A data center is a heat-management system as much as a computing system.

Heat has to go somewhere

Rows of servers turn electrical energy into computation and heat. Fans can move that heat from chips into a room, but the building still needs a way to release it outdoors. Some facilities use air-cooled chillers or outside air during favorable weather. Others use cooling towers, where evaporation carries heat away efficiently but consumes water. Liquid-cooling systems bring a fluid closer to hot components, which can move heat more effectively than air as chip power rises. A closed loop can recirculate that fluid, yet the equipment that finally rejects heat may still use air, water, or both.

There is no universally best cooling system because water and energy trade places. Evaporative cooling can reduce electricity demand in some climates while increasing onsite water consumption. Dry cooling can reduce direct water use but may require more power, more equipment, or reduced efficiency during very hot weather. Hybrid systems switch modes as conditions change. Engineers also have to protect equipment reliability, water quality, worker safety, and cost. The meaningful comparison is not a dramatic label such as “waterless,” but how a design performs in its actual climate, grid, and watershed throughout the year.

Count the facility, then widen the boundary

Operators often report water-use effectiveness, or WUE, as liters of onsite water used per kilowatt-hour consumed by computing equipment. That can reveal whether a fleet is becoming more efficient, provided the method and boundary stay consistent. It does not automatically include the water associated with electricity generation, chip fabrication, construction, or upstream materials. Total annual withdrawal answers another question, while consumption describes water that is not quickly returned to its source. Reclaimed wastewater may reduce demand for drinking-quality water even when a facility still reports a substantial total volume.

Independent research makes the need for context especially clear. Lawrence Berkeley National Laboratory found that modeled water use across data-center workloads can differ by more than 10,000-fold as hardware, utilization, cooling, electric grids, location, climate, and timing change. That makes a single industry-wide “water per search” or “water per prompt” number a poor guide to a local project. Better disclosure pairs fleet metrics with site-level information: the source of water, monthly or seasonal demand, local stress, drought plans, discharge or reuse arrangements, and the portion of electricity-related water included in the estimate.

Siting therefore becomes part of system design. A project can be efficient on paper and still strain a small utility during a dry season, while a larger facility using reclaimed water in a better-prepared basin may present a different tradeoff. Communities need enough information to discuss drinking-water capacity, wastewater infrastructure, power demand, emergency reserves, rates, ecosystem flows, jobs, and future growth together. National averages are useful context, but they do not replace a local water budget or a durable agreement with the people operating the watershed’s infrastructure.

Current initiatives show several routes to lower impact

Company reporting points to a portfolio rather than one cure. Microsoft says closed-loop direct-to-chip designs can avoid continual evaporation for operational cooling and reports that much of its owned fleet now uses low- or zero-water cooling approaches. AWS reports greater use of outside-air cooling, reclaimed water at operating sites, and lower fleet WUE. OpenAI describes closed-loop cooling at its Abilene Stargate campus. These initiatives can reduce pressure at a facility, but their published figures remain first-party claims with different definitions, years, and scopes. A design announced for one campus should not be projected across an entire company.

Google and other operators also report watershed projects intended to replenish water through wetland restoration, irrigation improvements, leak repair, aquifer recharge, or similar local work. Those projects may create genuine community and ecological value, and public project portfolios make scrutiny easier. They are still different from using less water onsite. Estimated replenishment can occur at another time or place, and a global percentage can hide a stressed facility. The strongest strategy combines efficient computing, climate-appropriate cooling, reclaimed water where safe and practical, careful siting, transparent local data, and projects chosen with the watershed—not as a substitute for reduction, but as another layer of stewardship.