NVIDIA has officially unveiled a major structural overhaul for AI data centers, debuting a 100% liquid-cooled reference design that promises up to a 100% reduction in facility water usage in favorable climates.

Announced via the company’s new NVIDIA DSX AI factory reference blueprint, the design completely removes internal server fans and traditional evaporative cooling towers. Instead, it relies on a specialized high-temperature, closed-loop liquid system to manage the extreme thermal loads of its next-generation Vera Rubin AI architecture.

1. The 45°C “Hotter Than a Hot Tub” Breakthrough

The core engineering innovation hinges on raising the temperature threshold of the cooling fluid. Traditional air-cooled or basic liquid data centers require chilled air or cold water to prevent silicon from throttling. NVIDIA’s new architecture flips this approach entirely:

  • The Fluid Mixture: The closed loop pumps a chemical recipe consisting of 75% water and 25% propylene glycol (similar to automotive antifreeze) through custom-engineered cold plates sitting directly on top of the processors.
  • Operating Temperature: The system runs this liquid at an inlet temperature of up to 45°C (113°F)—which is significantly hotter than the average commercial hot tub (usually 38°C to 40°C).
  • Direct-to-Chip Thermal Transfer: Because silicon processors operate at incredibly high internal temperatures, the 45°C fluid enters the chip, absorbs the extreme localized heat, and exits at roughly 55°C (131°F) without damaging the hardware.

2. Going Chiller-Less: Wiping Out Facility Water Consumption

By validating that its servers can run reliably on 45°C fluid, NVIDIA has effectively eliminated the need for massive mechanical chillers and evaporative cooling towers, which typically vent millions of gallons of water into the atmosphere.

Conventional Cooling ──► Evaporative Towers ──► Consumes ~2.6M Gallons / MW / Year
NVIDIA DSX Loop      ──► Outdoor Dry Coolers ──► Near ZERO Water Consumption (Closed Loop)

In compatible climates, the heated 55°C return fluid can be cooled down back to 45°C entirely by passing it through outdoor dry coolers using ambient air. Ali Heydari, NVIDIA’s Director of Data Center Cooling and Infrastructure, noted that this closed-loop design slashes a standard facility’s cooling water requirements from roughly 2.6 million gallons per megawatt (MW) per year to near zero. Mechanical chillers will only turn on for a fraction of the year (roughly 1%) during extreme, peak-summer heatwaves in specific geographic regions.

3. Shaking Up Data Center Footprints

Shifting to a 100% liquid-cooled blueprint completely changes the physical layout and environmental footprint of modern AI infrastructure:

  • No Fans, No Noise: By relying entirely on liquid to draw heat directly from the source, the design eliminates the thousands of high-decibel chassis fans that define legacy data centers.
  • Dense Clustering: Removing fans allows server blades to be stacked tightly against one another without requiring traditional “hot-aisle/cold-aisle” floor layouts, significantly shrinking the physical land footprint needed to deploy massive clusters.
  • Waste Heat Recovery: Because the loop expels fluid at a predictable, high-temperature 55°C, data center operators can capture this residual heat and redirect it into local municipal grids to warm nearby commercial or residential buildings, transforming a major energy waste product into a community utility.

4. The Climate Bottleneck

While NVIDIA’s Chief Sustainability Officer, Josh Parker, declared that “the water consumption challenge for data centers is largely solved” ahead of London Climate Week, industry analysts note a few distinct geographic hurdles:

The system achieves its maximum efficiency via “free cooling” from ambient outdoor air. In naturally cooler regions, the system runs completely dry. However, in regions experiencing extreme heat and high humidity (such as Nevada, parts of Texas, or regions across Southeast Asia), dry coolers struggle to shed heat efficiently into the surrounding air when outdoor temperatures approach or exceed the 45°C mark. In these hot zones, operators will still have to fall back on secondary mechanical refrigeration, meaning real-world water savings will heavily depend on where the AI infrastructure is physically built.

Furthermore, environmental groups point out that while the design drastically mitigates on-site water consumption, the massive electricity demands of running million-GPU clusters still carry a heavy, indirect water footprint at the regional power plants generating the grid electricity.

Because NVIDIA is embedding this liquid-cooling standard directly into its upcoming Rubin platform architecture, the design sets an immediate, aggressive compliance target for hyperscalers like Microsoft, Google, and AWS as they race to balance surging AI computing demand with tightening environmental regulations.