissions should be optimized together 44, although battery state-of-charge management, microgrid co-design and standardized benchmarks remain open technical gaps 44. A proposed operating framework uses adaptive uncertainty-regulation parameters for load and environmental-temperature risks 6, while another is designed to preserve mission-critical services and shift degradation to non-critical workloads 44. These concepts are promising, but they do not establish mature commercial capability.
The environmental outcome depends on electricity source, cooling method, facility utilization, embodied equipment, redundant capacity and the cost of maintaining both air and liquid systems during transition 38. Hyperscale facilities consume water and energy and generate carbon emissions and toxic waste 45. Cooling chemicals, refrigerants, wastewater and fuel storage create hazardous-material, pollution, spill, fire and contamination risks 55. Excess heat and inefficient heat disposal remain material concerns 37, including the risk of discharging heated water into warm rivers or natural bodies 48,61. Biodiversity damage is another potential consequence of expansion 40.
Implications for NVIDIA
For NVIDIA, this is best understood as a demand-quality and deployment-friction issue rather than a direct water-cost thesis. Escalating AI rack density is a structural driver of thermal-management demand 75. NVIDIA’s accelerator-led growth increases the likelihood that customers will require liquid cooling, higher-capacity power delivery, more sophisticated controls and expanded water or wastewater infrastructure. The opportunity therefore extends beyond GPUs into the broader data-center stack: thermal-management suppliers, coolant-distribution units, pumps, heat exchangers, monitoring software, reclaimed-water systems, municipal infrastructure and integrated power-management solutions stand to benefit from the same secular trend. Providers of reclaimed water, advanced cooling, water-efficiency, environmental consulting and municipal infrastructure are specifically identified as potential beneficiaries 36, while demand relevant to EMCOR is shifting toward water and wastewater infrastructure 50. Veolia likewise sees sustained demand from water scarcity, climate stress, pollutant elimination, industrial licensing, local energy and essential infrastructure 58.
The negative implication is that physical constraints can slow the conversion of GPU demand into installed compute capacity. Water scarcity may limit or redirect AI-infrastructure growth 36, constrain facility expansion 69, and increase costs, delays, alternative-water requirements or operating-capacity limits 17. Cooling choices affect reliability, energy costs and compliance, while water constraints can force capacity away from traditional clusters 10. Because data gravity, network architecture, sovereignty, cooling conditions, software and accelerator utilization limit workload mobility 54, customers cannot always move workloads to the cheapest or most water-secure region. NVIDIA’s ability to support efficient, dense and thermally manageable systems therefore has strategic value, but its customers remain exposed to local permitting and infrastructure bottlenecks.
NVIDIA’s competitive position should consequently be assessed on total system performance rather than accelerator performance alone. Liquid cooling, improved PUE, workload orchestration, monitoring and heat recovery can help customers deploy more compute within power and water limits 5,53,74. Adoption, however, depends on integration, standards, qualification, reliability and operating data 5,75. The absence of a mature, universally applicable solution for gigawatt-scale liquid cooling, high rack density, low PUE, networking and operational stability is identified as a technology risk for IREN 56. Although company-specific, that example highlights a broader ecosystem constraint relevant to NVIDIA’s customers and partners.
The most important financial risk is not necessarily the absolute price of water, but the possibility that water and power availability become gating factors for capacity additions. In India, water costs may be small while availability remains a major operational constraint 7. In Virginia and Texas, water and grid limitations can raise capital and operating expenses, delay approvals and redirect investment 11,36,69. Investors should therefore monitor customer-backlog quality, campus permitting status, water sourcing, cooling architecture, grid interconnection and deployment geography—not merely announced GPU orders. A regulatory review or groundwater moratorium could affect valuations of data-center developers and cloud and AI-infrastructure companies 10, with second-order implications for NVIDIA’s growth expectations if customers defer or resize projects.
The subject also creates an opportunity for NVIDIA to reinforce ecosystem standards around liquid-cooling reference designs, telemetry, predictive thermal control, energy-aware scheduling and workload flexibility. Claims concerning proprietary or advanced cooling concepts remain uncertain: critics question the feasibility and long-term water performance of Google’s proposed technologies 17, while proponents of the Armenian project claim closed-loop cooling without providing verified consumption figures 46. Similar uncertainty surrounds water-neutral claims 63 and reported consumption figures. Investors should distinguish demonstrated reductions in net water consumption from marketing claims based primarily on recirculation or avoided potable-water use.
Finally, social license may affect the pace of AI-infrastructure growth. Public inquiries about data-center water and energy use have become routine 47, while opposition increasingly encompasses environmental damage, electricity prices and community control 41. Examples from Toronto, Mississauga, Ohio, Texas, Virginia, Wisconsin and India indicate that scrutiny is geographically diverse rather than isolated. For NVIDIA, this increases the value of customers and partners that can document water performance, secure community consent, internalize infrastructure costs and demonstrate reliable grid behavior. It also raises the possibility that sustainability reporting, environmental permitting and local opposition become binding constraints on the addressable market.
Key Takeaways
- Water is emerging as a physical and regulatory bottleneck for AI infrastructure, particularly where evaporative cooling, drought, groundwater restrictions and summer grid stress coincide 1,2,4,7,26.
- The central technology trade-off is higher water use versus higher electricity use. Closed-loop, wastewater, liquid and immersion cooling can reduce direct water demand, but they introduce capital, integration, standards and reliability risks 37,63,65.
- For NVIDIA, the issue is primarily one of ecosystem execution: GPU demand must be converted into permitted, powered and thermally manageable capacity. Customer geography, cooling architecture, water sourcing and grid access are therefore important indicators of growth quality 36,54,65.
- Monitoring, heat recovery, workload orchestration and integrated water-power solutions are potential enablers and investment themes, but forward-dated measurement claims and project-specific water figures remain insufficiently corroborated 5.