Liquid-to-Air vs. Liquid-to-Liquid CDU: Choosing the Right for AI Data Centers

AI infrastructure is changing the thermal design assumptions of the data center.

For decades, most enterprise and colocation facilities were designed around air cooling. Typical rack densities ranged from 5 kW to 15 kW, with CRAC or CRAH units supplying conditioned air through raised floors, cold aisles, or containment systems.

That model is now under pressure.

Modern GPU servers, AI training clusters, HPC platforms, and accelerated computing environments are pushing rack densities far beyond traditional air-cooling limits. In many deployments, rack loads of 30 kW, 50 kW, and beyond are becoming increasingly common. At the highest end of the market, next-generation AI platforms are driving data center operators toward liquid cooling as a core infrastructure requirement rather than an optional upgrade.

At the center of this transition is the Coolant Distribution Unit, or CDU.

A CDU manages the coolant loop between IT equipment and the facility cooling system. It typically includes pumps, heat exchangers, controls, sensors, filtration, and safety components that regulate coolant temperature, pressure, and flow. In direct-to-chip liquid cooling deployments, the CDU provides the controlled coolant supply required to remove heat from CPUs, GPUs, memory, and other high-power components.

However, not all CDU architectures are the same.The two most common approaches are:

Liquid-to-Air CDU (L2A CDU)

Liquid-to-Liquid CDU (L2L CDU)

The right choice depends less on which technology is “better” and more on the condition of the existing facility, target rack density, deployment timeline, and long-term expansion strategy.

What Is a Liquid-to-Air CDU?

A Liquid-to-Air CDU transfers heat from the server coolant loop into the surrounding data center air.

In this configuration, coolant circulates through cold plates installed on GPUs, CPUs, or other high-power components. The warmed coolant returns to the CDU, where an internal air-to-liquid heat exchanger removes the heat. Fans inside the CDU then reject that heat into the data center environment, where it is handled by the existing room cooling system.

In practical terms, the server is liquid-cooled, but the facility still relies on air as the final heat rejection medium.

This makes Liquid-to-Air CDU systems especially useful in facilities that were not originally designed with chilled-water infrastructure or facility-water loops. Eaton notes that L2A CDUs can provide localized liquid cooling in air-cooled facilities without requiring full-scale facility-water implementation, making them particularly relevant for retrofits, edge sites, and colocation environments.

Key Advantages of Liquid-to-Air CDUs

Liquid-to-Air CDUs are often selected because they minimize infrastructure disruption.

They can allow operators to introduce liquid-cooled servers without installing a new chilled-water loop, cooling tower connection, or extensive piping network. This can significantly reduce construction complexity, shorten deployment time, and avoid major disruption to existing operations.

For many enterprise data centers, this is the most practical first step toward liquid cooling.

Key benefits include:

No facility-water connection required

Faster deployment in existing air-cooled spaces

Lower upfront infrastructure cost

Reduced need for mechanical and civil modifications

Suitable for phased AI deployment

Useful for edge, enterprise, and colocation environments

Some current L2A CDU platforms are designed for much higher capacities than earlier generations. For example, CoolIT Systems lists a liquid-to-air CDU designed for up to 240 kW of cooling capacity without facility-water integration. This illustrates that L2A is no longer limited only to small pilot projects; however, its practical suitability still depends on the room’s ability to absorb and reject the resulting heat.

The Limitation: Heat Still Ends Up in the Room

The main constraint of a Liquid-to-Air CDU is straightforward: the heat must still be removed by the data center’s air-cooling infrastructure.

Even if the CDU efficiently cools the server coolant loop, it transfers that heat into the white space. The existing CRAC, CRAH, in-row cooling, or perimeter cooling system must then handle the additional thermal load.

As a result, L2A CDU systems are most effective when:

The facility has unused cooling capacity

AI deployments are limited to a defined number of racks

The project requires rapid deployment

Building water is unavailable or too expensive to add

The goal is to extend the useful life of an existing air-cooled site

They are generally less attractive for large-scale AI campuses where hundreds of high-density racks must operate continuously at very high utilization.

What Is a Liquid-to-Liquid CDU?

A Liquid-to-Liquid CDU transfers heat from the server coolant loop to a separate facility-water loop.

The server-side loop, often called the technology cooling system or secondary loop, remains isolated from the building-side loop. Heat is transferred through a plate-and-frame heat exchanger inside the CDU, while the facility water carries that heat to chillers, cooling towers, dry coolers, or other external heat-rejection equipment.

This separation is important because it allows the server coolant loop to maintain tightly controlled water quality, pressure, temperature, and cleanliness while keeping it isolated from the broader facility cooling infrastructure.

According to nVent, the CDU helps control coolant pressure, flow, and temperature while maintaining separation between the technology cooling system and facility-water system.

Key Advantages of Liquid-to-Liquid CDUs

Liquid-to-Liquid CDU systems are designed for higher-density and more scalable liquid-cooling environments.

Because heat is rejected directly to facility water rather than the data hall air, L2L systems reduce the thermal burden on room air-conditioning equipment. They are typically better suited to large AI clusters, HPC deployments, GPU training environments, and new-build data centers designed around liquid cooling from the start.

Key benefits include:

Higher cooling capacity and scalability

Reduced dependence on room air cooling

Better support for high-density AI racks

More efficient heat transfer

Improved long-term expansion capability

Better alignment with campus-scale cooling infrastructure

Potential for lower long-term cooling energy use

Eaton states that liquid-to-liquid CDUs are designed for facility-water-based cooling and can support rack densities from approximately 50 kW to more than 200 kW per rack, depending on the system architecture and operating conditions.

Large in-row L2L CDU systems can also support multiple racks at megawatt-scale capacities. For example, Eaton’s enterprise CDU portfolio includes systems designed to leverage facility water for high-capacity in-row cooling.

The Trade-Off: More Infrastructure Required

The main disadvantage of L2L CDU systems is that they require a suitable facility-water infrastructure.

For an existing air-cooled data center, that may mean adding:

Primary and secondary water loops

Facility piping

Pumps and valves

Heat-rejection equipment

Cooling towers, dry coolers, or chillers

Water treatment and monitoring systems

Leak detection and containment measures

Controls integration with the building management system

This increases capital cost and engineering complexity. It may also require scheduled shutdowns or phased construction if the facility is already operating.

For this reason, L2L CDU systems are generally the preferred choice for new data centers, major expansions, or dedicated AI halls where high-density compute is expected to scale over time.

Liquid-to-Air vs. Liquid-to-Liquid CDU: Key Comparison

Why “Liquid-to-Air vs. Liquid-to-Liquid” Is Not a Simple Capacity Debate

It is tempting to assume that Liquid-to-Liquid CDU systems are always the superior option because they support higher density and more efficient heat rejection.

That is not always true.

A data center operator with an existing air-cooled facility may not need to build a full facility-water system to deploy a small AI cluster. If the goal is to install several liquid-cooled GPU racks quickly, a Liquid-to-Air CDU may provide the most economical path.

Conversely, deploying L2A CDUs across a large AI environment without considering the room-level cooling impact can create a bottleneck elsewhere in the system. The CDU may cool the IT equipment successfully, but the facility may still struggle to remove the heat from the white space.

The decision should therefore be based on the entire thermal chain:

Chip → cold plate → server loop → CDU → room or facility cooling system → external heat rejection

A CDU should never be selected in isolation.

When Should You Choose a Liquid-to-Air CDU?

A Liquid-to-Air CDU is usually the better choice when the project involves an existing air-cooled facility and the objective is fast, low-disruption deployment.

It is particularly suitable for:

Enterprise AI inference clusters

Small to medium GPU deployments

Existing colocation suites

Edge data centers

Proof-of-concept AI deployments

Phased migration from air cooling to liquid cooling

Sites without facility-water availability

Projects with limited capital expenditure budgets

This approach is often ideal when operators need to deploy liquid-cooled servers now but are not yet ready to redesign the entire mechanical infrastructure of the facility.

When Should You Choose a Liquid-to-Liquid CDU?

A Liquid-to-Liquid CDU is generally the better option when high-density computing is expected to become a long-term part of the facility strategy.

It is particularly suitable for:

New AI data centers

Large-scale GPU training clusters

HPC and supercomputing environments

High-density colocation halls

Hyperscale AI campuses

Facilities targeting 50 kW, 100 kW, or higher rack densities

Projects with long-term expansion plans

Sites seeking to reduce dependence on traditional room air cooling

L2L systems require more planning, but they provide a more scalable foundation for high-density computing.

Modern CDU designs are increasingly available in rack, in-row, and perimeter configurations, with options for liquid-to-air and liquid-to-liquid heat exchangers depending on facility requirements.

A Hybrid Cooling Strategy May Be the Most Practical Path

For many operators, the best strategy is not choosing one architecture forever.

A phased approach can reduce risk and spread capital expenditure over time:

Deploy Liquid-to-Air CDUs for initial AI racks.

Validate workload growth, cooling demand, and operational requirements.

Upgrade selected halls with facility-water infrastructure.

Introduce Liquid-to-Liquid CDUs for larger AI clusters.

Transition toward a mixed air- and liquid-cooled data center model.

This hybrid strategy allows operators to support immediate AI demand while preserving a path toward larger-scale liquid cooling in the future.

Lianli’s CDU Liquid Cooling Solutions:

Lianli independently develops and manufactures CDUs—including in-rack CDU and in-row CDU models—offering both air-to-liquid CDU and liquid-to-liquid CDU cooling options. Our comprehensive data center liquid cooling solutions can reduce PUE to as low as 1.15 and achieve an overall energy saving rate exceeding 30%, positioning us at the forefront of the industry. Our expert team possesses deep insight into global market standards and local requirements; we welcome you to visit our factory for an on-site inspection.

Final Thoughts

Liquid cooling is becoming a core part of data center infrastructure planning, especially as AI and accelerated computing continue to increase rack density.

Liquid-to-Air CDUs provide a practical, fast, and lower-cost way to introduce liquid cooling into existing air-cooled facilities. They are particularly valuable for retrofit projects, enterprise AI deployments, edge environments, and colocation sites where facility water is unavailable.

Liquid-to-Liquid CDUs provide a more scalable and efficient architecture for high-density AI data centers, HPC environments, and new-build facilities. They require more infrastructure investment, but they offer stronger long-term capacity for large-scale liquid-cooled computing.

The best CDU is not necessarily the one with the highest cooling capacity.It is the one that matches your facility infrastructure, AI deployment roadmap, rack density target, budget, and long-term expansion strategy.