Artificial intelligence is changing the way data centers are built. Just a few years ago, most facilities were designed around air cooling, raised floors, and relatively predictable server loads. Today, those assumptions no longer hold true.
Modern AI servers packed with GPUs can draw 50, 100, or even more than 200 kW per rack. At those densities, cooling is no longer just a support system. It becomes one of the most important parts of the facility’s infrastructure.
That shift is forcing owners, engineers, and operators to rethink how they approach cooling from the very beginning of a project. Instead of selecting cooling equipment after the IT design is complete, cooling infrastructure now influences everything from mechanical layouts and utility planning to expansion strategies and long-term operating costs.
The good news is that designing liquid cooling infrastructure does not have to be overwhelming. Whether you’re building a new AI data center or preparing an existing facility for higher-density workloads, the process follows a logical sequence.
This guide walks through that process step by step. Rather than focusing on individual products, we’ll look at how experienced engineers approach the design of scalable liquid cooling infrastructure that can support today’s AI workloads while remaining flexible enough for tomorrow’s.
Why Does AI Require a Different Cooling Strategy?
Traditional enterprise servers produced relatively little heat. Cooling them with computer room air conditioners and carefully managed airflow worked well for decades because the heat loads were manageable.
AI changes that equation.
The latest GPU platforms generate enormous amounts of heat in a very small footprint. As rack densities continue to climb, simply moving more cold air through the room becomes increasingly difficult. Air is still important, but it is no longer the primary method of removing heat from the most powerful servers.
Liquid changes what’s possible.
Water can absorb significantly more heat than air, making direct-to-chip cooling a practical solution for high-density AI deployments. Instead of relying on chilled air to cool every component, liquid captures heat directly where it is produced, then transfers it through the cooling infrastructure and ultimately to the facility’s heat rejection system.
This is why cooling has become an infrastructure discussion rather than simply an equipment discussion.
Choosing liquid cooling is only the beginning. The real challenge is designing an infrastructure that delivers reliable cooling today while allowing the facility to expand over the next decade.
What Should You Understand Before Designing a Data Center?
One of the most common mistakes is starting with the equipment.
People often ask questions like:
“What size CDU do I need?”
“Which cooling technology should I buy?”
Those are important questions, but they come later.
The first question should always be much simpler.
How much heat will this facility need to remove, both today and in the future?
Every design decision depends on understanding the thermal load. Before selecting any cooling equipment, answer these questions:
- What are your expected rack densities? – Designing for 30 kW racks is very different from designing for 100 kW or higher.
- How much total IT load will the facility support? – Look beyond the first deployment and consider the long-term buildout.
- How quickly will AI capacity grow? – Expansion plans should influence pipe sizing, electrical infrastructure, and equipment placement.
- Is this a new facility or a retrofit? – Existing infrastructure often creates different design constraints than greenfield projects.
- What level of redundancy is required? – Reliability goals will affect everything from pump configuration to CDU deployment.
Taking the time to answer these questions early helps prevent costly redesigns later in the project. That means looking beyond the current AI deployment โ consider expected rack densities, future GPU generations, anticipated tenant growth, and the likelihood of needing additional capacity quickly.
Many facilities begin with only a handful of AI racks before expanding into entire halls. Others are designed for a phased campus buildout over several years. Those growth plans should influence the cooling strategy from day one.
It is also important to distinguish between greenfield and retrofit projects. A new AI campus offers far more flexibility because mechanical rooms, piping, utility infrastructure, and equipment locations can all be optimized during design. Existing facilities often require designers to work within available space, electrical limitations, and existing cooling systems.
Despite those differences, the design philosophy remains the same. Start with the thermal requirements โ everything else follows from there.
How Does Modern Liquid Cooling Infrastructure Work?
One of the easiest ways to understand liquid cooling infrastructure is to follow the path of the heat.
Everything begins inside the server.
High-performance processors transfer heat into cold plates mounted directly on the CPUs and GPUs. Coolant flowing through those cold plates absorbs the heat before it can build up inside the equipment.
From there, the warmed coolant flows through the rack distribution system into the facility’s liquid-cooling infrastructure.
This is where the cooling distribution unit, often called a CDU, plays a critical role.
Rather than sending facility water directly into sensitive IT equipment, the CDU separates the building water loop from the IT cooling loop through a heat exchanger. That separation helps protect expensive computing equipment while allowing each loop to operate under its own optimal conditions.
Once the heat transfers across the CDU, it moves into the building’s cooling system, which may include cooling towers, dry coolers, chillers, or a combination of heat rejection technologies depending on the facility design.
The cooled water then begins the cycle again.
While the concept sounds straightforward, every part of that system must work together. Pump sizing, water temperatures, flow rates, redundancy, controls, and monitoring all influence overall performance.
That is why experienced designers think about the entire cooling ecosystem rather than individual pieces of equipment.
How Should You Design for Growth Instead of Today's Workloads?
Perhaps the biggest challenge in AI infrastructure is uncertainty.
Very few organizations know exactly what their AI environment will look like five years from now.
GPU technology continues to evolve. Rack densities continue to increase. New workloads appear faster than most facilities can be expanded.
Trying to predict every future requirement is impossible.
Instead, good infrastructure is designed to adapt.
That starts with modular thinking.
Rather than building one massive cooling system sized for an unknown future, many facilities are deploying infrastructure that grows alongside AI demand. Additional CDUs, pumping capacity, and distribution piping can be added as new AI halls come online without disrupting existing operations.
Planning for expansion also affects decisions that may seem minor during initial construction.
- Pipe routing should leave room for future connections.
- Mechanical rooms should be designed to allow additional equipment to be installed later.
- Electrical infrastructure should accommodate future cooling capacity.
- Isolation valves should simplify maintenance and expansion without taking the entire system offline.
These decisions may add modest costs during construction, but they can save substantial time and expense when expansion becomes necessary.
The facilities that struggle with AI growth are rarely the ones that lack cooling capacity.
More often, they lack the flexibility to add it.
What Data Center Design Decisions Have the Biggest Long-Term Impact?
Equipment specifications matter, but infrastructure decisions usually have a much greater influence on long-term performance.
Some of the most important infrastructure decisions include:
- Cooling architecture: Determine whether a centralized or distributed approach best supports the facility.
- Facility CDU placement: Position equipment to simplify maintenance while allowing room for future expansion.
- Water temperature strategy: Balance efficiency, cooling performance, and compatibility with the IT equipment.
- Heat rejection systems: Select cooling towers, dry coolers, chillers, or hybrid systems based on climate and operational goals.
- Monitoring and controls: Design for real-time visibility into temperatures, flow rates, pressures, and equipment health.
- Maintenance access: Ensure pumps, valves, heat exchangers, and controls can be serviced without disrupting critical operations.
These decisions often have a greater impact on the facility’s lifespan than the selection of any individual cooling component.
One example is determining where cooling equipment should be located. Some facilities benefit from centralized cooling infrastructure that serves multiple data halls. Others gain operational flexibility from distributing cooling capacity closer to the white space. Neither approach is universally correct โ the right answer depends on the size of the deployment, operational goals, maintenance strategy, and future expansion plans.
Water temperatures are another important consideration. Many modern AI cooling systems can operate using warmer water than traditional chilled-water systems. Higher supply temperatures may improve overall facility efficiency while reducing dependence on mechanical chilling under certain operating conditions.
Water quality is equally important. Data center liquid cooling systems require proper filtration, monitoring, and water treatment to maintain long-term reliability. Small issues that go unnoticed early can create larger maintenance challenges years later.
Instrumentation and controls deserve the same level of attention. Operators need clear visibility into temperatures, pressures, flow rates, pump status, and system performance. Better monitoring allows facilities to detect problems before they affect critical workloads.
These decisions rarely receive the same attention as GPU specifications or server hardware, yet they often determine how reliable and maintainable a cooling system becomes over its operating life.
What Are the Most Common Mistakes When Designing Data Center Liquid Cooling Infrastructure?
Most successful AI facilities avoid problems by thinking beyond the first deployment. Several design mistakes continue to appear as organizations adopt liquid cooling for the first time.
The first mistake is designing around today’s hardware. AI hardware evolves rapidly, and infrastructure designed only for current rack densities may become a limitation much sooner than expected. Building flexibility into the cooling system provides room for future technologies without requiring major reconstruction.
The second mistake is treating liquid cooling as an isolated mechanical project. Cooling infrastructure affects electrical systems, facility layouts, controls, maintenance planning, and future capacity. The most successful projects bring mechanical, electrical, IT, and operations teams together early in the design process.
The third mistake is overlooking serviceability. Eventually, every pump, valve, heat exchanger, and control component will require maintenance. Equipment should be accessible, isolated without affecting the rest of the facility, and easy for technicians to service safely.
Infrastructure that is simple to maintain is often infrastructure that remains reliable for many years.
Where Does the Facility CDU Fit Into the Overall Design?
As AI deployments become larger and more complex, the facility CDU has become one of the most important components of liquid-cooling infrastructure.
Rather than acting as a standalone piece of equipment, it serves as the connection point between the building’s cooling system and the dedicated cooling loop supporting the IT equipment.
That separation provides several advantages.
It allows each water loop to operate under conditions best suited to its purpose.
It simplifies maintenance by isolating the IT cooling system from the building infrastructure.
It also creates a modular foundation that can grow as AI capacity expands.
For organizations planning phased AI deployments, facility-scale CDUs make it possible to add cooling capacity alongside new compute infrastructure instead of redesigning the entire mechanical plant every time additional GPUs are installed.
This is the philosophy behind Nautilus’ EcoCore XCD platform.
Rather than viewing the CDU as simply another mechanical component, EcoCore XCD is designed to become part of a scalable liquid cooling architecture. Its modular approach supports phased growth, simplifies maintenance planning, and helps operators expand AI capacity without disrupting existing operations.
The result is infrastructure that evolves alongside the data center instead of becoming a bottleneck.
Why Is Cooling Infrastructure Becoming a Strategic Investment?
The conversation around AI data centers often focuses on GPUs, processors, and networking.
Those technologies are certainly important.
But none of them can reach their full potential without reliable thermal management.
Cooling infrastructure now influences how much compute can be deployed, how efficiently facilities operate, how quickly expansions can occur, and how confidently operators can adopt the next generation of AI hardware.
That makes liquid cooling infrastructure far more than a mechanical utility.
It has become a strategic asset.
Organizations that invest in scalable designs today will be better positioned to support higher rack densities, new processor generations, and evolving customer demands for years to come.
Those who design only for current requirements may find themselves rebuilding infrastructure much sooner than expected.
Final Thoughts
Designing liquid cooling infrastructure for AI data centers is no longer about selecting individual pieces of equipment. It is about creating an integrated system that connects servers, facility water, heat rejection, and controls, and that supports future expansion, all within a reliable, scalable architecture.
The best designs begin with understanding thermal loads instead of product specifications. They account for growth before it happens. They prioritize maintainability alongside performance. Most importantly, they recognize that cooling has become a core part of AI infrastructure rather than an afterthought.
As AI continues to push rack densities higher, that mindset will separate facilities that can scale with confidence from those forced into costly redesigns. A well-planned liquid cooling infrastructure gives data center operators the flexibility to support new technologies, protect their investments, and keep pace with the rapid evolution of AI computing.
At Nautilus Data Technologies, that philosophy drives the design of our liquid cooling solutions. The EcoCore XCD facility cooling distribution unit is engineered to serve as the foundation of scalable AI cooling infrastructure, helping operators deploy liquid cooling efficiently today while preparing for future expansion. Its modular design supports phased AI deployments, simplifies maintenance, and enables facilities to add cooling capacity as compute demand grows, without requiring a complete redesign of the mechanical plant.
Whether you’re planning a new AI campus or upgrading an existing data center, designing the right liquid cooling infrastructure today will have a lasting impact on performance, efficiency, and scalability. If you’re evaluating how to support higher rack densities or build a future-ready cooling strategy, the Nautilus team can help you design an infrastructure solution that meets your operational goals today and continues to support your AI roadmap for years to come.
FAQ
What is liquid cooling infrastructure in an AI data center?
Liquid cooling infrastructure is the system that removes heat from high-density AI servers using coolant instead of relying primarily on air. It includes components such as cooling distribution units (CDUs), pumps, piping, heat exchangers, controls, and heat rejection equipment that work together to keep GPUs and CPUs operating within safe temperatures.
Why are AI data centers moving from air cooling to liquid cooling?
AI servers generate far more heat than traditional enterprise hardware, with some racks exceeding 100 kW or even 200 kW. At these power densities, air cooling alone often cannot remove enough heat efficiently. Liquid cooling transfers heat directly from processors, making it a more effective solution for modern AI workloads.
What should you consider before designing liquid cooling infrastructure?
The first step is understanding your facility’s thermal requirements. Key considerations include expected rack densities, total IT load, future AI growth, redundancy requirements, available utility capacity, and whether the project is a new build or a retrofit. These factors determine how the cooling system should be designed and sized.
What role does a cooling distribution unit (CDU) play in liquid cooling?
A cooling distribution unit separates the building water loop from the IT cooling loop using a heat exchanger. This protects sensitive server equipment, maintains proper water quality and operating conditions, and allows the cooling system to expand more easily as AI capacity grows.
How can you future-proof an AI data center's cooling infrastructure?
Designing for future growth means planning for higher rack densities, modular expansion, and easier maintenance. Leaving space for additional CDUs, oversizing key infrastructure where appropriate, designing flexible piping layouts, and implementing comprehensive monitoring systems can help facilities support future generations of AI hardware without major redesigns.
