Water Availability, Sourcing & Climate-Driven Cooling Strategy (Siting Gate)

Power is the first gate. Water is the second, and in a growing list of markets it is the gate that kills deals, because you often cannot buy your way past it. You can pay for a faster interconnection, a behind-the-meter turbine or a pricier PPA. You cannot conjure a firm water right in an over-allocated basin in a drought year. That asymmetry is why this guide treats water as a mandatory gate, not a scored criterion: run it at desktop diligence, like power, before land is controlled and long before the cooling vendor shows up.

Water also sits inside a trilemma. A site optimizes for at most two of low water consumption, low energy (PUE) and low capital cost, and the climate decides which corner you are pushed toward. Evaporative cooling is cheap and energy-efficient but drinks water. Dry heat rejection, with no evaporation, uses almost no water but burns more energy and costs more to build. In hot weather a dry plant needs more cooler capacity, or some mechanical trim, to keep the coolant at its setpoint. Hybrid splits the difference. Choose your corner deliberately rather than inheriting it from whatever the EPC proposes. Stewardship (replenishment, WUE targets, disclosure) lives in Chapter 15.4, and the cooling hardware is engineered across Chapter 5.7 and Chapter 5.8.

A data center touches water in two places, and a serious diligence pass counts both. Direct (Scope 1) water is what the site itself uses. Count withdrawal and consumption separately. Makeup water to the cooling towers replaces what leaves as evaporation, blowdown and drift, and the blowdown goes back to the receiving system under a permit. Your supplier and withdrawal permit have to cover the peak intake, and your discharge path has to take the wastewater flow. Scope 2 (indirect) water is what evaporates at the power plants feeding your load. In LBNL’s 2023 US accounting, the water consumed generating the sites’ electricity came to nearly 800 billion liters, roughly 210 billion gallons and about twelve times what the sites consumed directly. HARC’s Texas estimate tells the same story: about 25 billion gallons in 2025, with roughly 8 billion used for on-site cooling and 17 billion, about double that, evaporated at the power plants. A site that looks water-light because it runs dry coolers can still carry a large embedded footprint through a thermoelectric-heavy grid. A hot, dry, solar/wind-rich grid can be the water-frugal answer once you net the two. Decide which scope your stakeholders are optimizing: the regulator and the community usually mean Scope 1, while full stewardship accounting counts both.

Scale is why water has moved up the screen. Direct water consumption by US data centers came to roughly 17 billion gallons in 2023, according to LBNL's December 2024 report. Its 2028 scenarios run to roughly 38–73 billion gallons, over 40% of it from hyperscale facilities. Per unit of energy, LBNL modelled the 2023 US average at about 0.36 L of site water per IT kWh across all cooling types. LBNL's June 2026 update revised its energy estimates but did not revise these water figures. Microsoft reported about 0.27 L/kWh for its owned and controlled fleet in FY2025, and a plant that evaporates nothing runs near zero.

Virginia DEQ's July 2026 report puts a typical 100 MW-IT campus at about 300,000 gallons per day, based on Fairfax water-use intensity. At 0.36 L/kWh the same campus running flat out would consume about 230,000 gallons a day, so the two agree in order of magnitude. Both are planning figures, not peaks. To size an evaporative plant's peak withdrawal, convert the design-day heat load into evaporation, then add blowdown and drift (Chapter 5.8). Water is now the variable that triggers moratoria, disclosure laws and the opposition coalitions covered in Chapter 3.11.

Not all water is the same water. Build the cooling plant around reclaimed (non-potable) water when its quantity, treatment duty, service continuity and discharge obligations beat the alternatives. It can keep you out of competition for drinking water, and it can earn an incentive or meet an approval condition. Locking in that supply does not end the political fight, and it does not tell you whether the source suits your plant. That turns on the water’s fouling, scaling and biological load, which set the treatment capex and the blowdown chemistry. Each source is a different gate with a different lead time; the table lays them out.

The table is a priority stack. In a water-stressed market, the default is to design out of potable dependence. Move to reclaimed supply where it keeps delivering through a drought and its chemistry suits the plant; where it does not, move to dry, non-evaporative heat rejection. The expensive end of the table (desal, RO of produced water) is where operators go when water is the hard gate and power is plentiful, as in the Gulf and parts of the US Southwest. It turns a siting blocker into a capex line. Chapter 5.7 details the mechanical loops and the makeup-water treatment train each source demands.

What makes a site durable is rarely whether water is available in an average year. It is where you rank when there is not enough. In prior-appropriation regimes (the US West), seniority is everything: junior rights are curtailed first. A new appropriation is junior to every older right. Trace the priority of whichever right actually supplies the facility, including a transferred right or the municipal supply, and read the curtailment terms in your own supply contract. In riparian and permit-allocation regimes, the regulator can tighten or decline to renew an industrial allocation when a basin tips into shortage. The penalty is severe and lopsided. If its cooling depends on that water right and it has no dry-mode fallback, a facility that loses the right cannot just throttle back. It loses goodput at the moment the grid is stressed as well.

Curtailment belongs in the design. A site whose cooling can ride through a water-supply interruption turns a hard curtailment risk into a graceful energy penalty. The tools are hybrid towers that pivot to dry operation, on-site storage sized for a defined outage, and a dry heat-rejection plant with evaporative assist only on the hottest days, sized to carry the load in dry mode during a curtailment. A site whose density depends on year-round evaporation has no fallback. Model the failure explicitly: the water-curtailment FMEA in Appendix F treats loss of makeup water as a first-class failure mode with defined detection, ride-through and load-shed responses, the way the electrical chapters treat loss of a utility feed.

Climate decides which corner of the trilemma a site can occupy, through one number: free-cooling hours, the share of the year when ambient wet-bulb (evaporative/adiabatic) or dry-bulb (dry coolers) is cold enough to reject heat without mechanical refrigeration. A cold climate can run the economizer for a large share of the year. Work out the actual hours from the site's hourly weather, your coolant setpoint and your plant capacity (Chapter 5.8). A cold, dry site (the Nordics, the US Upper Midwest, high-altitude interior) can get low water and low energy at once with a dry design. That is the rare site where you do not have to choose. Hot, humid sites (Singapore, the Gulf, the US Gulf Coast) offer fewer economizer hours for cold-water designs. Warm-water liquid cooling widens the hours that dry rejection can cover. NVIDIA says its Rubin platform accepts coolant up to about 45 °C (NVIDIA blog, June 2026); confirm the inlet limit for the rack and OEM configuration you buy. The facility-water setpoint, the CDU approach and the dry-cooler capacity decide which hours run without refrigeration. Outside those hours, air-cooled mechanical trim can hold the coolant setpoint without evaporative makeup, but it costs power and capital.

A cold climate makes dry cooling almost free in water and energy, so take it. A hot-dry climate gives the sharp version of the fork. Evaporative cooling is low in energy and high in water, a bet against the drought. Dry heat rejection evaporates nothing and so uses almost no water. It pays in energy, and in the heat it needs more cooler capacity or mechanical trim to keep the setpoint. Hybrid/adiabatic evaporates only on the hottest days and runs dry the rest of the year, the pragmatic 2026 default for water-constrained warm sites. In a hot-humid climate, evaporation is least effective exactly when you need it most, which is why those markets increasingly turn to seawater, district cooling, or simply accept a higher PUE.

A water-light site fixes how much rejection capacity you need, what coolant temperature you run and how much compressor power you budget. So the water call you make at siting feeds straight into the compute decision in Chapter 1.1. Running the loop warmer uses up thermal margin, and when dry capacity falls short, mechanical trim can hold the setpoint. The industry's move toward warm-water direct-to-chip loops (45 °C-class coolant on NVIDIA's Rubin platform) is partly a water strategy: warmer loops add free-cooling hours and shrink evaporative makeup, so hot-climate sites can stay water-frugal without surrendering as much density. The warm-water loop engineering is in Chapter 5.7, and heat rejection across towers, dry coolers, adiabatic and economizers is the subject of Chapter 5.8.

Every watt the facility consumes leaves as heat, and the default is to pay, in water and energy, to throw it away. The siting question is whether a nearby heat offtaker can take it instead: a district-heating network, a greenhouse, an aquaculture or industrial process, a swimming complex. Where one exists, waste heat flips from a rejection cost into a revenue line and, more importantly for siting, a social-license asset. A community that heats its homes from the data center is a very different political environment from one that only sees the water bill. Nordic markets, with mature district-heating grids and cold climates, accordingly treat heat reuse as a primary siting advantage rather than a sustainability garnish.

The catch, and the reason this is a siting decision rather than a retrofit, is temperature. District-heating supply temperature depends on the network generation: above 100 °C for 2G, 80–100 °C for 3G (IEA DHC TS1, 2017) and typically 50–70 °C for 4G (Yuan 2025), and the offtaker's actual seasonal schedule sets the lift. Air-cooled facilities reject heat far too cold to use directly, and even warm-water DLC loops typically need a heat pump to lift the grade, which costs energy and capital. A site planned for a high-temperature warm-water loop near a district-heating offtaker can deliver usable heat with a small lift; a conventional air-cooled hall far from any network cannot, at any reasonable cost. Screen for the offtaker and loop-temperature compatibility together, at siting. Offtake economics (temperature-indexed offtake, heat-pump integration, the Stockholm-class case studies) are built out in Chapter 5.9, and the sustainability and economics framing in Chapter 15.5.

Heat reuse changes land and utility decisions when the obligation is real. In Germany, EnEfG §11 requires data centers that start operating from July 1, 2026 to reach an annual PUE of 1.2 or better and to reuse at least 10% of their energy, within two years of starting. Sites starting from July 2027 carry a planned reuse share of 15%, and sites starting from July 2028 one of 20%, each due within two years of that site’s own start; the law defines exceptions. Before you reserve the heat route, record which start-date cohort the site falls in, its boundary and its compliance period. For the wider EU, the EED Article 26 guidance treats the heat-use and feasibility obligation separately from the installation's cost-benefit analysis. Get the offtaker’s required temperature, seasonal demand, reliability and price, and settle who is responsible for the connection. Reserve the route and the interface, but do not assume the customer will buy heat in every hour.

The German government's EnEfG amendment, bill 21/8027, is still a proposal as of 16 October 2026, after a committee hearing on 5 October. The bill keeps the PUE and reuse percentages for new sites, stretches the compliance ramp to four years and lets internal heat use count. It also proposes exempting a site from the reuse share when a cost-benefit analysis finds no connection to an existing or planned heat network that is both technically and economically reasonable. A site that takes the exception must still leave space for future heat-transfer infrastructure. Check whether the bill has passed (Chapter 15.7) before you change the site's compliance basis.

Water enters the site-scoring playbook as a hard pass/fail gate first, a weighted criterion second. The gate has two parts, supply and discharge. Does the site have a firm, permitted supply that survives drought? If not, can a dry, non-evaporative plant carry the load, with enough capacity for peak weather and the electrical headroom to run it? And is there a compliant discharge path? A site that fails is out, however good its power or fiber looks, because no cooling plant runs without heat rejection. Only sites that clear the gate get scored on the softer water dimensions that separate a good water site from an adequate one: WUE achievability, reclaimed availability, heat-reuse upside and regulatory durability.

The deepest coupling is back to power and density. A water-rich, cold-climate site lets you run evaporative or dry cooling at low PUE and push maximum density, the two-corner win. A water-poor, hot-climate site can still hold its coolant setpoint by adding dry-rejection capacity or mechanical cooling. That extra power and capital competes with the compute budget. Either way the water decision reaches the power bill and the compute plan, which is why it cannot be left to the mechanical contractor after the land is bought. Water-light siting also shows up as a portfolio strategy: pair cold, water-rich training sites with latency-sensitive online/edge inference sites that accept a higher cooling cost, as the market-clusters and scoring playbook of Chapter 3.13 formalizes.