Nvidia's 45°C Hot-Water Cooling: How the Rubin Platform Aims to Solve AI's Data Center Water Crisis

Nvidia's Rubin platform directly tackles the data center water crisis by enabling cooling with water at 45°C (113°F), a massive leap from the conventional ~6°C supply temperature . This temperature shift allows the cooling loop to reject heat via simple dry coolers instead of energy-intensive chillers and evaporative cooling towers, virtually eliminating on-site facility water consumption in Nvidia's reference design . The approach is a genuine engineering breakthrough, but it does not solve the AI industry's broader resource footprint. Below is a fact-checked breakdown of the technical details, cost savings, industry context, and limitations.

Technical Details

Nvidia's Rubin generation introduces the industry's first 100% liquid-cooled AI infrastructure . Key specifications include:

• 100% closed-loop liquid cooling: Every chip and networking component in the Rubin generation is cooled by liquid in a sealed loop; the system eliminates server fans entirely . Nvidia moved power modules out of the integrated cabinet to create an independent power rack, achieving full liquid cooling coverage for the first time .
• Single-phase direct liquid cooling (DLC): The coolant remains in liquid phase and operates at an inlet temperature of up to 45°C . The return temperature from the rack can reach up to 65°C .
• High power density: Individual Rubin GPUs approach roughly 2,000W TDP, and a Vera Rubin NVL72 rack configuration can reach 140 kW to 200 kW per rack . The warm-water approach makes that density manageable without chillers.
• DSX reference design: Nvidia claims the DSX AI-factory blueprint has "zero water consumption" because it uses no evaporative cooling inside the facility . Ali Heydari, Nvidia's director of data center cooling and infrastructure, stated: "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" .

Potential Cost Savings

The warm-water approach drives savings in three major areas:

• Energy savings: Eliminating mechanical chillers and server fans can reduce cooling energy by an estimated 30–50%, pushing power usage effectiveness (PUE) toward 1.03 . Cooling accounts for roughly 30% to 40% of a typical AI data center's electricity consumption . Rubin also delivers up to 10x energy efficiency over the prior Blackwell architecture .
• Capital expenditure savings: Mechanical chillers and complex heat-rejection infrastructure are no longer needed in most climates, lowering upfront build costs . Nvidia's announcement initially caused shares of major HVAC manufacturers to fall sharply .
• Operational savings: Less water consumption reduces both water bills and regulatory risk in drought-prone regions. Nvidia's earlier Blackwell generation already quoted 300x water efficiency over air-cooled racks ; Rubin improves further.

Industry Context: Microsoft and Amazon

The industry is rapidly pivoting toward warm-water liquid cooling, with major cloud providers aligning with Nvidia's approach:

• Microsoft: Azure publicly stated it is planning for "seamless integration of NVIDIA Vera Rubin NVL72 racks" across its next-gen AI superfactories at current Fairwater sites in Wisconsin and Atlanta and future locations . Custom liquid-cooling heat exchanger units are already in development . In 2024, Microsoft launched a new datacenter design that consumes zero water for cooling, adopting chip-level cooling solutions that recycle water through a closed loop .
• Amazon (AWS): AWS has been independently driving water efficiency, claiming its data centers are 7x more water-efficient than the industry average and is 75% of the way to its "water positive by 2030" goal . AWS has not formally announced Rubin deployment, but its trajectory toward warmer-temperature liquid cooling aligns with Nvidia's approach. Nvidia's press release lists AWS among the adopters of the Rubin platform .
• Broader market signal: In March 2026, within one week, Ecolab acquired CoolIT Systems for $4.75 billion, Trane Technologies bought immersion-cooling specialist LiquidStack, and Nvidia confirmed Rubin's 45°C spec — collectively signaling the industry's rapid pivot to warm-water liquid cooling .

Limitations and Caveats

Despite its promise, Nvidia's warm-water cooling approach faces several important limitations:

• Embodied water and energy are not solved: Eliminating operational water consumption inside the data center does not address the water used to manufacture chips, build facilities, or generate the electricity powering them . TechCrunch notes that while Nvidia's system delivers on its facility-level promise, the broader AI water problem extends far beyond data center cooling .
• Climate dependence: In very hot and/or humid climates, dry coolers alone may not be sufficient; auxiliary adiabatic or chiller systems may still be needed during peak summer days . Outside of favorable climates, the system may need supplemental cooling for roughly 1% of the year .
• Single-phase cold plate limits: Current single-phase cold plates top out at roughly 1,500W heat dissipation per component, which may be insufficient for the highest-power Rubin variants without further engineering advances . Nvidia is reportedly developing microchannel liquid cold plates to meet the thermal demands of the Rubin series .
• Legacy retrofit challenge: The 100% liquid, fanless design is a clean-sheet approach; existing air-cooled or hybrid data centers cannot easily adopt it without major infrastructure overhauls .
• Systemic water crisis persists: Even if every new AI data center adopted Rubin's design, the sheer growth in AI compute demand means absolute energy and resource consumption will keep rising, limiting the net environmental benefit .

In short, Nvidia's 45°C warm-water cooling is a genuine engineering breakthrough that can virtually eliminate on-site water consumption and chiller energy in new-build AI factories. It aligns with Microsoft's and Amazon's own water-efficiency pushes. However, it does not solve the AI industry's broader resource footprint, and its full benefits depend on climate, workload density, and the shift to entirely new data center designs.