Hot tubs sit at around 38 to 40 degrees Celsius — warm enough that most people can only soak for about 15 minutes before they need to climb out. NVIDIA's newest AI servers run their cooling liquid hotter than that: up to 45 degrees Celsius, or 113 degrees Fahrenheit. It sounds backwards. You would assume that hotter coolant means a hotter, more stressed machine. In fact, that higher temperature limit is precisely what makes these systems more energy efficient, and it is one of the more consequential design decisions in recent data center history.
NVIDIA describes its Rubin-generation infrastructure as the first to achieve 100% liquid cooling — every chip and every networking component cooled entirely by liquid in a closed loop, with no fans anywhere in the system. The approach is documented in the company's DSX AI factory reference design, a blueprint for how to build and operate the full infrastructure stack around these servers. To understand why running coolant at bathwater-plus temperatures is a feature rather than a flaw, it helps to start with a misconception the industry has carried for decades.
For a long time, a cold data center was treated as a well-run data center. If you walked onto the floor and it did not feel like a walk-in freezer, people assumed something was wrong. That instinct made sense in an air-cooled world, where you were trying to flood an entire room with chilled air and hope enough of it reached the components that needed it.
But silicon does not actually need a freezer. Modern processors generate enormous internal heat and tolerate far warmer surroundings than that instinct suggests. In a fully liquid-cooled Rubin system, coolant enters the chip at 45 degrees Celsius and exits at roughly 55 degrees, having absorbed the heat load directly off the chip surface. Crucially, performance does not degrade. The liquid-cooled cold plates sitting on top of each processor keep the device within its validated operating limits even though the coolant arriving at the rack is hot enough to be uncomfortable to touch. The chip stays happy; the room does not have to be cold.
Once you accept that the silicon is fine at these temperatures, the entire economics of cooling change. The room temperature becomes flexible. Warm summer air is acceptable, because nothing in the server depends on cool air anymore. The liquid does all the work.
Cooling has historically been one of the largest line items in a data center's power budget — accounting for up to 40% of total electricity consumption. That makes it the single biggest target for efficiency gains, because every watt you do not spend on cooling is a watt you can spend on compute, or simply not draw from the grid at all.
The leverage comes from a well-known rule of thumb: raising the temperature a chiller plant has to hit by just one degree cuts cooling energy costs by roughly 4%. The reason NVIDIA's 45-degree figure matters is that it pushes past the point where you need mechanical chillers at all in many climates. Instead of running power-hungry compressors to manufacture cold water, a facility can reject heat using outdoor dry coolers — essentially large radiator coils positioned outside the building — that rely on ambient air. For much of the year in a temperate location, the outside air is cool enough to do the job on its own.
NVIDIA puts hard numbers on this. A 50-megawatt hyperscale facility, it estimates, can save over $4 million annually in cooling-related energy and water costs by moving to this liquid-cooled infrastructure. At CES earlier in the year, NVIDIA's CEO went further, suggesting the broader shift could save around 6% of the world's total data center power — a meaningful figure in a market where access to electricity has become a genuine constraint on how fast AI can grow.
Energy is only half of it. Traditional cooling-tower-based systems lose enormous amounts of water to evaporation — roughly 2.6 million gallons per megawatt per year. As AI data centers have multiplied, that water consumption has become a flashpoint in local planning disputes, with dozens of proposed builds delayed or blocked over resource concerns.
A closed liquid loop largely sidesteps this. The system is filled once and recirculated for the life of the facility, so no new water is consumed to cool the chips. NVIDIA's reference design targets zero water consumption outside of perhaps 1% of the year, when a small number of climates might still need chillers on the hottest days. In favorable geographies, the design can cut facility cooling water use to near zero — up to a 100% reduction. For operators trying to build in water-stressed regions, that is not a minor optimization; it can be the difference between getting a permit and not.
The shift to full liquid cooling changes what the hardware actually looks and sounds like. Walk into a traditional data center and two things hit you immediately: the noise, where cooling fans push total sound levels to 85 decibels or more — loud enough to require ear protection — and the elaborate choreography of hot aisles and cold aisles engineered to steer chilled air across components.
Rubin removes both. With no fans, the roar goes away. The careful aisle management goes away too, because there is no air path to protect. The coolant — a mix of 75% water and 25% propylene glycol — flows through cold plates pressed directly onto the processors, pulling heat out at the source rather than trying to chase it through the air.
There are visible consequences. Rubin servers have clean, sealed front panels where air-cooled servers have perforated bezels designed to let air through. And because liquid cooling is so much more effective at moving heat, the servers can pack far tighter: a system that previously occupied six rack units now fits into two. That is more compute in less space, generating less noise, in a smaller building footprint.
Liquid cooling itself is not new, but earlier liquid-cooled servers were hybrids. The GPUs and CPUs got cold plates, while the rest of the board — memory, networking, power components — kept relying on finned heat sinks and moving air. That left fans in the system and air paths to manage, which capped how far the efficiency gains could go.
Getting to genuinely 100% liquid cooling meant redesigning how every remaining component sheds heat. NVIDIA's thermal team reworked the cooling loops so that liquid is routed to multiple high-power chips on a board through a single inlet and outlet, producing a cleaner tray-level architecture rather than a tangle of separate loops. It is the unglamorous plumbing work that had to be solved before the headline temperature numbers could mean anything in practice.
Another benefit falls out of capturing all that heat in liquid: it becomes recoverable. Residual heat from an AI factory can potentially be piped to warm nearby commercial or residential buildings, turning a waste stream into a usable resource — though that depends heavily on having the right buildings nearby and the infrastructure to connect them.
None of this is unconditional. The chiller-free promise lives and dies on where the building sits. A data center in the Scottish Highlands and one in Phoenix, Arizona face completely different outdoor realities. In cool climates you can lean almost entirely on dry coolers; in hot ones, you still need chillers for the days when ambient air exceeds what the loop can reject. Even there, 45-degree coolant helps — it moves operators much closer to the chiller-less ideal, where the compressors might fire only a handful of days a year instead of running constantly.
It is also worth keeping the announcement in proportion. Several cooling specialists have pointed out that this is less an overnight breakthrough than the next step on a path the industry was already travelling. Blackwell-generation systems already ran inlet temperatures around 40 degrees, and some vendors had been designing for the mid-40s before Rubin was announced. Liquid cooling memory and certain other components remains expensive and technically awkward, which is part of why chillers will not disappear entirely. The direction of travel — higher operating temperatures, more liquid, fewer chillers — is real, but it is a continuation rather than a clean break.
The reason any of this gets attention is the trajectory of AI compute. Demand is growing faster than almost any other category of infrastructure investment, and power has become the binding constraint. If cooling efficiency stood still, the energy cost of running AI at scale would climb in lockstep with the hardware — and the public and regulatory backlash over data center power and water would climb with it. Pushing coolant hotter than a hot tub is, oddly, one of the most practical levers the industry has to keep that curve from running away. Hotter for the chips, cooler for the planet.
If you want to read more about how it works, you can read the full breakdown on the official nvidia link.
