As AI, large language models, and GPU clusters continue to scale rapidly, data centers are entering an era of higher power density and higher computing density.

In the past, air cooling was sufficient for most servers. However, as AI GPU power and rack density continue to increase, efficiently removing the heat generated by high-density computing is becoming a critical challenge for AI data center infrastructure.

In 2026, demand for power and cooling infrastructure driven by AI continues to grow rapidly. Liquid cooling is gradually moving beyond its traditional role in high-performance computing and becoming an important infrastructure technology for next-generation AI data centers.

NVIDIA's GB200 NVL72 is a representative example. It integrates 36 Grace CPUs and 72 Blackwell GPUs into a rack-scale liquid-cooled system. NVIDIA's subsequent GB300 NVL72 also adopts a fully liquid-cooled rack-scale architecture.

So, what is direct liquid cooling? What is direct-to-chip cooling? And why is it becoming increasingly important for AI data centers?

What Is Direct Liquid Cooling?

Direct Liquid Cooling is a data center cooling method that uses liquid to directly remove heat from IT equipment.

Traditional air cooling mainly relies on air circulation:

CPU / GPU → Air → Fan → Cooling System

Direct liquid cooling, by contrast, uses coolant to transfer heat:

CPU / GPU → Liquid Cooling Loop → Cooling System

Because liquids can transfer high-density heat more efficiently than air, direct liquid cooling is particularly suitable for high-power computing equipment such as AI GPUs and HPC servers.

It is important to note that Direct Liquid Cooling is a broad term, while Direct-to-Chip Liquid Cooling is one of its key implementation methods.

What Is Direct-to-Chip Liquid Cooling?

Direct-to-Chip Liquid Cooling (D2C), also known as chip-level liquid cooling, is a cooling method in which the liquid cooling system directly serves high-power chips such as CPUs and GPUs.

Instead of relying primarily on air to remove heat, D2C typically uses a Cold Plate mounted on or directly connected to the chip to transfer heat into the coolant.

The basic cooling path can be simplified as:

GPU / CPU
↓
Cold Plate
↓
Coolant
↓
CDU
↓
Facility Cooling System

Therefore, when users search for:

• what is direct to chip cooling
• what is chip-level liquid cooling

they are essentially looking for information about a cooling technology that transfers heat directly from the chip into a liquid cooling loop.

Why Do AI Data Centers Need Direct-to-Chip Liquid Cooling?

The reason is straightforward:

Higher AI computing power → Higher GPU power → More concentrated heat → Greater cooling challenges

NVIDIA's GB200 NVL72 integrates 72 Blackwell GPUs into a liquid-cooled rack, showing that liquid cooling is no longer limited to individual high-performance servers. It is increasingly being deployed at the rack level for high-density AI infrastructure.

At the same time, the industry is moving toward increasingly high-power AI racks. While a traditional server rack might operate at several tens of kilowatts, high-density AI infrastructure is moving toward 100 kW+ racks and even higher power densities.

This raises an important question:

Can traditional air cooling continue to handle this concentrated heat economically and efficiently?

The value of liquid cooling therefore extends beyond simply improving heat dissipation. It also enables:

• Higher computing density
• Improved energy efficiency
• Greater data center scalability

How Does Direct-to-Chip Liquid Cooling Work?

A typical D2C liquid cooling system can be understood through three key components.

1. Cold Plate: Directly Capturing Heat

The Cold Plate is installed directly on or near high-power CPUs and GPUs.

Heat generated by the chip is first transferred to the cold plate and then carried away by coolant flowing through its internal channels.

GPU → Cold Plate → Coolant

This is the core heat-transfer step in chip-level liquid cooling.

2. CDU: Managing the Liquid Cooling Loop

After absorbing heat from the chips, the coolant flows into the CDU (Coolant Distribution Unit).

A CDU can perform several important functions, including:

• Coolant circulation
• Flow distribution
• Temperature control
• Pressure management
• Heat exchange
• Liquid cooling system monitoring

The basic flow can be represented as:

Cold Plate → Liquid Loop → CDU

3. Facility Cooling: Rejecting Heat

The CDU transfers the heat from the IT liquid cooling loop to the data center's facility-side cooling system, where the heat is ultimately rejected.

Therefore, a complete D2C architecture can be simplified as:

Chip → Cold Plate → CDU → Facility Cooling

AMD Socket SP5 Cold Plate | CPU Cold Plate

Cold Plate: A Core Component of Chip-Level Liquid Cooling

In a Direct-to-Chip Liquid Cooling system, the Cold Plate is one of the components located closest to the CPU/GPU heat source.

Its primary function is to efficiently transfer heat generated by the chip to the coolant.

For AI GPUs and high-performance CPUs, cold plate design needs to consider several factors:

• Heat Transfer
• Thermal Resistance
• Coolant Flow Rate
• Pressure Drop
• Chip Temperature
• Long-Term Reliability

A Cold Plate is therefore more than simply a "heat sink."

It plays a critical role in determining how efficiently heat generated by the chip can enter the liquid cooling loop.

For AI, HPC, and high-density server projects requiring Direct-to-Chip Liquid Cooling, cold plates can be designed according to parameters such as chip power, heat flux density, flow rate, and system pressure.

Cold Plate → Chip-Level Heat Removal

This is the first layer of the liquid cooling system.

CDU: Connecting Chip-Level Cooling to the Data Center Cooling System

If the Cold Plate is responsible for removing heat from the chip, the CDU is responsible for managing the coolant within the liquid cooling system.

A CDU (Coolant Distribution Unit) is typically positioned between the IT liquid cooling loop and the data center's facility cooling system.

Depending on the system architecture, a CDU can provide:

• Coolant circulation
• Flow distribution
• Temperature control
• Pressure management
• Heat exchange
• System monitoring

As AI GPUs scale from individual servers to multiple servers and racks, CDUs are becoming an increasingly important part of liquid cooling infrastructure.

Cabinet CDU vs. Rack CDU: Which One Should You Choose?

Different liquid cooling architectures can use different types of CDUs.

Cabinet CDU

A Cabinet CDU is designed to integrate with liquid-cooled server cabinets.

A typical architecture is:

Liquid-Cooled Server Cabinet
↓
Cabinet CDU
↓
Facility Cooling

This configuration is suitable for cabinet-level liquid cooling systems.

Rack CDU

A Rack CDU is positioned closer to the liquid cooling management requirements of high-density AI racks.

A typical architecture is:

AI GPU Servers
↓
Rack Liquid Loop
↓
Rack CDU
↓
Facility Cooling

For high-density GPU clusters, Rack CDU design can be based on parameters such as Rack Power, Cooling Capacity, Flow Rate, and supply/return coolant temperatures.

In other words, the CDU is not simply an independent cooling device. It serves as a critical fluid-management and heat-exchange node within the Direct-to-Chip liquid cooling system.

600kW In-Row Coolant Distribution Unit (CDU) | Data Center Liquid Cooling Solution

In other words, the CDU is not simply an independent cooling device. It serves as a critical fluid-management and heat-exchange node within the Direct-to-Chip liquid cooling system.

How Does Direct Liquid Cooling Improve Data Center Energy Efficiency?

This is the key question behind searches such as:

"how direct liquid cooling improves data center energy efficiency"

and

"how does direct-to-chip liquid cooling enhance data center efficiency"

Liquid cooling does not automatically produce the same energy savings in every data center.

Actual results depend on factors such as:

• IT load
• Coolant temperature
• CDU efficiency
• Data center climate
• Heat rejection method
• Control strategy

However, from a system perspective, Direct-to-Chip Liquid Cooling can improve efficiency in three major ways.

1. More Direct Heat Removal

Coolant removes heat directly near the GPU/CPU, reducing reliance on air as the primary heat-transfer medium.

2. Lower Air-Cooling Load

When more IT heat is transferred into the liquid cooling loop, the load on fans and air circulation systems can be reduced.

3. Support for Higher Computing Density

This is particularly important for AI data centers.

Liquid cooling is not only about using less energy to remove heat. It also enables more computing capacity to be deployed within limited space and infrastructure constraints.

NVIDIA's GB200 NVL72 architecture highlights the role of liquid cooling in supporting higher compute density while reducing energy use, carbon emissions, and physical footprint.

As a result, future AI data centers are increasingly focused not only on:

Cooling Efficiency

but also on:

Compute Density

and:

Compute per Megawatt

Direct-to-Chip vs. Air Cooling

This does not mean that air cooling will be immediately replaced.

For conventional servers and lower-density data centers, air cooling can still offer advantages in terms of cost and maintenance.

Liquid cooling becomes particularly important for high-power, high-density computing environments.

From Cold Plate to Liquid-Cooled Server Cabinet

As Direct-to-Chip Liquid Cooling expands from individual chips to complete servers and then to AI racks, the complexity of the cooling system also increases.

The architecture can be understood as:

Chip Level
GPU / CPU
↓
Cold Plate

Server Level
Liquid-Cooled Server
↓

Rack Level
Liquid-Cooled Server Cabinet
↓
CDU

Facility Level
Data Center Cooling System

The Liquid-Cooled Server Cabinet therefore serves as an important link between chip-level cooling and rack-level liquid cooling.

For high-density AI data centers, the focus is no longer on designing a single cold plate, but on integrating:

Cold Plate + Liquid Loop + CDU + Liquid-Cooled Rack

into a complete cooling system.

48U Liquid-Cooled Server Rack | Leak-Proof & Energy Efficient AI Rack Solutions

Applications of Direct-to-Chip Liquid Cooling

As AI infrastructure continues to evolve, Direct-to-Chip Liquid Cooling is particularly suitable for:

AI Data Centers

AI training, inference, and GPU clusters.

HPC

High-performance computing and scientific computing applications with high heat flux densities.

High-Density GPU Clusters

Computing environments with large numbers of GPU servers deployed in close proximity.

High-Density Data Center Racks

High-power racks where traditional air cooling becomes increasingly difficult or inefficient to scale.

Modular / Containerized Data Centers

Liquid cooling systems can also be integrated into modular and containerized data center infrastructure.

What Does a Complete Direct-to-Chip Liquid Cooling System Include?

Ultimately, chip-level liquid cooling is not a single product. It is a complete system.

A typical architecture can be understood as:

GPU / CPU
↓
Cold Plate
Chip-level heat transfer
↓
Liquid Loop
Coolant circulation
↓
Cabinet CDU / Rack CDU
Fluid distribution and heat exchange
↓
Liquid-Cooled Server Cabinet
Server- and rack-level integration
↓
Facility Cooling
Data center-level heat rejection

In other words:

From Cold Plate to CDU and then to the Liquid-Cooled Server Cabinet, liquid cooling is evolving from individual thermal-management components toward complete data center liquid cooling solutions.

https://cn.lianliwork.com/

FAQ: Direct-to-Chip Liquid Cooling

What is direct liquid cooling?

Direct Liquid Cooling is a data center cooling method that uses liquid to directly remove heat from IT equipment.

What is direct to chip cooling?

Direct-to-Chip Cooling is a liquid cooling technology that uses components such as Cold Plates to directly remove heat from high-power CPUs and GPUs.

How does direct-to-chip liquid cooling work?

Coolant absorbs heat from the CPU/GPU through a Cold Plate, flows through the liquid cooling loop and CDU, and ultimately transfers the heat to the data center's facility cooling system.

How does direct liquid cooling improve data center energy efficiency?

By removing heat more directly from high-power chips and reducing part of the air circulation and fan cooling load, Direct Liquid Cooling can improve overall cooling efficiency while supporting higher computing density.

What is a CDU in liquid cooling?

CDU stands for Coolant Distribution Unit. It manages coolant circulation, flow distribution, temperature control, pressure management, and heat exchange within a liquid cooling system.

Conclusion

AI data centers are entering an era of higher power density and higher computing density.

From liquid-cooled AI systems such as NVIDIA GB200 and GB300 to increasingly high-density GPU racks, the industry is shifting from simply asking:

How do we cool a server?

to a broader question:

How do we efficiently manage heat across an entire AI data center?

Direct-to-Chip Liquid Cooling connects chip-level heat removal with rack-level and facility-level cooling through:

Cold Plate → Liquid Loop → CDU → Liquid-Cooled Rack

For AI, HPC, and high-density data centers, the future value of liquid cooling goes beyond heat removal. It also involves compute density, energy efficiency, and infrastructure scalability.

For projects planning Direct-to-Chip Liquid Cooling, key components such as Cold Plates, Cabinet CDUs, Rack CDUs, and Liquid-Cooled Server Cabinets can be selected and engineered according to actual GPU power, rack power, cooling capacity, flow requirements, and coolant conditions.

From chip to rack. From products to complete liquid cooling solutions.