Planning for Liquid Cooling in AI/HPC Data Centers

Designing a new data center has always required looking years ahead. The rapid growth of artificial intelligence (AI) and high-performance computing (HPC) makes that challenge even greater.

Higher-density compute is changing the relationship among power, cooling, connectivity, and the physical infrastructure inside the rack.

For organizations building a new data center, the key question is not simply, "Do we need liquid cooling today?" A more useful question is, "How do we design the facility so it can support liquid-cooled infrastructure as our compute requirements evolve?"

That infrastructure could include rear-door heat exchangers (RDHx), single-phase or two-phase Direct-to-Chip (DTC) cooling, and, in some environments, immersion cooling. A new facility offers the opportunity to evaluate these technologies alongside power, networking, cabinets, and mechanical systems before major design decisions become difficult or expensive to change.

Here are eight areas to consider.

1. Design Around Tomorrow's Compute, Not Just Today's

Begin with the IT roadmap. What types of AI, GPU, and HPC systems might the facility support? How could rack densities change over the life of the data center?

• Which server platforms are likely to be deployed?
• What rack power densities should the facility accommodate?
• How much of the IT load may eventually require liquid cooling?
• Will air-cooled, rear-door-cooled, and direct-to-chip-cooled equipment need to coexist?
• How could these requirements change as future generations of compute arrive?

Design principle: Do not size the physical infrastructure solely around the first generation of equipment expected on opening day.

2. Plan the Facility-Water Architecture Early

A greenfield data center provides an important advantage: the mechanical infrastructure can be designed with liquid-cooled IT in mind.

Consider the complete heat-removal path early in the design:

Server cold plates → rack manifold → cooling distribution unit (CDU) → facility-water system → heat rejection

Facility water, CDUs, distribution piping, controls, and heat-rejection systems are interconnected. Even if only a portion of the racks use RDHx or DTC cooling on day one, planning the architecture early can preserve options for future expansion.

3. Determine Where the CDU Fits

A CDU is a key component in many DTC architectures. It transfers heat between the technology cooling system serving the IT equipment and the facility-side cooling system.

During facility design, consider:

• Where will CDUs be located?
• How will coolant reach liquid-cooled rows and racks?
• How will additional capacity be added?
• How will the design provide safe maintenance access?
• What level of cooling redundancy does the application require?

The objective is not simply to reserve floor space for a CDU. The cooling-distribution architecture should be part of the data center design from the beginning.

4. Design the Rack as a Complete System

The transition from air to liquid does not stop at the CDU. Inside a high-density AI cabinet, servers may need to coexist with:

• Cooling manifolds
• Supply and return hoses
• Quick-disconnect fittings
• Multiple power distribution units (PDUs)
• Power cords
• High-density network connectivity
• Cable-management infrastructure

Together, these components create a significant space and serviceability challenge.

Design principle: Do not design cooling, power, connectivity, and the cabinet independently. Treat the rack as a complete physical-infrastructure system.

5. Account for the Remaining Air-Cooled Load

It can be tempting to frame the decision as air cooling or liquid cooling. In practice, a new data center may need both.

DTC cooling typically removes heat from the server's highest-power components, but other components may still rely on airflow. Designers should understand the expected liquid heat-capture percentage of the intended IT equipment and size the room-level air-cooling system for the remaining load.

The goal is a cooling architecture that can support the expected mix of IT equipment rather than assuming every rack will operate in the same way.

6. Engineer for Maintenance, Not Just Installation

Consider what happens after commissioning. A technician may need to replace a server in a fully populated AI rack surrounded by rear door heat exchangers, coolant hoses, quick disconnects, manifolds, network cabling, and multiple power feeds.

Serviceability should therefore be a design requirement. Evaluate:

• Access to servers and other field-replaceable equipment
• Manifold and PDU placement
• Hose and cable routing
• Component replacement procedures
• Isolation of individual servers or rack sections
• Future expansion and reconfiguration

A rack that is easy to install but difficult to maintain can create unnecessary operating risk and downtime later.

7. Design the Response to a Liquid Event Before One Happens

Introducing liquid near electronic equipment naturally raises concerns. A new facility offers the opportunity to address those concerns in the original design rather than adding controls later.

Treat liquid cooling as a managed system. Consider leak detection, flow and pressure monitoring, alarms, isolation, drainage, and maintenance procedures.

The important question is not simply, "Can a leak occur?" It is, "How will the data center detect, isolate, and respond to an abnormal condition?"

Those requirements should become part of the facility's operating and emergency-response strategy.

8. Build for Expansion

Initially, only a small portion of the facility may support liquid-cooled AI infrastructure. What happens when that requirement grows?

Consider whether the initial architecture can accommodate:

• More liquid-cooled racks
• Higher rack densities
• Additional cooling capacity
• Additional manifolds and PDUs
• New generations of servers
• Expansion into additional rows or data halls

The objective is flexibility, not overbuilding: create an architecture that can accommodate changing rack densities, cooling technologies, and capacity requirements without unnecessarily constraining future options.

Think Beyond Cooling

Liquid cooling is sometimes discussed as though it is simply a replacement for moving air through a server. For a new data center, the implications are much broader.

Compute → Power → Cooling → Cabinets → Connectivity → Operations

Early coordination among IT, facilities, electrical, network, and operations teams is essential. Treat expected compute requirements, rack densities, facility water, CDU architecture, rack-level distribution, residual air cooling, power distribution, connectivity, maintenance, and future expansion as part of the same infrastructure strategy.

Direct-to-chip cooling is more than a new way to remove heat. It changes the relationship between the server, cabinet, power, network, and facility cooling infrastructure.

Plan for tomorrow’s cooling needs today.
 Explore Panduit’s Advanced Cooling Solutions for AI and high-density data centers, and see how cooling, power, cabinets, and connectivity work together to support your evolving infrastructure needs.