NVIDIA faces no greater test than powering the explosive growth of artificial intelligence. Its latest platform, Vera Rubin, doesn’t just pack more performance into silicon. It forces a complete rethink of the facilities that will run it. The shift centers on liquid cooling taken to a new extreme. Servers run with coolant entering at 45 degrees Celsius. Hotter than a hot tub. Yet this deliberate choice slashes water consumption and trims energy bills in ways air-based systems never could.
The numbers tell a stark story. Traditional cooling towers in data centers consume roughly 2.6 million gallons of water per megawatt per year. Rubin-based designs drop that figure to near zero in suitable climates. A single 50-megawatt facility stands to save more than $4 million annually on cooling costs alone. Cooling has long eaten up to 40 percent of a data center’s electricity. That burden shrinks dramatically here.
Rubin demands full liquid cooling across every component.
Direct-to-chip cold plates sit on processors. Coolant, a mix of 75 percent water and 25 percent propylene glycol, flows through closed loops. No fans spin inside the racks. No perforated front panels admit air. The entire system seals tight. Heat exits at around 55 degrees Celsius before reaching outdoor dry coolers that look like oversized radiators. In many locations these handle the load year-round without mechanical chillers. Only in extreme conditions might operators tap chillers for perhaps 1 percent of the year.
Jensen Huang drove the point home at CES 2026. “With 45 degrees C, no water chillers are necessary for data centers,” he said. “We are basically cooling this supercomputer with hot water.” The approach builds on Blackwell but takes the concept further. Every chip and networking element runs liquid cooled. Racks achieve higher density. What once occupied six units now fits in two. Serviceability improves. Noise drops below the 85 decibels common in air-cooled halls.
The original Verge report captured early details of this transition. NVIDIA’s own analysis goes deeper. Ali Heydari, director of data center cooling and infrastructure at NVIDIA, described the reference design. “The NVIDIA DSX reference design for AI factories has zero water consumption — we have eliminated massive amounts of power usage and pretty much all water usage,” he explained. The closed loop fills once and runs indefinitely in the right conditions.
But the change ripples far beyond NVIDIA’s labs. Data center operators must redesign halls for in-row coolant distribution units and higher flow rates. Pipes replace some traditional infrastructure. Power delivery evolves too, with liquid-cooled busbars appearing in newer racks. Hyperscalers such as Microsoft have signaled readiness. Their Azure infrastructure already incorporates upgraded thermal headroom and heat exchanger units tuned for Rubin densities.
Supermicro moved quickly. The server maker announced expanded manufacturing capacity for liquid-cooled solutions supporting Vera Rubin NVL72 and HGX Rubin NVL8 systems. Its designs feature in-row CDUs that support warm-water operation while minimizing both energy and water use. Dell followed with plans for high-density servers aimed at the same architecture. These vendors don’t simply bolt on cooling. They rebuild trays around board-to-board connectors that eliminate internal hoses and cables.
Market reactions proved immediate and brutal. When Huang spoke at CES, shares in cooling equipment providers tumbled. Reuters documented the sell-off. Johnson Controls fell more than 7 percent. Trane dropped over 5 percent. Carrier Global declined 1.1 percent. Investors feared shrinking demand for chillers and traditional HVAC. Yet analysts spotted nuance. nVent Electric, with stronger exposure to liquid cooling components, held steady or edged higher. Vertiv, which offers both air and liquid solutions, saw only modest pressure.
Richard Whitmore, president and CEO of Motivair, part of Schneider Electric, put the shift in perspective. “Once the watts per chip crossed a certain level, liquid cooling became mandatory,” he said. “In the right geographic location, with the right system design, you don’t need any refrigeration equipment. You can just put big radiator coils outside and use the air temperature for all your cooling. It’s incredibly efficient.”
Power density climbs relentlessly. Rubin racks head toward 300 kilowatts or more in coming generations. Some projections eye 600 kilowatts within two years. Air cooling hits physical limits well before that. Liquid cooling doesn’t just manage heat. It enables continued scaling. Processors stay within validated temperature ranges even as they deliver more compute per watt. Each generation extracts greater efficiency from the same energy envelope.
Yet challenges remain. Not every data center sits in a climate that favors dry coolers. Retrofitting existing facilities carries cost. New builds must plan for higher amperage busways, revised rack geometries and sophisticated coolant distribution. Water quality still matters in the closed loops. Propylene glycol mixtures require monitoring. And while facility water use plummets, the broader electricity demand for AI training keeps rising.
Recent coverage highlights these tensions. A Bloomberg analysis from June 2026 maps the race to redesign facilities amid surging power needs. Liquid cooling delivers roughly 15 percent better energy efficiency in studied cases, according to Nvidia and Vertiv data. Emissions tied to purchased power fall about 10 percent. Those gains compound at hyperscale.
Frore Systems demonstrated advanced cold plates at CES 2026 capable of handling Rubin’s projected 1950-watt GPUs. Their LiquidJet design improved thermal performance by more than 50 percent on Rubin compared with prior solutions while cutting weight. Such component-level advances will determine which suppliers capture share in the new stack.
The bigger picture extends past efficiency metrics. Waste heat from these systems reaches temperatures high enough for reuse. Operators talk of piping it to nearby buildings for district heating. AI factories could transform from energy consumers into community assets. In favorable locations they consume almost no water and return heat instead of expelling it wastefully.
NVIDIA positions the Vera Rubin platform, including its Vera CPU rack for agentic workloads, as the foundation for next-generation AI factories. Availability begins in the second half of 2026. Cloud providers and enterprise operators have little choice but to adapt. The platform’s liquid-cooled MGX architecture sets a new baseline. Those who delay risk falling behind on both performance and operating costs.
Critics once warned that AI would devour water resources and strain power grids. Early data from Rubin deployments suggest a different outcome. Water usage in U.S. data centers already represents just 0.2 percent of daily consumption and continues to decline with these methods. Energy intensity per unit of compute falls. The infrastructure evolves to match the ambition of the models it trains.
Still, the transition carries real costs. Billions will flow into new cooling infrastructure, electrical upgrades and facility redesigns. Supply chains for CDUs, cold plates and high-flow pumps must scale rapidly. Construction timelines for gigawatt-scale AI sites stretch years. NVIDIA’s design choices narrow some of those bottlenecks but cannot eliminate them.
So the industry watches closely. Rubin doesn’t merely cool chips. It redefines what a modern AI data center looks like. Cleaner. Denser. Quieter. Far less thirsty. Whether the broader ecosystem can build enough of them fast enough will shape AI progress for years ahead. The hot water gamble appears to be paying off.
WebProNews is an iEntry Publication