As rack densities increase in AI and high-performance computing systems, liquid cooling systems are becoming critical. At the center of these systems is the CDU, a critical mechanical equipment unit that transfers heat from IT equipment to the facility cooling system and manages coolant flow rate and temperature.
In our previous article, we examined why AI data centers are moving from air cooling to liquid cooling systems.
One of the key pieces of equipment at the center of this transformation is:
CDU – Coolant Distribution Unit
CDU stands for Coolant Distribution Unit. In some ASHRAE documents, the term Cooling Distribution Unit is also used for similar equipment. In the current AI Data Center Energy Performance Framework, the CDU is defined as an interface that provides heat transfer, pumping, and temperature control between the IT-side cooling circuit and the facility water system.
Simply put, the CDU:
Acts as a bridge between the facility-side cooling system of the data center and the technology cooling circuit serving the GPUs.
However, it would be incomplete to think of a CDU simply as a heat exchanger.
A modern CDU may contain many components, including:
- heat exchanger,
- pumps,
- control valves,
- filtration,
- temperature sensors,
- flow measurement,
- pressure measurement,
- automation,
- leak detection,
- redundant pumps and power systems
and many other components.
Why Is a CDU Necessary?
First, let’s look at the entire system. A simplified AI data center liquid cooling system can be considered as follows:
Chiller / Dry Cooler
↓
Facility Water Loop
↓
CDU
↓
Technology Cooling System – TCS
↓
Rack Manifold
↓
Cold Plate
↓
GPU / CPU
The heat generated by the GPU is transferred to the coolant through the cold plate. The heated coolant returns to the CDU. Inside the CDU, heat is transferred to the facility-side cooling circuit through the heat exchanger. The cooled technology cooling fluid is then sent back to the racks. This cycle continues continuously. ASHRAE data center design documents also state that facility chilled water is supplied to the CDU, while the CDU connects to the data center equipment through the TCS.
How Does a CDU Work?
It is more accurate to think of the CDU’s operating principle in terms of two separate water/fluid circuits.
1. Facility Water Loop
This is the main cooling system on the facility side. For example:
Chiller → Pump → Facility Water Loop → CDU → Chiller
or, depending on the design:
Dry Cooler → Facility Water Loop → CDU → Dry Cooler
The circuit is connected to the facility’s heat rejection system.
2. Technology Cooling System – TCS
The second circuit is directly associated with cooling the IT equipment. For example:
CDU → Rack → Manifold → Cold Plate → Rack → CDU
This circuit is called the Technology Cooling System (TCS). Depending on the system design, the fluid used within the TCS may be chilled water, deionized water, different coolants, or other specialized fluids. ASHRAE states that the TCS fluid may vary depending on the system architecture. The most important point here is:
Facility Water and the cooling fluid on the IT side do not have to be the same circuit.
One of the CDU’s most important functions is to separate these two sides while providing controlled heat transfer between them.
How Does Heat Transfer Occur Inside a CDU?
One of the central components inside a CDU is the heat exchanger. To simplify the process:
Hot TCS Fluid
↓
Heat Exchanger
↓
Facility Water
↓
Chiller / Dry Cooler / Heat Rejection
Heat transfer takes place in this manner. The heat generated by the GPUs has already been transferred to the TCS fluid. When this hot fluid reaches the CDU, it transfers its heat to the facility water side. The cooled TCS fluid is then sent back to the GPUs. The CDU’s role is not simply to transfer heat from one side to the other. It also manages the answer to the question: “At what temperature and at what flow rate should fluid be supplied to the GPUs?”
Why Is There a Pump Inside a CDU?
The fluid on the TCS side must be continuously circulated to the racks. Therefore, one or more pumps may be installed inside the CDU.
The pump:
- provides the required flow rate,
- establishes system pressure,
- provides the fluid circulation required by the cold plates in the racks.
However, there is an important engineering consideration here:
Rack load is not constant.
Computational workloads can vary in AI systems. As GPU load increases, heat generation also changes. Therefore, operating a fixed-speed pump continuously at maximum capacity is not always the optimum solution. Modern CDU systems can use variable-speed pumps and control systems to adjust the flow rate in real time according to demand. The current ASHRAE AI data center framework also identifies variable-speed pumping, differential pressure control, and flow control as key control strategies for modern TCS systems.
Why Is Flow Rate So Important in a CDU?
To transport the heat generated by a GPU through a liquid, the fundamental relationship can be expressed as: Q = ṁ × Cp × ΔT
Here:
- Q = heat transferred,
- ṁ = mass flow rate,
- Cp = specific heat capacity of the fluid,
- ΔT = supply-to-return temperature difference.
This equation is particularly important for understanding the fundamentals of CDU design. For example, when you need to transport the same heat load, you can:
- increase the flow rate,
- increase the ΔT,
- use different fluid properties.
However, every choice has hydraulic and thermal consequences. Increasing the flow rate beyond what is necessary can increase pump energy consumption. Increasing ΔT too much can affect the temperature conditions on the chip/cold plate side. Therefore, the objective in CDU design is not simply:
“Achieve the highest possible flow rate”
The objective is:
Transport the required heat within the required temperature range, with minimum auxiliary energy consumption and sufficient reliability.
How Is Temperature Controlled in a CDU?
Temperature control is one of the critical issues in AI data center liquid cooling systems. The temperature of the fluid supplied to the GPU can affect:
- cold plate performance,
- processor temperature,
- free cooling potential,
- chiller operating time,
- system efficiency
Information received from sensors installed on the CDU can be evaluated by the automation system to manage parameters such as:
- pump speed,
- control valves,
- fluid temperature,
- flow rate,
- differential pressure
The supply temperature optimization approach is specifically emphasized in ASHRAE’s 2026 AI Data Center Framework. Higher suitable fluid temperatures can potentially increase the duration of free cooling operation and reduce compressor usage.
Why Does a CDU Use Two Separate Circuits?
The answer to this question lies in data center reliability. Establishing a direct connection between the IT equipment cooling circuit and the facility’s main cooling system is not appropriate for every project. A CDU creates a controlled heat transfer point between the two sides. Its advantages include:
- managing different pressure levels,
- allowing different temperature levels to be used,
- controlling the fluid on the IT side,
- managing filtration and water quality,
- facilitating maintenance,
- enabling control according to different rack densities
For this reason, it is more accurate to consider a CDU not simply as a “heat exchanger,” but as a controlled thermal interface between two different hydronic systems.
Why Is Filtration Important in a CDU?
Fluid cleanliness can be critical in liquid cooling systems. The channels inside cold plates, manifolds, control valves, and other components may have sensitive hydraulic structures. Particles within the system can:
- clog filters,
- reduce flow rate,
- increase pressure loss,
- affect the operation of control valves,
- reduce cold plate performance.
Therefore, filtration is an important component of CDU design. However, the conventional mechanical engineering principle also applies here: Installing a filter alone is not enough. The filter’s:
- capacity,
- mesh/micron rating,
- pressure drop,
- maintenance method,
- bypass arrangement,
- filter replacement procedure
must all be evaluated together.
Why Is Leakage Such a Major Problem in a CDU?
This brings us to one of the most discussed issues in liquid cooling systems:
Leak Detection
Liquid is being circulated in the data center directly next to electronic equipment. Therefore, even a small leak can have serious consequences. For this reason, modern systems may use safety functions such as:
- leak detection sensors,
- automatic isolation,
- pressure monitoring,
- flow monitoring,
- alarm systems
Some next-generation systems are also developing alternative approaches such as negative-pressure liquid cooling instead of positive-pressure systems. For example, Chilldyne technology, acquired by Daikin in 2025, uses a negative-pressure direct-to-chip liquid cooling approach intended to reduce leak risk. This demonstrates an important point: in AI data center liquid cooling, not only heat transfer but also how the fluid is controlled is part of the engineering challenge.
Why Is Redundancy Necessary in a CDU?
A few minutes of downtime for an air-conditioning unit in an office is not the same as the failure of a critical cooling system in an AI data center. In high-density GPU racks, loss of cooling can directly affect IT performance. Therefore, CDU systems may incorporate solutions such as:
- redundant pumps,
- redundant power supplies,
- parallel hydraulic paths,
- automatic transfer systems,
- equipment capable of continuing operation during maintenance
For example, Carrier’s new CDU design announced in 2026 includes redundant pumps and power supplies among the features supporting continued cooling during maintenance or unexpected events. The objective here is not simply to “install a redundant pump.” The real objective is:
Prevent the failure of a single piece of equipment from taking down a critical IT load.
Therefore, redundancy concepts such as N+1 become important in CDU design.
What Is the Relationship Between a CDU, BMS, and DCIM?
A CDU is no longer considered merely a mechanical piece of equipment. In modern data centers, mechanical systems need to operate together with automation and IT infrastructure. Through the CDU, information such as:
- temperature,
- flow rate,
- pressure,
- pump status,
- valve position,
- alarm status,
- leak status,
- filter status
can be monitored.
This information can be transmitted to the building automation system or data center infrastructure management system. The current ASHRAE AI Data Center Framework specifically emphasizes the integration of temperature, pressure, flow, and leak detection sensors with BMS/DCIM platforms. Therefore, the CDU is becoming a point where three elements converge: mechanical equipment + automation + data.
What Are the Types of CDUs?
CDUs do not have a single physical configuration. Different architectures can be used depending on project size and rack density.
In-Rack CDU
The CDU is installed directly inside or close to the rack. It can be preferred for more localized and modular solutions.
In-Row CDU
The CDU is positioned within the row where the server racks are located. With this approach, liquid distribution can be located closer to the rack group.
Room-Based CDU
Larger-capacity CDUs can be installed in an appropriate area of the data hall or mechanical space. ASHRAE’s data center handbook describes a modular room-based liquid cooling architecture in which facility chilled water is supplied to a CDU located in or near the data hall as a common configuration.
How Is CDU Capacity Determined?
CDU selection is not based solely on the question:
“How many kW will it cool?”
The following parameters must be evaluated together:
1. Total IT Heat Load
The total thermal load that the CDU must handle is determined.
2. Rack Density
For example, 50 kW/rack and 200 kW/rack do not require the same hydraulic and thermal design approach.
3. Supply and Return Temperatures
The supply and return temperatures on the TCS side are determined.
4. Flow Rate
The required fluid flow rate is calculated based on the required heat load.
5. Pressure Drop
The total pressure drop across the racks, manifolds, cold plates, piping, and valves must be considered.
6. Redundancy
The required redundancy level, such as N, N+1, or higher, is determined.
7. Heat Rejection System
The facility water system behind the CDU must be evaluated according to which system it will operate with, such as:
- chiller,
- dry cooler,
- cooling tower,
- heat pump
8. Future Capacity
If a system currently operating at 100 kW/rack is planned to increase to 200 or 300 kW/rack within a few years, the CDU and distribution infrastructure should be designed accordingly.
A CDU Is Not the Same as a Chiller
This distinction is particularly important. Chiller produces cooling. The CDU, on the other hand, ensures that this cooling is transferred to and distributed through the liquid cooling system on the IT side in a controlled manner.
In simplified terms:
Chiller = Cooling source
CDU = Thermal interface and distribution center
TCS = Technology cooling circuit
Cold Plate = Component that removes heat from the GPU/CPU
Therefore, it is not correct to think of a CDU as a “small chiller.”
Where Is CDU Technology Heading?
As AI computing power increases, CDU technology is also expected to evolve to support higher rack densities. Current examples of this trend can be seen in new products introduced in 2026. When Carrier introduced its new CDU for European data centers in 2026, it cited increasing rack densities and growing demand for liquid cooling among the key drivers. Its system includes precise secondary-circuit control, redundant pumps, and redundant power supplies.
Daikin and Delta Electronics also announced a collaboration in 2026 to develop CDU solutions for AI and HPC data centers. The companies stated that their solutions could support rack power densities ranging from 100 kW to 3000 kW. These figures demonstrate why the CDU is no longer considered a conventional auxiliary piece of equipment.
Conclusion: The CDU Is the Hydronic Heart of the AI Data Center
As liquid cooling systems become more widespread in AI data centers, the importance of the CDU continues to increase.
The CDU:
- connects the facility water system with the TCS,
- provides heat transfer through a heat exchanger,
- pumps the technology cooling fluid,
- controls the flow rate,
- manages temperature,
- provides filtration,
- collects data from sensors,
- can monitor leaks and fault conditions,
- can provide redundancy,
- can communicate with BMS/DCIM systems.
Therefore, it is not sufficient to define a CDU simply as:
“The heat exchanger in a liquid cooling system”
A more accurate definition is:
A CDU is critical hydronic equipment in an AI data center that manages heat transfer, fluid circulation, temperature control, and reliability between the facility cooling system and the technology cooling circuit serving high-density IT equipment.
And in fact, a major engineering transformation is taking place here. While data center cooling was traditionally approached largely through the:
Room → Air → CRAC/CRAH → Chiller concept,
high-density AI infrastructure is increasingly moving toward:
GPU → Cold Plate → TCS → CDU → Facility Water → Heat Rejection
In other words, artificial intelligence is not only changing the field of computer engineering. It is also redefining the role of mechanical installation engineering in the data center.
