Nvidia's Rubin generation of AI servers is the first to run on 100% liquid cooling, with coolant flowing into chips at up to 45 degrees Celsius — 113 degrees Fahrenheit, hotter than the 38-to-40-degree water in a typical hot tub. The counterintuitive choice is the point: by letting the loop run that warm, Nvidia eliminates fans entirely and, in many climates, eliminates mechanical chillers and evaporative water cooling along with them. Nvidia estimates a 50-megawatt hyperscale facility can save more than $4 million a year in cooling energy and water costs by switching to the design.
Cooling has historically consumed up to 40% of a data center's electricity, and industry rule of thumb says every one-degree increase in chiller plant temperature trims cooling energy roughly 4%. Pushing the inlet temperature to 45°C compounds those gains across an entire fleet. Conventional cooling-tower systems burn through about 2.6 million gallons of water per megawatt per year; Rubin's closed loop, paired with outdoor dry coolers, can take that figure to near zero — a 100% reduction in facility water use in favorable geographies.
The full specification is published in the Nvidia DSX AI factory reference design, which lays out how to build and operate the rack, row, and facility stack around the Rubin platform. Every chip and every networking component sits on a liquid-cooled cold plate. The coolant itself is a 75% water and 25% propylene glycol mix, filled once and recirculated for the life of the facility. The data hall no longer needs cold air at all — warm summer air is fine, because nothing in the server depends on it.
Key facts
- 01Nvidia's Rubin platform is the first AI infrastructure to run 100% liquid-cooled, with no fans anywhere in the system.
- 02Coolant enters chips at up to 45°C (113°F) — hotter than a 38–40°C hot tub — and exits at roughly 55°C.
- 03A 50MW hyperscale facility can save over $4M annually on cooling energy and water by switching to the design.
- 04Conventional cooling-tower systems use about 2.6M gallons per megawatt per year; Rubin's closed loop can cut that to zero.
- 05Rack density improves: a server that took six rack units in an air-cooled design now fits in two.
The shift was forced by chip power. Watts per package crossed the line where air cooling stopped being viable, and the ecosystem followed.
“Once the watts per chip crossed a certain level, liquid cooling became mandatory.”— Richard Whitmore, President and CEO of Motivair
Motivair, the advanced cooling division of Schneider Electric, has worked alongside Nvidia's roadmap for nearly a decade and has retooled around the Rubin requirement. Coolant exits the chip at roughly 55°C after picking up the heat load, then moves to coolant distribution units that hand it off to outdoor dry coolers — large radiator coils sitting outside the building. In a place like the Scottish Highlands, the outside air does all the work year-round. In Phoenix, mechanical chillers may still run a handful of days. Nvidia says some climates would need chillers no more than 1% of the year.
The zero-water claim is the headline number. Conventional hyperscale cooling sprays water across evaporative towers to shed heat; the closed loop in Rubin facilities never touches an evaporative stage.
“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.”— Ali Heydari, Director of data center cooling and infrastructure at NVIDIA
The engineering work to get there was not trivial. Prior generations of liquid-cooled servers were hybrids — GPUs and CPUs got cold plates, but memory, power components, and switches still ran air-cooled with finned heat sinks. Going to 100% liquid forced Nvidia's thermal team to redesign cooling for every part of the board, routing a single inlet and outlet through multiple high-power chips on each tray. The visible result is a sealed front panel where air-cooled servers have perforated bezels, and a density jump that fits a system previously occupying six rack units into two.
Noise drops too. Traditional data halls run at 85 decibels or louder from fan walls, loud enough to require ear protection. Rubin rows have no fans at all. The choreography of hot aisles and cold aisles, the elaborate air-handling architecture that defined a generation of facility design, becomes irrelevant. Operators recover floor space, electrical capacity that was going to fans, and the option to recover waste heat for nearby commercial or residential buildings.
The caveats are geographic and behavioral. The economics of chiller-less operation depend on outside air temperatures, and the water-elimination figure assumes operators actually deploy dry coolers rather than defaulting to cooling towers out of habit. Retrofits into legacy facilities designed around raised-floor air distribution are harder than new builds. And the higher rack density Rubin enables concentrates electrical load — a power-delivery problem that doesn't disappear just because the cooling problem got easier.
For the AI infrastructure market, the Rubin cooling spec sets a baseline that every hyperscaler buying the platform now has to meet, which pulls the entire data-center supply chain toward closed-loop liquid by default. The companies that have been quietly building dry-cooler capacity and direct-to-chip plumbing — Schneider's Motivair among them — are about to see demand they spent a decade preparing for. And the policy debate around AI data centers, which has fixated on water draws from local aquifers and reservoirs, loses its most-cited number when the reference design ships with a zero next to gallons-per-year.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
