A 45°C coolant ceiling in NVIDIA's new Vera Rubin reference design lets AI factories skip evaporative cooling towers; per chip efficiency is improving, but whether total AI water use falls or rises still depends on how much compute the industry builds.
AI data centers are quietly raising the temperature of the liquid running through their servers, to 45°C, or 113°F, hotter than a typical hot tub. That sounds like a problem. It is actually a water-saving lever. Warmer coolant widens the operating window where a facility can cool itself with air and dry coolers instead of evaporative cooling towers, the component that drains local water supplies to keep servers cold. NVIDIA's Vera Rubin DSX AI Factory Reference Design, released alongside an Omniverse digital twin blueprint, is built around exactly that bet.
The mechanism is unglamorous. Most data centers run chilled water loops at temperatures cold enough to require chillers, the refrigeration compressors that hand heat off to an evaporative cooling tower in hot or humid weather. The tower works by spraying water through hot air; some evaporates and the rest falls back as cooled water. In dry, hot climates it is the only realistic way to dump gigawatts of heat, and it can drain millions of gallons from local water supplies each year per facility. Raise the coolant temperature and the chiller drops out: a dry cooler, essentially a giant car radiator, can dump heat to ambient air on its own, with no evaporation and no tower.
That is what 45°C buys you. The NVIDIA DSX product page describes the reference design as covering the full AI factory stack, from power and networking to racks and cooling, paired with an Omniverse digital twin so operators can simulate facility behavior before pouring concrete. Ali Heydar, NVIDIA's director of data center cooling and infrastructure, told the company's blog that the DSX reference design for AI factories has "zero water consumption" and eliminates "massive amounts of power usage and pretty much all water usage."
The Rubin generation is the first NVIDIA platform with 100% liquid cooling, every chip and networking component cooled entirely by liquid in a closed loop with no fans anywhere in the system. Historically, cooling has accounted for up to 40% of a data center's electricity consumption, making it one of the most significant areas where efficiency improvements drive down operational expenses and energy demands.
Two caveats matter here. First, "zero" is NVIDIA's own characterization of a reference design, a blueprint for how to build a Rubin AI factory, not a measurement of any specific deployed site. The same source notes that chillers still kick in for roughly 1% of the year when needed in some climates. Second, 45°C is a DSX reference-design choice. Most of the existing industry still runs cooler. NVIDIA is betting that the rest of the cooling stack will catch up; whether it does is an open engineering question.
The aggregate math is where the story stops being a clean win. Per-chip efficiency is improving with each NVIDIA generation, and the Rubin platform announcement introduces six new chips and a unified supercomputer, all liquid-cooled. But total AI compute demand is still climbing faster. A reader asking "is AI getting cleaner?" has to hold two sentences at once: yes, per watt; probably not, in absolute terms. The 45°C shift makes the first sentence more true without resolving the second.
The ecosystem is at least committing to the bet. CRN reports nine hardware and cloud companies publicly building Vera Rubin systems, from chip packaging and switching to rack-level integration. That gives the reference design a chance to be tested outside NVIDIA's own modeling. If hyperscalers adopt the dry-cooler pattern in dry climates and reserve evaporative cooling for genuinely hot-and-humid sites, the water savings will be real. If they keep building at the same pace, even cleaner per-chip AI can still draw more water than last year's fleet.
Watch the dry-cooler tax. Cooling cost scales with the temperature gap between the chip and the air; raise the chiller setpoint by one degree and a hyperscale site saves roughly 4% of cooling energy costs. The 45°C ceiling is not the limit of physics. It is the limit of the surrounding equipment NVIDIA is willing to bet on. That equipment is what the next two years of procurement decisions will actually test.