NVIDIA Ruben – NVIDIA says its new Ruben AI server design is water-cooled in a closed loop that never has to run below 45°C, avoiding evaporative cooling. The company positions the approach as a way to cut cooling costs—potentially to the tune of $4 million a year for a 50MW
It’s the kind of debate NVIDIA doesn’t like hearing while AI headlines spiral: job loss, bigger power bills, and—especially—freshwater use. In the company’s latest pitch, the water problem is supposed to disappear.
NVIDIA’s Ruben AI server architecture is described as all-water cooled. with coolant kept in a closed loop that “won’t need any evaporative cooling.” The reason is built into the temperature target. NVIDIA says the coolant “can stay in a closed loop,” and “never needs to be cooled below 45 C, or 113 F.”.
That temperature range is meant to be familiar to anyone who has assembled a water-cooled PC or a PlayStation-style cooling setup: a glycol-water mix. water blocks attached to hot components. and a radiator that rejects heat to the surrounding environment. NVIDIA doesn’t say whether the Ruben servers include RGB lighting, but the practical distinction is straightforward. In these racks. “everything is on a waterblock. ” so heat from any chip generating thermal load on the motherboard is carried into the same cooling water.
NVIDIA’s numbers sketch the operating rhythm: the cold side of the loop sits at about 45 C, while coolant leaving the racks runs at about 55 C. The piece giving the details translates that to 113 F to 131 F for readers keeping their mental gauges in Fahrenheit.
The key tradeoff is how much cooling the system actually needs. NVIDIA argues the coolant just has to be cooler than the chips, and that the required drop is modest enough that there’s no need for the “evaporative chillers” that have made water usage a flashpoint for AI data centers.
Once the design philosophy is set, the implications follow. With heat removed by dry heat exchangers—described as “ambient-temperature air running over dry heat exchangers” and also called “big honkin’ radiators”—the setup doesn’t require losing water to evaporation. And because components sit on waterblocks. the servers don’t require cooling air either; the server farms only need air conditioning “to the degree required to make them comfortable to work in.”.
But if this sounds like a sudden act of environmental conscience, the language around the motivation points elsewhere. The press material frames the shift as a direct cost move. Cooling costs money, and running these systems hot saves it. The document puts a figure on the potential savings: “four mil US a year for a 50 MW hyperscaler.”.
That raises the question engineers worry about when you change thermal conditions: would hotter fluid shorten the life of the NPUs doing the work?. The same material acknowledges the possibility—suggesting the thermal regime “could limit the lifetime of the hard-working NPUs”—then immediately undercuts it with a different timetable. Since the NPUs are expected to be obsolete in a few years. it says the lifetime concern “isn’t likely a big concern. especially not for NVIDIA.”.
Water-based cooling in computing isn’t new either. The report mentions earlier examples, including coffee used to cool computers, and it notes that water cooling in data centers has been on the radar for years.
For now, the Ruben concept signals a coming fight over what “sustainable” cooling can mean for AI infrastructure: not just how much water is used, but whether it’s avoided by design—and whether the design changes because of water, or because of money.
NVIDIA Ruben AI servers water cooling closed loop coolant evaporative cooling data center cooling hyperscaler NPUs 45C 55C