NVIDIA’s Liquid Cooling Roadmap: From Blackwell GB300 to Rubin Ultra 1MW Racks

In the high-stakes world of AI infrastructure, the narrative has shifted. For years, the industry obsessed over TFLOPS and HBM capacity. Today, the conversation has moved to a more primal physical constraint: Heat. As NVIDIA transitions from the Blackwell architecture (GB300) to the upcoming Rubin and Rubin Ultra, we are witnessing more than just a chip upgrade. We are seeing the total "liquification" of the data center.

 

 

NVIDIA GB300: The "Hybrid" Bridge to Liquid Cooling

The NVIDIA Blackwell GB300 serves as the industry’s training wheels for the liquid-cooled era. It utilizes a Single-Phase Cold Plate system, where a coolant (typically deionized water or propylene glycol) circulates through micro-channels to pull heat away from the GPU and CPU.

 

The "L2 Autonomy" Analogy: Much like Level 2 self-driving cars that still require a steering wheel, the GB300 is a "hybrid." It introduces liquid cooling to the hotspots but maintains legacy air-cooling logic for secondary components.

 

Strategic Purpose: The GB300 allows the global supply chain—ODMs, CDU (Cooling Distribution Unit) manufacturers, and data center operators—to normalize liquid infrastructure without a total architectural shock.

 

 

 

Vera Rubin: The Era of 100% Liquid Standardization

With the Vera Rubin platform, the TDP (Thermal Design Power) of a single GPU is rumored to reach a staggering 1500W to 2000W. At this density, air cooling is no longer just "inefficient"—it is physically impossible.

 

1. The Death of the Fan

Vera Rubin represents the definitive pivot to 100% Liquid Cooling. Flagship rack systems will likely move toward fanless, cable-less, and modular designs. In these high-density AI clusters, air is no longer the medium of choice; liquid is the primary conductor.

 

2. 45°C Warm Water Cooling & Sustainability

Perhaps the most significant shift is the move toward Warm Water Cooling. By utilizing 45°C water, data centers can bypass energy-hungry chillers and rely on "free cooling" (evaporative or dry coolers). This redefines the TCO (Total Cost of Ownership) model from "refrigeration" to "heat transport."

 

3. Liquid Metal TIM

To bridge the gap between the silicon die and the cold plate, Liquid Metal Thermal Interface Materials (TIM) are becoming the standard. With Rubin, liquid metal isn't a premium feature—it’s a prerequisite to prevent thermal throttling at peak frequencies.

 

 

 

 

The Innovation: Micro-channel "Skin-Tight" Cooling

The real technological leap in the Rubin era is the Micro-channel Cold Plate Cover.

Traditional cold plates act like a "roof" over the chip. Micro-channels, however, bring the coolant directly to the "doorstep" of the silicon.

Manufacturing Shift: Creating these ultra-fine, high-precision fluid paths requires advanced Metal 3D Printing (Copper Additive Manufacturing).

Precision Targeting: Unlike traditional CNC-machined plates, 3D-printed versions can vary channel density based on the chip's "heat map"—denser channels for high-heat areas and sparser for lower ones, acting like a GPS for the coolant.

 

 

Rubin Ultra: Approaching the 1-Megawatt Rack

While Vera Rubin sets the standard, Rubin Ultra pushes into the realm of extreme systems engineering. We are now entering the era of the 1-Megawatt (1MW) Rack.

Systemic Power Density: A single rack will consume as much power as a small neighborhood. This forces a rethink of the entire facility, from High-Voltage Direct Current (HVDC) power delivery to reinforced floor loading and specialized fire suppression.

The New Infrastructure: Liquid cooling is no longer a server "accessory"; it is a foundational layer of the data center, as vital as the electricity itself.

 

 

Conclusion: Stability is the New Performance

The roadmap from Blackwell to Rubin reveals a fundamental truth: The second half of the AI race isn't about who has the largest chip—it's about who has the most stable system.

NVIDIA isn't just selling "hotter GPUs"; they are selling an incredibly difficult-to-replicate Integrated System. The value in the future supply chain won't just reside in chip design, but in the mastery of manifolds, leak-proof quick disconnects, and the orchestration of liquid-cooled megaclusters.

The age of air cooling is ending faster than anyone predicted. The Liquid-Driven Compute Factory has arrived.

 

 

Point:

If you are in need of liquid cooling solutions for data centers, consider Lian Li Liquid Cooling. We offer a comprehensive, full-chain liquid cooling solution backed by experience from over 5 million deployed units and projects exceeding 10GW in total power capacity. 
Our products are deeply optimized for compatibility with NVIDIA GB200, GB300, B200, B300, H100, H200, 5090, and 4090. 
We enable a PUE of less than 1.15, thereby reducing the total cost of ownership (TCO) for data centers, and hold authoritative certifications such as RoHS, CE, and UL.

 

 

FAQ About  NVIDIA GB300 

What are the key performance specifications of the NVIDIA GB300 compared to previous generations? 
The NVIDIA GB300 (Blackwell Ultra) represents a significant leap in AI reasoning performance, delivering up to 1.5x higher dense FP4 throughput than the standard GB200. With 288GB of HBM3e memory and a single-package design that integrates the Grace CPU with a Blackwell Ultra GPU, the GB300 is specifically optimized for frontier AI models exceeding one trillion parameters. Its architecture focuses on reducing "time-to-token" and increasing context window capacity for complex agentic workloads.  

How does the NVIDIA GB300 NVL72 architecture handle extreme power density and cooling? 
The NVIDIA GB300 NVL72 is a rack-scale system that unifies 72 Blackwell Ultra GPUs into a single massive GPU via NVLink 5. To manage a rack power density ranging from 132kW to 180kW, it utilizes an advanced direct liquid cooling (DLC) system. This setup typically employs a 90/10 split—where 90% of heat is captured via liquid cold plates (using 45°C warm water) and the remaining 10% is handled by air—allowing the system to operate at hyperscale efficiency without traditional high-energy chillers.

 

What is the estimated NVIDIA GB300 Price and its impact on Data Center TCO?
While official NVIDIA GB300 Price points vary by configuration and volume, market estimates for a full NVL72 rack system range from $3 million to $4 million, depending on integrated networking and storage options. Despite the higher upfront CAPEX compared to the Hopper generation, the GB300 significantly lowers the Total Cost of Ownership (TCO) by providing 5x higher performance per megawatt. This energy efficiency, combined with liquid cooling's ability to maximize compute density, makes it the most cost-effective solution for large-scale AI factories.