Nvidia Rubin NVL72 Cost Breakdown:Liquid Cooling Are Reshaping AI Infrastructure Economics

The global AI industry is entering a new phase of infrastructure expansion. While previous generations of AI clusters were primarily defined by GPU performance, the latest generation of systems reveals a different reality: memory, storage, networking, power delivery, and liquid cooling are becoming equally important cost drivers.

According to recent industry estimates, Nvidia's Vera Rubin NVL72 rack may reach a total system cost of approximately $9.1 million per rack, significantly exceeding earlier market expectations. More importantly, the increase is not driven solely by GPUs. Rising prices of HBM4, LPDDR5X, and enterprise NAND storage are fundamentally changing the economics of AI infrastructure.

For hyperscalers, AI cloud providers, and data center operators, the industry is rapidly transitioning from a GPU-driven market to a memory-driven market.

What Is Nvidia Vera Rubin NVL72?

Vera Rubin NVL72 represents Nvidia's next-generation AI training and inference platform designed for ultra-large-scale AI workloads.

Key Specifications

Compared with previous Blackwell-based systems delivering roughly 720 PFLOPS per rack, Rubin offers approximately 3.5× higher AI performance density, pushing rack-level power consumption and thermal loads to unprecedented levels.

Why Has the Cost Increased So Dramatically?

The primary reason is that AI infrastructure is no longer just a collection of GPUs.

Modern AI racks require:

Massive HBM memory capacity

High-speed CPU memory

Large-scale local storage

Advanced networking fabrics

High-voltage power systems

Rack-scale liquid cooling

Every subsystem is becoming more expensive simultaneously.

Rubin NVL72 Rack Cost Breakdown

Total Estimated Rack Cost: $9.1 Million

The data demonstrates that GPUs remain critical, but memory and storage are rapidly becoming the second-largest contributor to total rack cost.

HBM4 Is Becoming the New Bottleneck

HBM Cost Explosion

High Bandwidth Memory (HBM4) has emerged as one of the most constrained components in the AI supply chain.

The projected increase reflects growing demand from AI training clusters and limited production capacity among leading suppliers.

Why HBM Matters

HBM directly impacts:

Model training throughput

GPU utilization efficiency

Large language model scaling

Inference performance

Without sufficient HBM supply, even advanced GPUs cannot reach their full performance potential.

LPDDR5X and NAND Are Also Driving Costs Higher

HBM is not the only contributor.

LPDDR5X Memory

CPU-side memory costs have increased significantly due to:

AI server demand growth

Supply prioritization toward enterprise customers

Premium pricing for high-capacity modules

Estimated Rubin rack LPDDR5X cost:

$802,000 per rack

Enterprise NAND Storage

AI infrastructure increasingly requires:

High-speed checkpoint storage

Local dataset caching

Distributed training storage

Vector database workloads

As a result, enterprise SSD demand continues to rise.

Estimated direct-attached storage cost:

$1.282 million per rack

Networking Costs Have Doubled

AI clusters rely on massive east-west traffic between GPUs.

Networking Cost Breakdown

Compared with previous-generation AI racks, networking costs have roughly doubled.

Without high-speed interconnects, large-scale AI training clusters cannot effectively scale.

The Hidden Cost Driver: Liquid Cooling

As rack densities exceed 200 kW, traditional air cooling becomes increasingly difficult to deploy efficiently.

Why Rubin Requires Liquid Cooling

220 kW per rack creates thermal loads that conventional cooling systems struggle to handle.

Benefits of liquid cooling include:

Higher cooling efficiency

Lower fan power consumption

Improved rack density

Better GPU performance stability

Reduced data center PUE

Cooling Cost Comparison

The nearly 3× increase reflects the industry's transition toward rack-scale liquid cooling infrastructure.

What Does a 1GW AI Cluster Cost?

The impact becomes even more dramatic at hyperscale deployment levels.

1GW AI Infrastructure Economics

For comparison:

The trend is clear: AI infrastructure costs continue to rise despite improvements in performance.

Annual Operating Costs Remain Enormous

Capital expenditure is only part of the equation.

Estimated Annual Operating Expenses

Depreciation remains the largest ongoing expense for hyperscale AI operators.

Why Liquid Cooling Will Become Standard Infrastructure

The Rubin generation highlights a broader trend across the industry.

As rack densities move toward:

150 kW/rack

200 kW/rack

300 kW/rack and beyond

Liquid cooling is no longer optional.

Future AI facilities will increasingly rely on:

Cold plate liquid cooling

Coolant Distribution Units (CDUs)

Rear-door heat exchangers

Direct liquid cooling loops

High-voltage power architectures

This transition creates substantial opportunities across the liquid cooling supply chain.

Future Outlook: The Era of Memory-Priced AI Infrastructure

The AI industry is entering a period where infrastructure costs are determined by more than GPUs.

Future competitiveness will depend on:

Supply Chain Control

Securing HBM, LPDDR, NAND, and networking components.

Energy Efficiency

Reducing power consumption through advanced cooling technologies.

Deployment Speed

Accelerating delivery through modular and prefabricated data center designs.

Infrastructure Optimization

Maximizing performance per watt and performance per dollar.

Organizations that can balance these factors will be best positioned to scale AI operations profitably.

Conclusion

The Vera Rubin NVL72 platform demonstrates that AI infrastructure is becoming increasingly complex and capital intensive. While GPUs remain at the center of AI computing, memory, storage, networking, power delivery, and liquid cooling are now equally critical components of the overall cost structure.

As AI clusters scale toward gigawatt-level deployments, infrastructure efficiency—not just compute performance—will become the defining competitive advantage.

For data center operators, cloud providers, and AI investors, understanding these emerging cost dynamics will be essential for navigating the next generation of AI infrastructure development.

Lianli Liquid Cooling’s One-Stop Liquid Cooling Services for Data Centers:

As a leading provider of liquid cooling solutions for AI data centers, Lianli Liquid Cooling delivers products that reduce PUE to below 1.15 and achieve an overall energy-saving rate exceeding 30%. Its product portfolio—ranging from cold plates and liquid-cooled server racks to CDUs and containerized data centers—is deeply compatible with NVIDIA’s product series (such as the H100/H200 and GB200/GB300). These products hold authoritative international certifications, including RoHS, CE, and UL, and are available for rapid delivery.