NVIDIA's Rubin AI Factory Runs Hot to Use Almost No Water. The Open Questions Are Chillers and Power Plants.

NVIDIA used London Climate Week to make a sustainability claim that lands directly on the cooling industry. In a reference design for liquid-cooled AI factories, the company says its Vera Rubin generation is the first NVIDIA AI infrastructure to run on 100 percent liquid cooling, with no fans anywhere in the system, and that the design eliminates onsite water consumption entirely. The number it put on the change is large. Conventional cooling-tower facilities draw roughly 2.6 million gallons of water per megawatt per year. The DSX reference design, NVIDIA says, takes that to near zero.

The engineering is real and the direction is correct. The framing around it is where the work for an operator begins. NVIDIA Chief Sustainability Officer Josh Parker told reporters the water-consumption challenge for data centers is "largely solved." A warm-water closed loop does solve a specific, important part of that challenge. It does not touch the larger part, which sits outside the fence line at the power plant. This report walks through what NVIDIA actually built, why running hot uses almost no water, and the two questions the cooling industry is already pressing on: whether the design is truly chiller-free, and where the rest of AI's water footprint actually lives.

What NVIDIA Actually Built

The Vera Rubin DSX is a full AI-factory reference design, a blueprint that integrates compute, networking, storage, power, and cooling rather than a single product. On the thermal side, the change is a move to direct-to-chip liquid cooling for every heat-generating component, including the networking gear that earlier designs still cooled with air. Cold plates sit on the processors and capture heat at the source. A closed loop carries it to outdoor dry coolers, passive radiators that reject heat to ambient air. NVIDIA developed the architecture with Motivair, the advanced cooling division of Schneider Electric.

The driver is power density. Blackwell-class GPUs already dissipate on the order of 1.2 to 1.4 kilowatts each, and Rubin racks push past 150 kilowatts. Air cannot move that much heat out of that small a volume. "Once the watts per chip crossed a certain level, liquid cooling became mandatory," Motivair CEO Richard Whitmore said. The secondary effects are substantial in their own right. NVIDIA says the design collapses rack height from six rack units to two, removes roughly 85 decibels of fan noise, and replaces perforated bezels with sealed front panels. The hot-aisle and cold-aisle choreography that has organized data center floors for two decades goes away.

Why Running Hot Uses Almost No Water

The water savings are a consequence of temperature, not a separate feature. Conventional facilities use evaporative cooling towers, which throw heat away by boiling off water, and they consume that water continuously. NVIDIA's loop runs warm enough to skip evaporation. Coolant enters the cold plates at up to 45 degrees Celsius, 113 degrees Fahrenheit, hotter than a hot tub, and leaves around 55. At those temperatures a dry cooler can shed heat straight to the outside air across most of the year, the way a car radiator does, with no water boiled off in the process.

The loop itself is filled once and runs closed for the life of the facility. The working fluid is a mix of roughly 75 percent water and 25 percent propylene glycol, closer to engine coolant than to tap water, and it is recirculated rather than consumed. The efficiency logic is the same lever the industry has pulled for years, turned into an architecture. Raising chiller-plant temperatures by a single degree trims cooling energy cost by about 4 percent, and warm-water direct-to-chip raises the supply temperature permanently. NVIDIA puts the facility-level result at more than 4 million dollars a year in combined cooling energy and water savings for a 50-megawatt site, and cites closed-loop efficiency gains of up to 25 times on energy and more than 300 times on water against conventional designs. Those last figures are NVIDIA's own, drawn from its marketing, and worth reading as vendor claims rather than independent measurements.

The same warm return that eliminates the cooling tower also makes the waste heat useful. Coolant coming back at 45 to 50 degrees and above is hot enough to feed district heating, industrial preheating, or water treatment, which is why operators like Nautilus Data Technologies argue data centers do not have to consume water, they choose to. We covered the underlying shift toward 45-degree hot-water cooling that designs the chiller out when NVIDIA first signaled it, and the Rubin reference design turns that signal into a specification the rest of the supply chain will build against.

The Chiller-Free Claim, and the Vendors Who Disagree

NVIDIA frames 45-degree operation as a path to a "chiller-less ideal," a facility where mechanical chillers might run only a handful of days a year. The company is careful to note that geography matters. A data center in the Scottish Highlands and one in Phoenix, Arizona face very different realities, in NVIDIA's own words, though it argues even warm climates move significantly closer to the chiller-free target. It has real third-party support on this point. Microsoft's Steve Solomon, a vice president of data center engineering, said the approach could eliminate the need for any type of mechanical chiller in most climates most of the time, even in hot places such as Arizona, and called it "a big deal for everybody" if every chip could be cooled this way.

Cooling vendors who build this equipment for a living are more measured, and their objection is specific. CoolIT Systems put it bluntly: warm-water cooling reduces chiller dependency, but it does not magically eliminate the need for mechanical cooling. The limiting factor is the temperature delta between the facility water and the outside air. A site can run chiller-free only where ambient air can reliably produce 45-degree water. Data centers are engineered for worst-case conditions, the heat waves and humid afternoons, so chillers stay installed as a risk-management backstop for the hottest days even if they sit idle most of the year.

The physics is unforgiving in hot regions. CoolIT's own example: 45 degree water against 35 degree ambient leaves a 10 degree delta to work with, while against 5 degree ambient the delta is 40, and a fourfold change in delta fundamentally changes how the system performs. Advanced Cooling Technologies, the engineering firm also known as 1-ACT, raises a different concern. Pushing inlet temperatures up strips away the thermal-headroom buffer that used to hide small deviations between components, so the difficulty does not disappear, it concentrates upstream as plate-to-plate variance, pump cycling, and harder-to-tune controls. Both companies sell the cold plates, CDUs, and two-phase hardware that this complexity creates demand for, so their read carries a commercial interest. It is still the more accurate description of what an operator outside a cool climate will actually experience.

There is a serviceability tension as well. A hermetically sealed, fanless loop is elegant until a server or GPU fails and a technician needs to pull it without draining a cluster. Quick-disconnect fittings with dripless seals exist to manage that, and they are a tradeoff rather than a clean solution. None of this makes the design wrong. It means the headline of a chiller-free, sealed facility describes the best case, and the cooling plant an operator specifies still has to survive the worst one.

The Water That Moves Upstream

The larger asterisk is about where the accounting stops. NVIDIA's zero-water claim is measured at the facility boundary. It is accurate inside that boundary and incomplete outside it, because the largest share of AI's water footprint was never onsite to begin with. It sits at the power plants generating the electricity, and to a lesser degree in chip manufacturing. By the analysis in the coverage that followed the announcement, offsite water can double or triple a facility's total footprint, which leaves NVIDIA's design addressing roughly a quarter to a third of the real number.

The water hides in the generation mix. Thermoelectric power plants consume water for their own cooling, and the intensity varies sharply by source: about 1.17 liters per kilowatt-hour for natural gas and 2.2 for coal, against roughly 0.01 for wind and 0.03 for solar. Even hydropower, usually filed under clean, loses about 6.8 liters per kilowatt-hour to reservoir evaporation. U.S. fossil-fuel power plants consume an estimated 2.7 billion gallons of water a day for cooling, and fossil generation still supplies about half of data center electricity, with the IEA projecting gas and coal will cover more than 40 percent of the new power AI demand requires through 2030. As long as those plants feed the building, the water a sealed loop no longer evaporates in the facility yard evaporates instead at the station upstream. We laid out that boundary problem in detail in our analysis of the offsite water data centers do not count.

MIT Sloan's regional analysis called the dynamic a sustainability paradox: a near-zero onsite number can mask the broader upstream cost, and cooling technology alone cannot close AI's water footprint without a parallel move to low-water generation that does not yet exist at the scale the buildout needs. There is a rebound risk on top of it. Making each facility cheaper and lighter on water can accelerate how many get built, which is how an efficiency gain can still raise the industry's aggregate draw. The honest version of NVIDIA's claim is conditional on the grid, and the design cannot control that variable. Our companion news brief on the announcement covers the same boundary in short form.

What It Means for the Cooling Industry

Strip away the sustainability framing and the Rubin design is a standard-setting event. The company that defines the rack is now defining the facility around it, and the reference architecture assumes a sealed warm-water loop with high supply temperatures and outdoor dry coolers. Evaporative cooling as the default heat-rejection method is on a clock. Vendors selling cooling towers, CDUs, dry coolers, manifolds, and quick disconnects will be specifying against this design whether or not NVIDIA is a named customer, the same way NVIDIA's watt roadmap has been writing the cooling industry's business plan generation by generation.

The competitive ground shifts toward components. As facility-level thermal headroom disappears, the reliability of a warm-water site comes down to precision at the cold plate, the manifold, the flow controls, and leak mitigation, which is exactly where 1-ACT argues warm-water operation exposes deviations that cooler water used to hide. That favors suppliers who can hold tight tolerances at volume and reopens the two-phase versus single-phase question, since two-phase cold plates run at lower mass flow and tolerate higher inlet temperatures. Higher-temperature direct-to-chip is now the industry's trajectory rather than a frontier, and the warm return it produces turns waste heat from a disposal problem into a salable output.

NVIDIA has built something the cooling industry should take seriously, and it has described it in terms an operator should read carefully. The onsite water reduction is genuine and worth the engineering. The chiller-free promise holds where the climate cooperates and needs a backstop where it does not. The total-water claim depends on a grid the design does not touch. For the people specifying thermal plants against this reference, the takeaway is to adopt the architecture and keep the boundaries visible, because the asterisks are where the real procurement decisions live.