AI Data Center Cooling Systems: From Rack Heat Density to Liquid Cooling and Heat Rejection

How Direct-to-Chip Cooling, CDUs, Facility Water Systems and HVAC Heat Rejection Work Together

AI data centers are creating a new cooling problem for HVAC engineers. The challenge is not simply that the building needs more cooling capacity. The bigger issue is that much more heat is being concentrated into individual racks, and that heat has to be removed reliably from very small areas.

According to the International Energy Agency, global data center electricity demand increased by approximately 17% in 2025, while electricity consumption from AI-focused data centers increased even faster. More electrical power entering computing equipment ultimately means more heat that the cooling system has to remove.

ASHRAE's current guidance for AI data centers discusses rack environments in the 50–100+ kW range and higher. At these densities, conventional room-level air cooling becomes increasingly difficult to use as the only cooling method.

1. Higher Rack Heat Density Changes the Cooling Design

Traditional data centers have relied heavily on air cooling. Servers transfer heat to the surrounding air, and CRAC, CRAH or other facility cooling systems remove that heat from the room.

This approach remains useful for many applications, but high-density GPU racks change the heat distribution inside the data center.

Instead of a relatively distributed room load, large amounts of heat may now be concentrated into a much smaller number of racks. A facility can therefore have enough total cooling capacity and still experience thermal problems if that capacity cannot remove heat where the heat is actually being generated.

For HVAC engineers, rack density therefore becomes just as important as total building cooling capacity.

2. Follow the Heat: From the Chip to the Outdoor Environment

A practical way to understand modern data center cooling is to follow the heat from the processor all the way to final heat rejection.

This thermal path makes one point very clear: liquid cooling does not eliminate the facility cooling system. It changes where heat is captured and how that heat is transported.

Modern data center cooling should be viewed as one continuous thermal path, from heat capture at the chip to final heat rejection outside the facility.

3. Direct-to-Chip Liquid Cooling Moves Heat Removal Closer to the Source

In direct-to-chip liquid cooling, coolant flows through cold plates attached directly to high-heat components such as GPUs or CPUs.

This allows a large portion of processor heat to be captured before it is released into the data center room air.

A typical liquid-cooling path may include:

• Cold plates
• Rack or server manifolds
• Technology cooling loop
• Cooling Distribution Unit (CDU)
• Facility water connection

The CDU is an important interface between the IT-side cooling loop and the building's facility cooling system. Depending on the system architecture, it can provide heat exchange, hydraulic separation, flow control, pressure control, temperature control and monitoring.

4. Liquid Cooling Does Not Mean Air Cooling Disappears

One common misunderstanding is that liquid cooling will completely replace air cooling in AI data centers.

In practice, many facilities are moving toward hybrid cooling. Liquid cooling can handle the highest-density processor loads, while air cooling continues to remove heat from components such as power supplies, memory, storage, networking equipment and other lower-density loads.

ASHRAE also describes hybrid strategies for existing facilities, where direct-to-chip liquid cooling is added for high-density GPU loads while existing air-cooling infrastructure continues to serve residual loads.

The design question therefore should not be:

Air cooling or liquid cooling?

A better question is:

Which cooling method should handle each part of the heat load?

5. The Facility Water System Becomes a Critical Part of the Design

Once server heat enters a liquid loop, the project becomes closely connected with traditional hydronic HVAC engineering.

Important facility-side design parameters include:

• Supply water temperature
• Return water temperature
• Water flow rate
• System temperature difference (ΔT)
• Pump head
• Heat exchanger performance
• Chiller or dry-cooler capacity
• Redundancy requirements

The basic thermal relationship remains familiar: higher flow or a larger water temperature difference allows more heat to be transported through the water loop.

But data center systems add another requirement: the cooling infrastructure must maintain performance continuously, often under rapidly changing IT loads.

6. Chiller, Dry Cooler or Cooling Tower?

There is no single heat-rejection solution that is best for every data center.

The choice depends on factors such as:

• Local climate
• Required facility water temperature
• Peak outdoor temperature
• Water availability
• Electricity cost
• Redundancy requirements
• Available installation space

The final cooling architecture depends on rack heat density, facility water temperatures, climate, energy cost, water availability and reliability requirements.

7. Energy, Water and Reliability Have to Be Considered Together

A data center cooling system cannot be evaluated only by cooling capacity.

Energy consumption, water use and system reliability are directly connected with cooling-system design.

Power Usage Effectiveness (PUE) remains a widely used data center metric:

However, PUE alone does not describe every aspect of cooling performance. Water use, available cooling capacity, operating conditions, resilience and redundancy also need to be considered.

For example, evaporative heat rejection can reduce electrical consumption in some climates while increasing water use. Dry cooling can reduce direct water consumption, but its performance depends strongly on outdoor temperature and the required water temperature.

The practical design target therefore becomes:

8. Data Center Cooling Controls Are Becoming More Important

High-density computing loads are not always constant. Cooling demand can change as server load changes, which makes coordination between cooling equipment increasingly important.

A modern cooling system may have to coordinate:

• Temperature sensors
• Flow sensors
• CDUs
• Pumps
• Control valves
• Air-handling equipment
• Chillers
• Dry coolers
• Cooling towers
• Heat-rejection equipment

The control sequence may gradually move from simple reactive operation toward:

Predictive control, fault detection, load forecasting and digital-twin tools can support this process, but they do not replace good mechanical design.

Correct sensor selection, hydraulic design, equipment redundancy, commissioning and fail-safe control logic remain fundamental.

9. What HVAC Professionals Need to Understand

Modern data center cooling increasingly requires coordinated monitoring of IT load, temperatures, flow rates, pumps, cooling equipment and heat rejection.

10. From Cooling Equipment to Complete Thermal Management

The biggest change in AI data center cooling is not one specific technology.

It is the need to look at the entire thermal path as one system:

For HVAC professionals, this creates a larger engineering role than conventional comfort cooling.

The next generation of data center cooling will require a better understanding of rack heat density, liquid cooling, hydronic systems, heat rejection, controls, energy use and mission-critical reliability.

The goal is no longer simply to provide more cooling. It is to move heat from where it is generated to where it can be rejected safely, efficiently and reliably.

YUJU's View on Data Center Cooling

YUJU continues to follow the development of high-density data center cooling and the growing connection between HVAC equipment, hydronic systems, intelligent controls and energy management.

For facility-side cooling applications, YUJU can support international partners with HVAC equipment selection, hydronic system evaluation, heat-rejection equipment, control integration and project engineering discussions according to the actual application requirements.