NVIDIA published a reference design for liquid-cooled AI factories built around its DSX architecture, and the headline number is a water figure. The Rubin-generation design moves to 100 percent liquid cooling, with coolant entering cold plates at temperatures up to 45 degrees Celsius and exiting near 55 degrees, hot enough that dry coolers can reject the heat to the air without mechanical chillers in many climates. The closed loop recirculates a mix of roughly 75 percent water and 25 percent propylene glycol and consumes no fresh water once filled. NVIDIA puts the reduction in facility cooling water at up to 100 percent, falling from about 2.6 million gallons per megawatt per year to near zero.
For a 50-megawatt hyperscale facility, NVIDIA estimates more than $4 million in annual savings on cooling-related energy and water. Ali Heydari, the company's director of data center cooling and infrastructure, framed it bluntly: the DSX reference design "has zero water consumption." Motivair, now part of Schneider Electric's cooling division, collaborated on the architecture, and its CEO Richard Whitmore captured the underlying driver: once watts per chip crossed a certain threshold, liquid cooling stopped being optional. The water savings are a byproduct of a thermal architecture that chip power density made mandatory.
Eliminating evaporative cooling towers removes a genuine and continuous draw on local water supplies, the part of the footprint that strains aquifers and municipal systems in places like Phoenix and the Mountain West. That matters. The problem is the boundary. Both TechCrunch and Fortune made the same point on June 22: drawing a line around the data center and zeroing out the water inside it leaves most of AI's water footprint untouched, because the largest share has never been onsite to begin with.
The water moves upstream to electricity generation. Thermoelectric power plants consume water for their own evaporative cooling, and the volumes scale with how the grid is built. A natural gas plant consumes roughly 1.17 liters per kilowatt-hour and a coal plant roughly 2.2 liters, while wind sits near 0.01 liters and solar near 0.03. Fossil generation supplies about half of data center power today, and U.S. fossil plants consume an estimated 2.7 billion gallons of water daily for cooling. By the analysis cited in the coverage, offsite consumption at the power plant and in chip manufacturing can double or triple the total footprint, which puts the onsite portion NVIDIA addresses at roughly a quarter to a third of the whole.
A closed-loop, dry-cooled data center trades onsite water for electricity and reassigns the water debt to whoever generates that electricity. If the supplying grid runs on gas and coal, the water that no longer evaporates from a cooling tower in the facility yard evaporates instead at the power station feeding it. If the grid runs on wind, solar, and hydro with low evaporative loss, the total footprint genuinely shrinks. The honest version of NVIDIA's claim is conditional on grid mix, and the design itself cannot control that variable. Andrew Chien of the University of Chicago's CERES Center noted that warm-water operation does let facilities shed heat without air conditioning, while cautioning that zero water use across the full system remains unrealistic.
The same dynamic shows up in cooling energy. Higher coolant temperatures reduce the lift required from the mechanical plant, and raising chiller setpoints by a single degree trims cooling energy by roughly 4 percent. Warm-water direct-to-chip turns that lever into an architecture rather than a tuning adjustment, the same logic that runs through NVIDIA's push toward 45-degree cooling that eliminates chillers. Less cooling energy means less generation, and on a fossil-heavy grid, less generation means less upstream water. The water and the power are the same problem viewed from two ends of the wire.
The strategic signal for the cooling industry is that warm-water closed-loop DTC is now the design operators will be asked to build to, because the company defining the rack is also defining the facility around it. Evaporative cooling as the default heat-rejection method is on a clock, and vendors selling cooling towers, CDUs, and dry coolers will be specifying against a reference that assumes a sealed coolant loop and high supply temperatures. This is consistent with how NVIDIA's watt roadmap keeps writing the cooling industry's business plan, one generation at a time.
The caution for operators and the marketing teams behind them is to keep the boundary visible. An onsite-water reduction is a defensible and valuable claim. A total-water claim requires accounting for the grid that powers the building, and conflating the two invites the exact scrutiny that arrived within a day of NVIDIA's announcement. The cooling architecture that ends evaporative draw onsite is real progress. The water ledger for AI gets settled at the power plant, and that is where the next argument about AI's water footprint will be fought. For the full breakdown of the Rubin design, the chiller-free debate, and the offsite water it does not touch, see our Special Report, produced in partnership with NVIDIA.