Separating the two halves of the data center cooling circuit lies the CDU – the critical component that keeps water where it belongs while transferring heat between them.

Crucially, water doesn’t come as a one-size-fits-all. The facility water system (FWS) operates as a completely separate loop from the technology cooling system (TCS), with each designed around a different application and set of requirements. The FWS typically carries standard industrial water or a specific water-glycol mixture, since its industrial piping and large heat exchangers can tolerate conventional building water chemistry.

The TCS, by contrast, requires a far more controlled and sensitive environment – an ultra-clean loop designed to route specialized coolant directly to the server racks and across CPU and GPU cold plates. The CDU helps maintain that separation while allowing heat to be transferred between the facility and technology cooling loops.

But zoom in on the cooling circuit and the importance of the valves becomes much clearer. At rack level, they provide the final layer of control and protection, allowing individual racks to be isolated quickly if an issue is detected.

That distinction reflects the different demands placed on each loop – and the increasingly granular level of control required as the cooling circuit reaches the rack. At this level, multiple components must be monitored and managed to ensure the system can respond precisely and perform as expected. Riccardo De Danieli, global data center systems application manager at Belimo, explains:

“When we look at the technical cooling system, many racks have a control valve to balance the flow. Then there are two more fast-acting on-off valves to close the rack if there’s a problem like a leakage. There’s even a leakage sensor, too.”

The power scarcity problem

Power – particularly electricity – is a resource that comes at a premium today. AI, cloud, and high-performance workloads are placing a level of pressure on grid capacity around the world that none of us could have imagined, while the infrastructure needed to meet that demand simply isn’t being built quickly enough.

Developers are racing to secure large-scale power allocations, but once that capacity is secured, they still have to make the most of what they have. The commercial imperative doesn’t then just disappear – operators need to use every available megawatt as efficiently as possible, maximizing IT output while still providing the power required to keep the supporting infrastructure – including cooling systems – running.

Every watt saved on cooling is another watt that can be directed toward the IT equipment generating the revenue.

“Liquid cooling is one of the most demanding systems in the data center with regard to power consumption,” says De Danieli.

Yet liquid cooling is also one of the most effective ways of enabling higher chip performance. According to De Danieli, the first step toward improving AI data center efficiency is to introduce liquid cooling, and once it’s in place, optimize the system itself.

Optimize to monetize

Supply temperature, differential pressure, and valve opening percentage are all important indicators of rack performance – and its return. Rack performance directly influences chip performance, and, therefore, token-generation capacity, which increasingly correlates with the revenue a facility can generate.

“A parameter that is often overlooked in the data center is the differential pressure of the rack. You can use this pressure to control the valve in order to rescale flow effectively.”

With electronic pressure-independent valves, the valve opening percentage can also provide an indication of how the pump is performing. If the valve is too far closed, the pump head is likely too high; if the valves are consistently fully open, the pump head may be too low.

“It’s crucial to get the system right because if you have a lower flow rate on each tracker, you aren’t transferring enough cooling power to each rack, which can damage the IT equipment,” says De Danieli.

The opposite can also be problematic. “If you have a higher flow rate than expected, you can accelerate wear and damage the cold plates – the small heat exchangers that are attached to the chips – and even the tubes that are connected to them. You need the right flow rate in order to prevent damage.”

But it’s not just a matter of causing damage. In an era defined by demand for maximum compute output, every component of the cooling system needs to support performance as efficiently as possible. Monitoring the right parameters can give operators a clearer picture of how the facility is behaving over time, allowing them to identify trends, intervene proactively, and optimize system performance before inefficiencies become costly.

As every component in the cooling system consumes energy, something as straightforward as precisely controlling when and how far valves open can reduce unnecessary consumption. Multiplied across a large-scale facility, those incremental gains can translate into significant reductions in cooling power – and, ultimately, more power available for revenue-generating IT.

From liquid cooling to intelligent cooling

The potential of water system optimization is not new. De Danieli points to a 2014 study of the Maui High Performance Computing Center, conducted for the US Department of Energy’s Federal Energy Management Program, which found that replacing air cooling with water cooling improved performance parameters.

“They estimated that water cooling would save $200,000 per year in operating costs. It is also very interesting that it shows the Riptide PUE – water cooling – is consistently lower than the MANA PUE – air.”

And that principle has only become more important as rack densities have increased.

Recent research is beginning to quantify the opportunity. A 2026 study by experts from the Institute for Building Energetics, Thermotechnology and Energy Storage found significant energy-saving potential from dynamically adjusting supply water temperature. Based on 21 months of operational data from a direct liquid-cooled data center, the research highlights the value of adapting the cooling set point to changing operating and environmental conditions, rather than keeping it fixed.

Better thermal management can protect high-value hardware, improve operating efficiency, and create the conditions for higher-performing IT. In other words, the rack is where the economics of cooling ultimately become tangible.

The value of valves

When discussing components for liquid cooling systems, it is easy to assume that one valve, sensor, or actuator is much like another, and that, where necessary, the components are interchangeable. So what actually differentiates a high-performance component from a standard alternative, and why does that distinction matter?

According to De Danieli, a high-performance component is not just that – it is a combination of all the products working together to form a high-performance system.

“A cooling system only works as well as its weakest component. It doesn’t make sense to have a smart valve but not a smart pump or pipe connectors – that’s why it’s important to have a holistic view of the data center.”

It is their intercommunicative capacity that leads individual items to a refined function. High-performance components do more than perform a mechanical function – they can measure, communicate, and respond. A conventional valve may be controlled, while a conventional sensor simply provides information. Intelligent components combine both capabilities, allowing operators to monitor conditions and adjust the system in real time.

This is where experience in HVAC becomes particularly relevant. Rather than developing an entirely new class of products for data centers, Belimo has adapted decades of experience in HVAC controls and equipment to the more demanding requirements of high-density liquid cooling. The result is a portfolio designed to not only control flow, but to measure what is happening within the system and communicate the information to the wider data center infrastructure.

“When you have a device that can measure the differential pressure or temperature, you need to communicate, to convey this information to the building management system (BMS) in the most effective way,” says De Danieli.

That communication architecture can have a significant impact on the complexity of the installation.

“If you have analog signals, you need a cable for each parameter. With bus communication, we have just one cable for all these devices.”

The Belimo Energy Valves, for example, combine flow measurement, differential pressure sensing and control capabilities with connectivity and cloud-based analytics. This enables operators to monitor and adjust flow at rack level, including when a server is removed for maintenance – a particularly important capability in high-density environments where cooling requirements can change as the IT load changes.

For electronic pressure-independent valves, two characteristics are particularly important – the communication protocol and the ability to measure flow directly. Rather than calculating flow from other parameters, these valves can measure it, giving operators a more accurate picture of what is happening at rack level and providing the data needed to control the cooling system more precisely.

The valve’s role has shifted. Before, it was simply a device that opens and closes to regulate flow, but now, it has become a sensing and control point within a connected cooling system.

The final calculation

The AI data centers of the future are not shaped just by how much compute they can deploy. Increasingly, they will be determined by how intelligently and efficiently they can cool it. Smarter cooling means less power spent on infrastructure, more available for compute, and ultimately, greater value from every rack.

For more information, visit belimo.com/datacenters.