AustinIO > Texas Energy Nexus > Water & Cooling

Water & Cooling in Large Facilities

Power is one axis of a large facility's grid dependence. Water is the other. The August 2026 Texas audit asks every project two self-sufficiency questions side by side — do you bring your own power, and do you bring your own water — and the second is where the harder physics lives. A campus can generate its own electricity behind the meter and still drain an aquifer a small town depends on. Reducing that draw is a different engineering problem, solved by cooling architecture and water reuse rather than by generators.

There is no true "behind-the-meter" for water. A meter is a physical device with a grid on one side, and water has no equivalent geometry. The right frame is water self-sufficiency: how little outside water a facility has to withdraw, and how much of what it uses it returns. Data centers are the loudest case here too, but fabs consume more, and the whole Triad is converging on the same answer — use less, reuse more, and site where the water math works.

Why these facilities drink

Two different mechanisms, often confused. A data center uses water almost entirely to reject heat — evaporative cooling towers throw heat to the sky by evaporating water. A semiconductor fab uses water mostly in the process itself, rinsing and cleaning wafers with water purified to a grade found nowhere in nature. The first is a cooling problem you can engineer around by not evaporating. The second is a chemistry problem you can only shrink, never eliminate.

The distinction matters because the fixes are different. A data center can go from thirsty to nearly dry by changing how it cools. A fab cannot stop needing ultrapure water — it can only recycle a larger and larger share of what it draws.

Data center cooling: the water-versus-power trade

Cooling architecture is now as consequential as compute architecture, because the choice of how to reject heat sets a facility's entire water footprint. The measure is Water Use Effectiveness (WUE) — liters of water per kilowatt-hour of IT load — and the options span two orders of magnitude:

Evaporative cooling — high water, low power. WUE around 1.5 to 2.5 L/kWh. A medium campus can consume 300,000 gallons a day, roughly 1,000 households, and 70 to 85% of that water evaporates and never returns. Best suited to water-abundant, cost-sensitive markets.
Closed-loop — the same coolant recirculates instead of evaporating, filled once and reused. On-site water drops to roughly 5 to 10% of an evaporative design. Best-in-class closed-loop reports below 0.05 L/kWh.
Air-cooled — near-zero on-site water, paid for in electricity. This does not eliminate water so much as move it upstream to the power plant, unless the campus also generates behind the meter.
Direct-to-chip and immersion — liquid delivered to the silicon, near-zero water and low power. The frontier for high-density AI in water-stressed markets; closed-loop liquid systems can reach far better water efficiency than air.

The industry shift is decisively toward the low-water end for new builds. Microsoft's next-generation designs use closed-loop, zero-evaporation cooling filled once at construction, cutting more than 125 million liters a year per facility. Location still dominates the number — the same operator reports Arizona WUE near 1.5 and Singapore near 0.02 — which is why arid-region campuses lead the move away from evaporation.

Fab water: the ultrapure floor

Semiconductor fabs are the heavier consumers, and their water cannot be engineered to zero. A large fab uses on the order of 10 million gallons a day, rivaling a small city, most of it as ultrapure water (UPW) — purified to roughly 18 MΩ·cm resistivity, virtually free of particles, ions, and organics, orders of magnitude cleaner than drinking water. Producing UPW is itself wasteful: it takes 1,400 to 1,600 gallons of municipal input to make 1,000 gallons of UPW, because the low contaminant tolerance sends much of the intake to a reject stream.

Because a fab cannot stop needing UPW, the entire lever is reuse. Conventional treatment recycles 40 to 70% of process water; best-in-class fabs now reach 85 to 90%, and advanced reclaim can push a facility that once drew 10 million gallons of fresh water a day down to a small fraction of that from outside. The catch is scale: SK Hynix set a claimed world record recycling 170,000 tons a day, and its single Yongin cluster expansion will consume more water annually than eight years of that conservation saved. Reuse rates rise while absolute demand rises faster — the same pattern the grid shows on the power side.

Sourcing: not all water is the same

Self-sufficiency is also about which water a facility draws. Pulling potable municipal supply is what triggers community opposition; drawing water nobody else can use does not. The Texas campuses leading on this pick their source deliberately: Google's Meitner center is air-cooled and limits water to domestic use; Microsoft's Pecos campus and several West Texas projects run closed-loop on brackish, non-potable groundwater that has no competing use; others build on reclaimed municipal wastewater, funding the treatment upgrades themselves. Sourcing from brackish or reused water is often the difference between a permit and a lawsuit.

Why it is the second axis of the audit

Power self-sufficiency is a speed play — bring your own generation and skip the grid queue. Water self-sufficiency is a permitting and community play — draw little, reuse most, and take it from a source no one contests, so a project clears local resistance and the state audit. The Texas review asks for annual and peak water consumption, the source, the cooling technology, and whether water is reused, weighting each project on exactly these choices.

Both axes point the same way and reward the same posture: facilities that provide their own power and reuse their own water are the ones Governor Abbott has praised and the ones the audit is built to advance. Across the Triad, the winning design is now self-sufficient on both counts — generation behind the meter, cooling in a closed loop.

Last updated September 11, 2026. Water figures are typical ranges; per-facility consumption varies with cooling design, climate, and node. See individual facility pages for detail.