NVIDIA's newest Blackwell chips are so power-hungry that traditional air cooling is becoming obsolete. The GB200 NVL72 rack system, which packs 72 Blackwell-generation processors, demands up to 140 kilowatts per rack, roughly five times more power than typical server racks from just a few years ago. This thermal intensity is forcing a fundamental redesign of how data centers operate, moving away from decades-old air-cooling infrastructure toward liquid-based systems that can handle the heat concentration.

How Much Heat Are We Actually Talking About?

To understand the scale of the problem, consider this: a single NVIDIA H100 GPU draws up to 700 watts of power, roughly equivalent to seven desktop computers worth of heat from one chip. A DGX H100 server holding eight of these GPUs pulls around 10.2 kilowatts under full load. But the newer GB200 NVL72 system takes that to another level entirely, demanding up to 140 kilowatts per rack.

This is not a minor engineering challenge. The heat output at modern GPU densities is comparable to a steel mill, making traditional air cooling fundamentally inadequate. Old data centers that handled email and basic web traffic relied on simple air circulation through server aisles. That model broke when AI accelerators arrived.

"Air has low heat capacity compared with liquids. It works well when heat is spread across many ordinary racks, each drawing a few kilowatts to perhaps tens of kilowatts. It becomes noisy, bulky and energy-hungry when a rack behaves like a small industrial machine," noted researchers analyzing the thermal engineering challenge.

Lawrence Berkeley National Laboratory and industry thermal analysis

What Cooling Solutions Are Data Centers Adopting?

Data centers now have several liquid-cooling options, each with different tradeoffs. The most dramatic approach gaining attention is two-phase immersion cooling, where servers are submerged in a dielectric fluid that boils at around 50 degrees Celsius. When the fluid reaches its boiling point on hot components, it vaporizes, rises to a condenser, cools back into liquid, and falls by gravity back into the tank. This creates a closed thermal cycle without traditional fans or compressors.

Other approaches include direct-to-chip cooling, where metal plates sit directly on processors with cool liquid flowing through them, and single-phase immersion systems that use liquid circulation without the phase-change mechanism. Each method has different implications for data center design, power consumption, and operational complexity.

Steps to Understanding Modern Data Center Cooling Strategies

• Evaporative Cooling: The most common method, where water passes through cooling pads and evaporates to remove heat, but loses about 80% of water to the air compared to households that return 90% of water back to supply systems.
• Direct-to-Chip Liquid Cooling: Metal plates attached directly to processors with cool liquid flowing through them, removing heat at the source before it spreads throughout the system.
• Two-Phase Immersion Cooling: Servers submerged in dielectric fluid that boils and condenses in a closed loop, enabling denser racks with less dependence on compressor-based cooling systems.

Why Is NVIDIA Redesigning Its Hardware for Liquid Cooling?

NVIDIA's decision to build Blackwell chips specifically for full liquid cooling signals that the company recognizes air cooling has reached its limits. This is not a minor design choice; it represents a fundamental shift in how accelerator hardware is engineered. The company is essentially saying that future AI infrastructure will be liquid-cooled by default, not as an optional upgrade.

Microsoft has already deployed two-phase immersion cooling in production at its Quincy, Washington data center, reporting a 5% to 15% power reduction for a given server and enabling denser cloud resources. This real-world deployment demonstrates that the technology is moving beyond experimental trials into mainstream operations.

What Does This Mean for Data Center Infrastructure?

The shift toward liquid cooling represents a broader transformation in how data centers are designed and operated. For decades, the standard data center was built around raised floors, cold aisles, hot aisles, and computer room air handlers. AI clusters have fundamentally changed that equation.

Air is losing its automatic status as the default cooling medium. Dense AI racks force data centers to choose between much more airflow, colder supply air, more containment, more fan energy, more floor space, or a shift toward liquid systems. For operators running Blackwell systems at 140 kilowatts per rack, liquid cooling is increasingly the only practical option.

The thermal problem extends beyond just the GPU itself. High-density AI nodes include CPUs, memory, NVLink interconnect devices, power conversion hardware, network switches, and storage. A rack-scale AI system is a complete thermal ecosystem, and immersion cooling addresses this by placing far more electronics in direct contact with coolant.

How Does This Connect to the Broader AI Infrastructure Challenge?

The cooling challenge is inseparable from the larger story of AI's infrastructure demands. Global data center electricity consumption is projected to double to about 945 terawatt-hours by 2030, with accelerated servers driven mainly by AI growing much faster than conventional server loads. U.S. data center electricity use rose from 58 terawatt-hours in 2014 to 176 terawatt-hours in 2023 and could reach 325 to 580 terawatt-hours by 2028, depending on growth assumptions.

Cooling is not an accessory to this equation. A facility that doubles its IT load must remove roughly double the heat unless servers become dramatically more energy-efficient per unit of work. AI does not only increase the electricity question; it increases the heat-rejection question simultaneously.

The shift toward liquid cooling for Blackwell systems reflects a hard physical reality: AI hardware is concentrating too much heat into too little space for ordinary airflow to remain viable. This is not a decorative engineering trick or a speculative future scenario. It is a direct response to the thermal limits of current GPU density, and it is reshaping how the world's most powerful data centers will be built and operated.