A hyperscale data center running at full load can consume 1–5 million gallons of water per day, comparable to a town of 10,000–50,000 residents. For decades, evaporative cooling towers remained attractive because they minimized capital costs. That tradeoff is breaking down as three constraints converge: AI rack densities exceeding the capabilities of air cooling, water-use disclosure turning private operating costs into public KPIs, and political resistance that can outweigh tower capex savings.
This is not primarily an ESG transition. It is a thermodynamic and permitting response. Roughly 25–30% of new global data center projects now incorporate advanced closed-loop liquid cooling, up from under 5% several years ago. Goldman Sachs estimates that 76% of AI servers deployed by the end of 2026 will be liquid-cooled. Reduced water consumption is a valuable consequence of solving the underlying performance problem: preventing thermal throttling at 50–135+ kW per rack.
The Physics Constraint
Open-loop cooling deliberately evaporates water to reject heat, requiring continuous makeup water. Traditional systems consume roughly 2–5 million gallons per megawatt annually, or 200–500 million gallons for a 100 MW facility.
Traditional racks operated at 5–10 kW. AI training racks now exceed 100 kW without a comparable increase in physical footprint. ASHRAE’s TC 9.9 committee recommends direct-to-chip cooling above 20 kW per rack, a threshold crossed by nearly every serious AI training deployment in 2025.
Recent benchmarking places row-based air cooling near a 29.5 kW ceiling. Direct-to-chip and immersion systems support 60–120+ kW, while some platforms target 200 kW. Above air-cooling limits, hardware throttles, computation slows, and expensive GPUs sit underutilized. Direct-to-chip cooling is therefore being adopted for performance assurance, with water efficiency as a structural co-benefit.
PUE and WUE Compression
Closed-loop performance appears in two metrics: Power Usage Effectiveness, or PUE, and Water Usage Effectiveness, or WUE.
Traditional air cooling operates at roughly 1.50–1.80 PUE. Rear-door heat exchangers reach about 1.20–1.40, direct-to-chip systems 1.15–1.30, and full immersion cooling 1.03–1.08. WUE can fall from roughly 1.8 L/kWh in evaporative-tower designs to under 0.05 L/kWh in leading closed-loop, dry-cooled facilities. Some spray-cooling platforms target zero WUE by pairing direct-to-chip systems with dry coolers.
Microsoft reports an 80% cumulative improvement in WUE across successive data center generations, reaching approximately 0.30 L/kWh in its newest facilities versus 0.49 L/kWh in 2021. Zero-water chip-level cooling designs may save more than 125 million liters per facility annually. One Illinois AI campus is designed to save over 277 million gallons each year compared with conventional evaporative cooling.
These savings result from sealed coolant loops and dry or adiabatic heat rejection. Coolant is continuously reused, while heat is discharged through dry coolers instead of evaporative towers.
WRAP 2.0 and Federal Policy
On April 16, 2026, EPA Administrator Lee Zeldin launched WRAP 2.0, a non-regulatory public-private framework building on the 2020 Water Reuse Action Plan. It emphasizes industry, technology, and reliable water supplies for data centers and semiconductor manufacturing.
Its three pillars cover domestic industrial reuse, water for the technology revolution, and energy development. Priority sectors include data center cooling, semiconductor manufacturing, electricity generation, oil and gas, agriculture, automotive manufacturing, and food production.
A significant provision identifies oil-and-gas-produced water as a potential cooling and mineral-extraction feedstock. That could be especially relevant in areas such as the Texas Permian Basin, where produced water has traditionally been treated as a disposal liability.
WRAP 2.0 does not create binding requirements or override state water law, tribal rights, or local zoning. Its value is directional: it may reduce technical and coordination barriers, but it is not a permitting shortcut.
Social License in Water-Stressed Regions
Water use is cited in more than 40% of contested U.S. data center projects. Over 200 state bills addressing data centers were introduced in 2025, with more than 40 enacted. By mid-2026, bans or moratoriums were under discussion in over 20 states, while more than $130 billion in projects had reportedly been delayed or abandoned during the first quarter.
Two-thirds of data centers built or under development since 2022 are located in water-stressed basins, including the Colorado River Basin, southern Arizona, and Texas. In these regions, data center demand intensifies existing political disputes over scarce water.
Local opposition has already altered development. Chandler, Arizona, imposed commercial water restrictions in 2021 that effectively halted new data center construction. Tucson’s Project Blue was reduced from 59 planned buildings to 11 after sustained opposition. Arizona lawmakers also passed what has been described as the nation’s longest freeze on industry tax incentives.
Permitting transparency has become another grievance. Investigative reporting found that no Arizona data center developers had sought Section 404 federal review since early 2024, while streamlined permits often became visible only after issuance.
Cooling as a Permitting Strategy
Closed-loop, non-evaporative cooling is now both an engineering choice and a community-acceptance strategy. Oracle publicly explains its direct-to-chip systems in response to concerns about local water use. Vantage Data Centers similarly promotes dry cooling for hot, water-scarce climates and uses reclaimed utility water where available.
New AI campuses that retain evaporative towers in stressed basins increasingly face combined permitting, reputational, and stranded-asset risks. Closed-loop cooling does not eliminate every objection, however. Electricity prices, grid strain, construction water, and distrust of opaque permitting remain significant concerns.
What Direct-to-Chip Does Not Solve
Three caveats matter:
- Heat rejection: A sealed chip-level loop connected to an evaporative tower still consumes substantial water. Major savings generally require dry or adiabatic final heat rejection.
- Retrofits: Installing direct-to-chip cooling in a legacy facility does not automatically remove the existing tower. “Waterless” claims should therefore be tested against the complete cooling architecture.
- Wastewater risk: Glycol, corrosion inhibitors, and possible heavy-metal leaching can create compliance issues when coolant is flushed or the system is decommissioned.
Strategic Watch Items
- WRAP 2.0 adoption: Watch how quickly Arizona, Texas, California, and other states operationalize reclaimed-water pathways.
- Section 404 transparency: New federal or state requirements could alter project timelines regardless of cooling technology.
- 2030 water-positive commitments: Google, Microsoft, Amazon, and Meta must close the gap between rising consumption and replenishment.
- Cooling supply-chain consolidation: Direct-to-chip adoption should support investment and M&A among cold-plate, coolant-distribution, and engineered-fluid suppliers.
- Fab recycling benchmarks: Taiwan’s 90%+ recycling rates are becoming a reference point for U.S. and European semiconductor facilities.
The closed-loop pivot is the mechanical result of thermodynamic limits, regulatory scrutiny, and social-license friction. Operators that treat cooling as both an engineering system and a permitting strategy will preserve development flexibility. Those that continue treating water as an afterthought risk seeing multibillion-dollar campuses halted before construction begins.