By Adam Kimmel

As lower-GWP refrigerants replace legacy materials, technicians are encountering a wider range of refrigerants with different safety characteristics. What hasn’t changed is the need for safe handling practices, though the details now vary by classification.

Refrigerants are categorized under ASHRAE Standard 34 and ISO 817 based on toxicity and flammability. These classifications (A1, A2L, A2, A3, and B2L) dictate how systems are designed, installed, and serviced, and dictate appropriate tools, PPE, and procedures to ensure both technician and end user safety.

Understanding the Classification System

Two factors define refrigerant safety class: toxicity and flammability. Each factor has a hierarchy of safety precautions.

• Toxicity
• A = lower toxicity
• B = higher toxicity
• Flammability
• 1 = no flame propagation
• 2L = mildly flammable
• 2 = lower flammability
• 3 = higher flammability

The result is a combined designation, such as A1 or A3. For technicians, these labels translate directly into how job sites must be managed. 

A1 Refrigerants: Established and Non-Flammable

A1 refrigerants include many traditional HFCs and earlier refrigerant families. They are non-flammable and have lower toxicity, which has made them the baseline for decades of HVACR service work.

What A1s provide:

• No ignition risk under normal conditions
• Standard servicing tools and procedures
• Emphasis on leak prevention, recovery, and environmental compliance

While A1 refrigerants remain widely used, regulatory pressure on high-GWP products is driving a gradual shift toward mildly flammable alternatives.

A2L Refrigerants: Mild Flammability with Manageable Risk

A2L, a common classification among newer HFOs and blended refrigerants, is the intended destination of many transition strategies.

They are classified as mildly flammable, with low burning velocity and higher ignition energy requirements compared to more flammable refrigerants. A2L refrigerants don’t eliminate flammability risk; they relocate it into procedures, tools, and technician awareness.

What technicians need to do:

• Control ignition sources (sparks, open flames, non-rated electrical equipment)
• Confirm tools are rated for A2L use
• Ensure adequate ventilation during service
• Prioritize leak detection and prompt repair

A2L refrigerants are designed for use in systems that account for these risks, but servicing practices must reflect their flammability characteristics.

Figure 1: ASHRAE Standard 34 Safety Classifications and Flame Propagation

A3 Refrigerants: High Flammability, Strict Controls

A3 refrigerants, including hydrocarbons such as propane, offer very low GWP but come with higher flammability.

They differ significantly from A2Ls in how easily they ignite and how quickly flames propagate.

What technicians need to do:

• Eliminate all ignition sources, including static electricity
• Follow established charge limits
• Maintain strict work zone control
• Use only equipment approved for flammable refrigerants

Because A3 refrigerants can ignite with less energy and burn more rapidly, the margin for error is smaller.

B2L Refrigerants (Ammonia): High Performance with Higher Toxicity

Ammonia remains common in industrial refrigeration and is classified as B2L: mildly flammable but more toxic.

What technicians need to do:

• Use appropriate PPE at all times
• Monitor for leaks, as exposure risk is significant
• Follow facility safety procedures and emergency protocols

Ammonia’s strong odor aids detection, but its toxicity requires a different level of preparedness compared to A-class refrigerants.

Why A2L and A3 Behave Differently

The key technical distinction between A2L and A3 refrigerants is burning velocity and ignition energy.

• A2L refrigerants have lower burning velocity and require higher-energy ignition sources
• A3 refrigerants have higher burning velocity and can ignite from lower-energy sources, including static discharge

As a result, the difference between A2L and A3 isn’t just flammability; it’s how easily ignition occurs and how quickly a flame can spread. This difference drives how each class is handled in practice and why A2Ls are being adopted more broadly in comfort cooling applications.

What Changes for Technicians

As refrigerant classes evolve, several practical shifts are occurring across the field:

1. Training Becomes Essential

Technicians need to understand classification differences and the associated risks before working on newer systems.

2. Tool Selection Matters

Recovery machines, leak detectors, and vacuum pumps must be approved for the refrigerant classification.

3. Leak Management Is Critical

Faster detection and response reduce both safety and environmental risks.

4. System Design Plays a Larger Role

Refrigerant selection is increasingly tied to:

• Efficiency requirements
• Application conditions
• Regulatory constraints

Understand Refrigerant ‘Classification’

Refrigerant classification is not just a technicality. It defines how systems are handled in the field.

A1 systems remain familiar, but A2L and A3 refrigerants introduce new variables that must be understood and managed. For technicians, the transition comes down to recognizing the differences, using the right tools, and applying the appropriate safety practices on every job.

Adam Kimmel is Senior Manager Business Development for Orbia Fluor & Energy Materials. He specializes in the transition to lower-GWP refrigerants and supports industry training and education on safe handling and system performance. orbia-fem.com

By Andre Patenaude

Andre Patenaude is one of the commercial refrigeration industry’s most well-known and recognized experts, who you will very likely see as a presenter at industry events. And, most likely, the topic nowadays will be related to carbon dioxide (CO₂) refrigeration. As more supermarkets install carbon dioxide-charged systems, servicing contractors must be aware of the major differences between CO₂ and the man-made gasses.

NOTE: This article is not intended to replace formal CO₂ training.

RSES Journal (RJ): Andre, thank you for your time. First, how does a technician begin to charge a CO₂ system?

Andre Patenaude (AP): When charging a CO₂ refrigeration system, it is critically important for technicians to be aware of CO₂’s (R-744) “triple point,” which has a saturation pressure of 60.4 psig. If liquid refrigerant is introduced into the system at or below 60.4 psig, it will turn to dry ice, stopping all refrigerant flow in the charging line.

To prevent this, contractors need to begin charging with vapor—not liquid—until the system is well above the triple point, as recommended by the manufacturer. Note: Some manufacturers recommend charging with vapor up to 150 psig to ensure sufficient system pressure. Then, technicians can switch to liquid R-744 to complete charging the system. Charging slowly is recommended to prevent pressure from building and potentially discharging pressure-relief valves.

RJ:  How does a technician detect leaks in CO₂ systems?

AP: The natural refrigerant R-744 is derived from carbon dioxide (CO₂), which exists abundantly in our atmosphere. Although it has a global warming potential (GWP) of one, detecting and preventing system leaks is essential for maintaining reliable performance. However, since R-744 is colorless, odorless and heavier than air, it can be difficult to detect and requires robust leak detection strategies.

Leak detectors should be mounted 18 in. off the ground (well below breathing level). To pinpoint small leaks at joints or valves, technicians can use handheld leak-detection devices specifically designed for R-744 detection.

RJ: What are common CO₂ refrigerant storage precautions/requirements?

AP: CO₂ tanks are designed to handle R-744’s high pressures and weigh significantly more than standard hydrofluorocarbon (HFC) tanks. These cylinders typically weigh nearly 150 lb empty and up to 200 lb full. Because many supermarkets+ prefer to have a full system charge onsite —which could be up to 2,000 lb of refrigerant—storage would require 40 cylinders weighing up to 8,000 lb (or 4 tons).

So, it is important for contractors to understand where to store the reserve refrigerant and whether it will meet local building codes. And if tanks are stored on a mezzanine, ensure it can handle the total storage weight. Otherwise, tank storage best practices are similar to those of legacy refrigerants, including stacking procedures, safety precautions and keeping tanks chained off in a designated storage area.

RJ: What is “trapped liquid” in a CO₂ refrigeration system?

AP: “Trapped liquid” refers to liquid CO₂ confined between two closed valves or in a section of piping with no pressure relief path. R-744 has a very high thermal expansion rate with a coefficient of expansion (COE) higher than that of legacy HFC refrigerants. Because pressure can increase rapidly, systems are often equipped with appropriate pressure relief valves in locations where liquid is likely to be trapped, aiding in system operation and servicing.

RJ: What are the major safety considerations when handling, charging and servicing a supermarket CO₂ rack system?

AP: R-744 has an A1 refrigerant safety classification, meaning it is nonflammable and non-toxic. However, in enclosed spaces, R-744 leaks could accumulate to higher concentrations, displacing oxygen and posing an asphyxiation risk. As a result, proper leak detection devices should always be used in machine rooms and walk-in boxes.

Because CO₂ refrigeration systems operate at significantly higher pressures than HFC systems, technicians should always take precautions when commissioning or servicing. System pressures are highest during the supercritical (i.e., transcritical) mode of operation and when a system is shut down from a power outage or routine servicing. When a CO₂ system is shut down—especially for an extended period—standstill pressures will rise and must be handled carefully. Auxiliary backup cooling systems are often used to maintain pressures below the maximum allowable setting of pressure-relief valves of the flash tank/receiver.

Another safety consideration is the potential for dry ice formation. Because the surface temperature of dry ice is -109.3°F (-78.5°C), physical contact with solid or liquid R-744 will cause freeze burns. Protective gloves and goggles should always be worn when working with R-744.

If dry ice forms, technicians or supermarket operators should dispose of it in a well-ventilated area that is off-limits to the general public. When dry ice sublimes, it releases a significant volume of carbon dioxide, which quickly displaces oxygen and presents an asphyxiation risk in confined spaces. Thus, outside, in a protected area, is a safe space to allow dry ice to sublime.

With proper training and equipment design, CO₂ refrigeration is a safe, reliable and future-proof technology for meeting retailers’ sustainability goals.

High efficiency motors are a high priority for data center installers. Selecting the right replacement motors and drives will position contractors to help customers reduce downtime and total cost of ownership via efficiency.

By Alex Wiegmann

The rapid growth of data centers across the country is offering HVAC&R contractors and technicians an increasing number of opportunities to install and maintain equipment. While some of the install and maintenance work is similar for other commercial or industrial applications, HVAC&R specialists should be aware of the newer, unique technologies now available for data centers.

Three primary differences exist when comparing traditional HVAC&R systems to specific cooling technologies for data centers:

1. Keeping data servers cool through heat removal is the primary goal, versus comfort. This is driven by the fact that overheating of servers can lead to costly downtime, averaging $9,000 per minute.
2. System runtime is 24/7 in data centers, rather than 8–14 hours a day in commercial facilities.
3. Humidity and temperature control are precise in data centers, whereas traditional HVAC&R targets a wider comfort range.

These differences illustrate why data center operators cannot afford to cut corners when it comes to reliability and efficiency of cooling systems, particularly the motors, which are the heart of the system. A faulty motor can shut down an entire operation. In terms of energy usage, motors that run HVAC&R and other cooling equipment are responsible for approximately 40% of the power consumed in the facility.

Manufacturers are responding with advanced cooling technologies that deliver various energy-efficient solutions, whether through chilled water, refrigerants or conventional forced air.

This article provides a general overview of the technologies and terms that contractors and technicians need to know about data centers and motors, as there are some notable differences when compared to traditional commercial and industrial HVAC&R applications.

Motors are the Heart of the System
Before diving into cooling applications, it is important to emphasize that one of the biggest areas of impact at data centers is the electric motors and drives that power these technologies, delivering airflow and liquid movement. Motors power not only fans in traditional HVAC&R equipment but also pumps and compressors that dissipate heat from servers.

Because high efficiency motors are a high priority for data centers, selecting the right replacement motors and drives will position contractors to help customers mitigate risks and cost in terms of downtime and total cost of ownership via efficiency.

Conserving valuable space in the data center footprint is also a top need, so contractors who can provide flexibility on equipment size will stay ahead of the game. Motor manufacturers are supporting this trend with higher power and smaller frame sizes.

In addition, redundant or back-up motors are very common for critical equipment and are highly recommended for data centers.

Computer Room Air Handlers
Computer Room Air Handlers (CRAH) are forced air units that provide filtered air for clean rooms in data centers. NEMA Premium induction motors are commonly used in this application. Synchronous motors are also recommended as they offer the highest efficiency needed for HVAC&R. These motors also offer reliability of operation, preventing dust particles from entering the clean room, and therefore allowing the server equipment to perform at its best.

Synchronous motors include variations such as synchronous reluctance, synchronous reluctance with permanent magnet assistance and brushless permanent magnet motors. When a synchronous motor is integrated with a Variable Frequency Drive (VFD), the industry refers to those as Electronically Commutated Motors (ECM). Synchronous motors can also be paired with separate VFDs. 

The need for high-efficiency performance has increasingly led to the use of variable-speed motors. Utilizing variable speed on variable torque systems such as fans and pumps allows the equipment to match the cooling needs, saving up to 40–60% in energy costs compared to fixed-speed systems. 

Coolant Distribution Units

Coolant Distribution Units (CDU) (see photo below) are liquid-based equipment that prevent server overheating and save space in server rooms. CDUs manage the circulation, filtration and temperature of coolants for servers and processors. They provide controlled coolant for heat exchangers, direct-to-chip and immersion cooling devices.

CDUs ensure that the primary facility water remains separated from the secondary loop directly cooling the servers, allowing for precise, targeted cooling.

“In-rack” CDUs are designed for seamless integration into a server rack, whereas “in-row” CDUs are made for high-capacity, large-scale cooling.

A typical CDU commonly cools up to approximately 2 MW of processing power while consuming approximately 100 kW to provide circulation of coolants at full load. Matching demand of the process heat is the goal and the reason that efficiency is a high priority. Two different motor types are currently used on CDUs—synchronous and induction motors. To maximize efficiency at partial loads, synchronous motor usage is quickly growing.

Synchronous motors have zero slip, and therefore are synchronized with the frequency of the power supply, and the rotor has essentially zero losses. Most synchronous motors are rated as NEMA Premium 4 or higher efficiency and will consume less energy at the same power output than induction motors. At partial loads the advantages of synchronous motors can expand to as many as 15 points of efficiency. These higher efficiency motors are more complex than the asynchronous motors, typically, and as a result may cost more.

The photo below shows a synchronous motor integrated with a variable frequency drive (VFD). It’s called SynRA^® synchronous motor. 

VFDs enable the motor variable speed operation to match system demand and can be integrated (ECM) or stand alone. VFDs can also offer precise cooling control, helping to prevent hotspots that could impact server performance. Operating at variable speed can also extend motor, fan and pump life due to less wear and tear over time.

Primary Loop Pumps

Data centers require the use of large pumping units to move water throughout the facility. This is considered the primary loop, which supplies water to CDUs to cool the secondary cooling loop.

Commonly, NEMA Premium efficient induction motors being run on a VFD are utilized to save energy over fixed speed. The momentum in the industry is to take the next step to variable speed synchronous motors which in addition to saving energy are often smaller, lower weight, and run cooler.

Traditional HVAC&R

Heating, cooling and ventilation of data center facilities requires the highest efficiency available to reduce the already considerable drain these operations place on the energy grid. A trend toward synchronous motors, VFDs and other energy-efficient technologies has been under way for some time, maximizing efficiency for the 24/7 operating environment.

These products use the same motor technology as traditional HVAC&R—permanent magnet motors, synchronous motors and integrated ECMs but in a slightly different mechanical package to meet lower horsepower ranges.

Fire Pumps

Like all commercial and industrial facilities, fire protection equipment is required in data centers. HVAC&R and plumbing contractors and technicians who install this equipment know that horizontal and vertical fire pumps deliver dependable, high-efficiency water movement, ensuring compliance with safety standards while safeguarding facility infrastructure.

Induction motors are generally used in fire pumps, rather than the energy-saving variable drive motors used in other data center applications. The straightforward design has a simple on/off switch that is activated when needed. While variable drive motors are not required, the use of VFDs is becoming more common to limit water hammer (a sudden start or stop in water flow) and control the system. 

Overall, no matter the application, contractors are advised to avoid the low-priced motor options, which are often unreliable. When millions of dollars of computing equipment are at stake, relying on a well-known, established motor is highly recommended.

Reliable, high-efficiency motors and drives will help data center facilities keep energy costs down and reduce equipment downtime. Having this basic understanding of motor and drive technologies will position contractors to offer the best solutions to meet their customers’ goals.

Alex Wiegmann is the Director of Marketing, HVAC at Nidec Motor Corporation.

Helping Industry Become Data Center Savvy

Data center construction ties in a wide array of code requirements. The ICC has started a process to provide guidelines to supplement—but not replace—existing codes.

By Terry McIver

The International Code Council (ICC), headquartered in Washington, DC, is the leading global source of model codes and standards and building safety solutions that include product evaluation, accreditation, technology, training and certification. ICC’s codes, standards, and solutions are used to ensure safe, affordable, and sustainable communities and buildings worldwide.  

The vast, worldwide growth of data center construction has given ICC a new challenge. ICC is currently developing its “G-12 Guideline on Data Centers,” to provide a clear, comprehensive and easy‑to‑navigate framework that aligns the most relevant code provisions for modern data center design and construction. As data centers evolve in scale, operational complexity and criticality, this guideline will bring together key requirements from multiple disciplines, electrical, mechanical, fire protection, structural, water efficiency and more, into a single, cohesive resource tailored specifically to these unique facilities.

The ICC team will integrate applicable codes and standards into its document to highlight how they relate within the context of data center operations. The completed G-12 standard will enhance present-day understanding of data center requirements for building officials, designers and developers. It will offer a clear pathway for achieving safety, reliability and sustainability in the built environment, supporting the development of data centers that meet today’s performance needs while preparing for tomorrow’s technological advancements.

As it begins what will be a deliberately methodical process, ICC is gathering information from existing codes and standards to avoid conflicts in standards language. To gain a better understanding of ICC’s role in this sector, RSES Journal interviewed Gabriel Maser, Executive Vice President of Innovation and Growth for the International Code Council, and Kevin McOsker, Vice President of Technical Resources with the International Code Council. Maser works across the organization to ensure products and services continue to evolve with emerging technologies, industry trends and member needs, while McOsker provides technical assistance with building codes and building department operations for the Innovation & Growth Department of ICC. 

“ICC wants to be sure that our codes and standards meet the demands of our industry in the United States and abroad. And, it’s no surprise that there’s been a dramatic increase in data center construction,” said Maser, “[T]his is a type of construction . . . with many unique attributes.”

Data centers are indeed a complex and unique technology, but they are also having a major cultural impact across the US. The complexity and size of many facilities now in operation, and those in the planning or construction phase, have earned accolades as well as scorn. Accolades from those inside the industry, and scorn from public interest groups who fear what data centers will do to communities in which they are located. In some regions, voters have been successful in voting down proposals to build in their backyards.

For its part, ICC may be able to help quell some emotions, as it focuses on what is required for safe and efficient data center construction and site management.

“There are a number of decisions [related to data center construction] that are left to the building safety official, and we want to land on some consensus best practices, for the government officials side, and for industry, so that both sides are well aware of the rules of the road, how to classify it and how to characterize it, because those decisions have second order effects on how you manage safety in the building,” said Maser.    

ICC’s membership is comprised of what is just about an equal split between public and private sector building industry professionals. The public sector side includes building officials, inspectors, plan reviewers, public works officers, energy code officials, fire code officials, sustainability directors and others with similar administrative titles. On the private sector side are builders, engineers, architects and materials manufacturers.

For that wide universe of building professionals, ICC realizes each has already been saddled with many other codes and standards, and it in no way wants to tread on what is currently in place.

McOsker said a preliminary data center standards framework has been developed and presented to members of the ICC board of directors, as a first step in formulating the technical language for G-12.

For the current wide universe of building professionals, ICC realizes each has already been saddled with many other codes and standards, and it in no way wants to tread on what is currently in place.

“It contains some broader topics we’re going to look at, like physical security and site control, occupancy and building issues, like area height, allowable area, allowable heights, passive and active fire protection, interior environmental conditions like cooling and heating, as well as structural, electrical and environmental. We’re going to facilitate this process so that we can identify all of our code applications and identify any ambiguity, so that our customers, whether they’re governmental members on the building side or fire side or energy side, and our private sector members—the architects, engineers and contractors—have a tool they can use [for] this new type of building application that didn’t exist maybe 10 or 20 years ago,” McOsker said.

The modern data center era began around 2006 with the advent of cloud computing, which led to the need for “hyperscale” data centers.

McOsker said the combination of data centers, AI and the huge advances in electronic information processing have brought the industry to a “technology threshold.”

“It’s like the Interstate Highway System, Industrial Revolution or the railroads,” he said. “They’re building these hyperscale projects all over the country . . . having a tool for the broad range of code officials and jurisdictions throughout the country will be a huge benefit to all ICC members. Some will be savvy on the topic, while others will be learning about it.”

The G-12 standard for data centers will not become anything like a “Ten Commandments of Data Center Construction.”

“We will stay out of any space for which there is an existing standard, where experts have worked on those applications. The guideline, in and of itself, will not be written in mandatory language, which gives us a lot more flexibility in providing things like best practices and lessons learned.”

McOsker said ICC welcomes the participation of HVAC&R contractors who would like to become involved in the development of these guidelines.

“They can be involved in meetings as interested parties, become active in the conversation and provide comments as well. There are many areas in data centers that are impacted by cooling; [center owners are] also doing some very unique things with how they want to cool the equipment.”

HVAC&R contractors interested in participating in the development of the G-12 standard are asked to sign up as interested parties through ICC’s committee application process. Visit https://bit.ly/icc-committee-application locate the G12 – Data Center Guideline [AD1] option on the committee list.  

A dedicated webpage at https://bit.ly/icc-guidelines is available to follow the progress of the guideline as it proceeds through the development process.

HVAC&R for AI Data Centers: What Contractors and Facility Managers Should Consider

AI-driven data centers demand advanced HVAC&R systems to handle the immense heat output from high-performance computing and liquid cooling technologies. Facility owners must prioritize energy efficiency, redundancy, and precise temperature control to maintain uptime and optimize costs. 

By Donnelly Mechanical

As artificial intelligence (AI) continues to shape industries and redefine technology, the demand for high-performance computing (HPC) infrastructure is soaring. AI data centers are at the heart of this evolution—housing powerful processors that run complex machine learning models and data analytics at lightning speed. But with that performance comes a major challenge: heat.

Traditional air-based cooling methods are no longer sufficient to keep up with the heat output of modern AI workloads. To maintain reliability, efficiency and uptime, facility owners are turning to advanced HVAC&R and cooling strategies, especially liquid cooling systems, like direct-to-chip cooling, to future-proof their data centers.

The Growing Demand for Efficient AI Data Center Cooling

AI-driven workloads have dramatically changed how data centers operate. Rack densities that once averaged around 5–10 kW are now reaching 20 kW or more, with HPC and AI racks quickly approaching 50 kW. These higher densities generate unprecedented heat levels that push traditional HVAC&R systems to their limits.

Moreover, there is increasing global pressure on data centers to reduce energy use and carbon emissions. Since cooling can account for up to 40% of a data center’s total energy bill, improving thermal management is no longer optional—it is essential.

For facility owners, the challenge is balancing three key priorities:

• Precision—Maintaining consistent temperatures at the chip and rack level.
• Scalability—Supporting future hardware without major infrastructure overhauls.
• Sustainability—Lowering energy use and operating costs while meeting environmental, social and governance (ESG) goals.

The Rise of Liquid Cooling in AI Data Centers

To meet the growing demands of AI, many operators are turning to liquid cooling, a method that uses water or dielectric fluids to dissipate heat directly from server components. Liquid cooling is roughly 3,000 times more effective than air alone for removing heat from high-density computing environments.

Data centers are adopting one of three strategies:

• Building new, fully liquid-cooled facilities;

• Retrofitting existing air-cooled spaces to support liquid cooling; and

• Integrating hybrid cooling systems that combine both air and liquid methods.

Most organizations opt for the hybrid model, as it offers a balanced path to higher capacity, faster deployment, and a better return on investment.

Direct-to-chip Cooling: Precision at the Source

Among liquid cooling methods, direct-to-chip cooling stands out as one of the most efficient and scalable solutions.¹ This technology circulates a dielectric coolant directly over the surfaces of processors through specialized cold plates, efficiently removing heat right where it is generated.

By targeting the heat source (the chips themselves), direct-to-chip cooling:

• maintains optimal operating temperatures under any workload;

• reduces the risk of overheating and downtime;

• supports higher rack densities without compromising performance; and

• improves system reliability and extends hardware lifespan.

This approach is particularly valuable for AI and machine learning systems, where computational intensity can fluctuate dramatically depending on training or inference loads.

Components and Functionality

A typical direct-to-chip cooling system includes:

• Coolant Distribution Unit (CDU)—Manages flow, pressure and temperature between racks and facility systems.
• Dielectric coolant—A non-conductive liquid designed for safe, efficient heat transfer.
• Cold plates—Metal interfaces attached to CPUs and graphic processing units (GPU) where coolant flows.
• Circulating tubes and pumps—Moves coolant through the system.
• Thermal interface material—Conducts heat from chips to cold plates.

Direct-to-chip systems can operate in single-phase (liquid remains in one state) or two-phase (liquid evaporates and recondenses) configurations. Both offer high efficiency, but two-phase systems provide superior heat removal for ultra-dense setups.

Integrating Direct-to-Chip Cooling into Existing Systems

For most existing data centers, transitioning to liquid cooling does not require starting from scratch. Facilities can integrate liquid-to-liquid, liquid-to-air or liquid-to-refrigerant CDUs into their HVAC&R systems to establish a dedicated cooling loop for AI racks.

• Liquid-to-liquid CDUs isolate the coolant loop from the main chilled water system, offering flexibility and control.
• Liquid-to-air CDUs work without chilled water access, using existing air-cooled units for heat rejection.
• Liquid-to-Refrigerant CDUs connect directly to refrigerant-based systems, maximizing modularity and speed of deployment.

This flexibility allows operators to incrementally add capacity, manage risk and tailor their approach based on available infrastructure and cooling loads.

Partnering with HVAC&R Experts for AI Readiness

Implementing advanced cooling technologies like direct-to-chip cooling requires expertise in both mechanical systems and data center operations.

Facility owners design, maintain and upgrade their HVAC&R systems to meet the evolving demands of AI data centers. Customized maintenance and service agreements should be built to minimize downtime, maximize efficiency and ensure smooth transitions from traditional to next-generation cooling systems.

A data center cooling contractor must provide the experience and precision needed to deliver reliable, energy-efficient performance, without disrupting critical operations.

Building the Future of AI Infrastructure

The shift toward AI is not slowing down, and neither is the need for smarter, more sustainable cooling. With innovations like direct-to-chip cooling, facility owners can confidently scale their operations, protect their hardware investments, and reduce energy waste.

This article was originally published online by Donnelly Mechanical, a premier provider of HVAC&R service, maintenance and construction based in New York City, and is reprinted here with permission. RSES Journal appreciates Donnelly Mechanical’s willingness to share this information with the HVAC&R industry.

Find the original article here on Google:

Direct link:
https://donnellymech.com/blog/commercial-hvac/hvac-for-ai-data-centers-what-facility-owners-should-consider/

¹ Vertiv.com: “Must-read research: It’s the data center economy, stupid!
Or is it?”
² National Laboratories of the Rockies (NLR.gov/news): “Reducing Data
Center Peak Cooling Demand and Energy Costs with Underground Thermal
Energy Storage.” 
³ Understanding direct-to-chip cooling in HPC infrastructure: A deep dive into liquid cooling

Frequently Asked Questions

What is direct-to-chip cooling?

Direct-to-Chip Cooling is a liquid cooling method that circulates coolant directly over processor chips using cold plates. This removes heat more efficiently than air cooling and supports higher rack densities common in AI data centers.

Why are data centers adopting liquid cooling?

AI workloads generate far more heat than traditional computing. Liquid cooling offers superior thermal transfer, lowers energy consumption and enables data centers to run more powerful servers within smaller spaces.

Can existing data centers be upgraded to use direct-to-chip cooling?

Yes. Facilities can retrofit existing HVAC&R systems using liquid-to-liquid, liquid-to-air or liquid-to-refrigerant CDUs. This allows for hybrid setups that integrate with existing air-cooled infrastructure.

What are the main benefits of liquid cooling for facility owners?

Liquid cooling enhances energy efficiency, system reliability and scalability while reducing total cost of ownership. It also helps organizations meet sustainability targets by cutting power usage.

To take on data center cooling, HVAC&R contractors must be up-to-speed on the nature of these projects and the technical and project management skills they require.

Compiled & Edited by Terry McIver

The growth of the U.S. data center market is arguably a phenomenon unlike any other in the history of technology, including copy machines, cable TV and cell phones. The PEW Research Center reported that spending on data centers for artificial intelligence has overtaken consumer spending as a share of U.S. GDP growth. In 2025, data centers in North America delivered a 36% increase in capacity. Demand for data centers continues to grow and is projected by PEW to require a 50% increase in power demand by 2027, and as much as 165% by 2030.

Data centers are “mission critical” projects with high service margins for commercial HVAC&R contractors with data center experience in installing computer room air conditioners (CRAC) and computer room air handlers (CRAH) but also experience servicing both types. 

Artificial Intelligence (AI) is pouring gasoline on this fire. The latest industry analysis shows that global data center capacity has grown five-fold over the past two decades, reaching 114 gigawatts as of 2025.

“That kind of growth doesn’t just strain infrastructure, it transforms it. Cooling strategies are being reinvented. BMS architectures are being reimagined, and the next generation of HVAC&R is being forced to evolve faster than any of us have ever seen. Even here at home in the U.S. market, it’s forecasted to grow by $276 billion by 2024, an average CAGR of about 10.4%,” said Ryan Sen, Vice President, Business Development, Distech Controls, as he moderated a data center panel discussion during the 2026 AHR Exposition in Las Vegas, NV.

“There has never been a more intense, fast-moving consequential moment in the data center world than right now. We’re living in a period of unprecedented acceleration.” - Ryan Sen

The qualified HVAC&R contractor serves as a consultant to the data center owner, not just a repairman. In this role, he or she can be the owner’s first contact when seeking advice on installing, servicing and expanding the data center. Specialized HVAC&R contractors will design a system that fits the server facility’s density, to ensure precise control of temperature and humidity. They will know which system is best for the situation: CRAC units for smaller loads or chilled water CRAH units that are the best choice for larger environments.

They must know about “redundancy,” which is basically a back-up cooling system that ensures continuity of cooling. They must design the hot/cold aisle containment system, which is a major element in cooling efficiency. Then there are constant retrofits and upgrades, because data center cooling technology and capabilities are constantly changing.

An HVAC&R contractor seeking to enter the data center cooling market must do so with eyes wide open and a clear understanding of these requirements. A major caveat is that they must complete the cooling project on time, because the owner must have the facility operational at the exact time stated in the construction contract.

“There has never been a more intense, fast-moving consequential moment in the data center world than right now. We’re living in a period of unprecedented acceleration,” said Sen.

The 2026 AHR Expo panel on “Data Center Demands” described the harsh realities behind the growth of data centers and data center cooling: what is working, what is breaking and what is coming. The panelists were:

• Ken Duncan, Lead Data Center Application Consultant, Belimo Americas’;
• Davin Sandhu, Global Portfolio Director for Data Centers, Johnson Controls;
• Justin Seter, Strategic Initiatives Officer, DLB Associates and Fluid2Chip;
• David Quirk, CEO, DLB Associates and Fluid2Chip; and
• Ryan Sen, VP Business Development, Distech Controls, moderator/inquisitor.

–Terry McIver[AD1]

What are the biggest cooling challenges emerging as rack densities continue to rise? 

David Quirk (DQ): The biggest challenge is how to keep all this stuff cool. As we go to higher densities on a per-rack basis, we still must cool with air cooling, but we’ve added another cooling loop in the mix. We must do liquid cooling to the chips and air cooling. And we have this dichotomy going on between these two cooling systems. Right now, we’re putting them on combined piping systems—what I refer to as a two-pipe system—and we’re quickly diverging into four-pipe land, and probably into six pipe land.

That means a four-pipe cooling system on the facility water system, and then a separate technology cooling system (TCS) loop for the liquid-to-chip that may be operating at a different temperature. NVIDIA announced they’re pushing to a higher temperature of up to 45°C for the supply or the inlet on the chips, so we’re going to see a divergence of the temperature classes, how to cool the equipment, and enabling other forms of heat recovery in the mix. But we’re also seeing a lot of higher density in air cooling. So, where we used to use fan walls, crawl galleries and other methods to push a lot of air down the data center aisles, with the average cloud site about 10 kilowatts of rack right now, we’re going to be blowing far past that. Think 20-, 30-, 40-, 50-kilowatts and higher of air cooling per rack, even in a liquid cooled environment.

Meanwhile, we must reject all this heat, so, if we went from 10-kilowatts-per-rack up to 100-kilowatts-per-rack, now it’s going to 500- to 600-kilowatts per rack with the Ruben architecture that NVIDIA just announced. And so, when you look at a megawatt rack, when the level of density increases at the rack level inside the building, you must ask, ‘Where’s all the heat going outside?’ So, we went from 10x now we’re talking like up to 100x. We must get rid of all that heat outside and we already shot ourselves in the foot now because we went all closed-loop cooling, with air-cooled chillers located outside. So, we need even more surface area to reject the heat. The challenges are literally on all fronts.

Davin Sandhu (DS): It’s going to come down to the ability to be flexible into the future. If you have all those different rack densities, how do you build something today that is ready for that future state, as you get denser racks? Where do you put that? How do you ensure that you have the right view all the way through, that you’re building the infrastructure today that can handle every one of those? And so, as you’re looking at this, each hyperscale data center, each colocation has to deal with that paradigm and see which way they want to proceed.

It’s almost like the CDU becomes so instrumental in the design of what that part hits and where you want to put that. And from there, which type of loops do you have serving that? Is it a two-, four- or six-pipe? Because there are different ways you solve that. And then, do you do a single loop or a bifurcated loop? There’s no single way to solve it, and there are probably many good designs out there. The other part of solving this is not just getting the thermal out, it’s also delivering that to the customer, because they want to have their ‘speed to tokens’. They need capacity to build as quickly as possible, using a solution that’s consistent and reliable, that gets them where they want to be as fast as possible. So, there are many different technical challenges, but we’re continuing to solve them all.

Ken Duncan (KD): We have three components in a liquid-cooled system: the facility water system; your chillers; and this ‘magical’ coolant distribution unit (CDU) unit that sits in the middle of it, and very few people know what it does. And then, you have the technology cooling loop. They’re all separate from each other, but they’re all intertwined, and not just on the mechanical side of it, but on the ownership side.

If you are a colocation owner who maintains the CDU, there are many ownership issues along with the mechanical side of it. We’re building so fast that, sometimes, some of the consultants don’t have time to future-proof the system. Belimo is [providing] a simple valve in the system, but if it doesn’t work, there are major implications. If you can load power into a field device, that will do its own thing autonomously, the overall system is more resilient. There are so many facets to it and it moves so fast; every week there’s something new.

Justin Seter (JS): To keep us on the same page a 10-kilowatt cloud computer rack that we’ve been ‘phoning in’ the design for from 2013 to 2023 puts out about as much heat as a residential fireplace, like what goes up the chimney. You could use air cooling to move enough CFM across that rack to cool it.

So now, if you have a 100-kilowatt rack in the same footprint, that’s like 10 fireplaces. So, you have to get the heat of two of them out by air and the other eight cooled by water. The megawatt rack is 100 fireplaces in the same footprint.

"The [nation’s] grid is pretty much out of power. We’ve got maybe 24 months, after which there’s not really going to be very many new projects starts that will be able to tap into the grid." - Justin Seter 

There is a data center campus just announced in Las Vegas recently that projects they might need between eight to 12 gigawatts of power over 10 buildings, or some similar big number. The Los Angeles metro region uses 12 gigawatts. The [nation’s] grid is pretty much out of power. We’ve got maybe 24 months, after which there’s not really going to be very many new projects starts that will be able to tap into the grid. They’ll go find whatever pockets of power they can. But after that, most projects that are on this gigawatt scale are going to build their own natural gas power plants and the data center will essentially be the utility. That’s the kind of broad landscape of where we are.

The other big number is, there is a huge opportunity here: how many people’s business development teams have attended this conference to see if they can get into data centers. So, for every gigawatt deal that gets announced, think of the 3,000 to 7,000 trade workers we need to build EVERY gigawatt.

           How are liquid cooling and hybrid cooling architectures evolving and where do you see the next major efficiency gains coming from?

JS:      The data center industry is scared out of its wits to use water. There’s been this backlash in the news about water usage, and the most efficient plants that we can use to build data centers are from water-cooled chillers. So, the industry’s gone almost completely to air-cooled chillers. So, you take an efficiency hit there, but there are other things you can do to mitigate some of that. But I think as the temperatures go higher on some of the IT, the software is what will drive the building design. For example, NVIDIA says ‘[AD2] we can use 45°C for this one thing that we need. But then you’ve got all the rest of the building that’s going to be at some lower temperature. And, if you’re a wholesale co-developer, how will you know what computers are going to roll into the space when you’re signing the deal two years before it’s energized? So, basically, there’s no safety factor left on the IT application or the hardware associated with that.

Historically, there has been a lot of conservatism and safety factors built into the designs, all the way back through the chiller plants, over-provisioning capacity, redundancy and all of that type of stuff. Those margins are getting tighter. I would be so bold as to say they’ve essentially evaporated.

KD:      We talk about thermal storage, and we have thermal storage at the chiller level to help meet instantaneous demands. And now, we’re putting storage tanks on the technical cooling loop. We have thermal storage tanks there to help offset that, because we can’t fully stroke a control valve in one second without causing the CDU pain, because it has a variable frequency drive in it, and it can’t change the speed fast enough. And the worst thing you want to do is to have one valve change position, knock a CDU offline and then all the rest of the racks by that CDU don’t have water flow.

We must understand what’s going on in the process. Do we define this as an HVAC&R process or as an industrial cooling process for code requirements? I don’t know how anybody could consider this to be an HVAC&R process, because we’re trying to keep chips from melting down. It’s an industrial process. But to go a little bit farther on Justin’s numbers, a rack the size of a home refrigerator would require a 34-ton air handler you physically can’t put in there; and you can’t blow 4,000 CF across your home refrigerator to keep it cool. And that’s where liquid cooling comes in. But then you have to worry about fault detection, and what happens if you lose power? Or what happens if the CDU can’t achieve the pressure? What happens when you’re filtering down to 20 microns and the filters in the CDU get plugged? You need to know that immediately. Part of my job is to distribute that intelligence out through the building so that we can gather that information in real time and avert disaster before it happens. We have to be preventive and know when things aren’t happening like they should be and catch it early.

DS:      As we talk about our TCS loop and how you want to handle it, and the rapid changes that the GPUs might be going through, you’re going to see whether it’s through inference or training, however those GPUs are reacting, and they’re going to get very hot very quickly, and then come back off…and how do you advance that rapid change? Do you have the right thermal loop from a thermal mass standpoint? How do you keep all of that together?

You talk about the difference between liquid cooling and hybrid cooling, where you’re doing a bit of air and a bit of liquid. NVIDIA wants to go one hundred percent liquid-cooled. Is that possible? You still have all the network gear that must be operated as well.

The next stage of efficiency gains is, how do you put all of those sensors and different systems together and have an intelligent view [of the system] so that you can predict what the GPUs might need from a power requirement of from a thermal view? And then be able to react and control that loop efficiently and make sure that you’re actually using the thermal storage effectively and you’re actually running it at the right paradigm of your efficiency gain. It’s ensuring that you have full visibility of all the different sectors within your data center and that you’re taking a broad view and ensuring that you’re operating at the different individual portions within the data center to really be most effective. It’s to make sure you’re working in concert and the system is operating as efficiently as possible. You can put a great system in place but if you’re not operating it efficiently, if you just let it go full bore, your Power over Ethernet (POE) budget goes through the roof. Then, your data center, ‘colo’ or hyper owners are writing big checks.

DQ:     I don’t think the bulk of the industry has their arms around what we’re dealing with today, never mind what’s coming a week or a year from now. But as an industry, we’re way behind the eight ball. We don’t have the standard of care established for most of what we’re doing right now. We went to a very high-quality process—for example, a water loop for the TCS system, and we must filter these systems down to below 25 microns, which is the width of a human hair, and that’s difficult to do. You can’t use the same trades person that does chilled water systems doing a process system. You wouldn’t do it in pharmaceutical or food and beverage. They’re as different as night and day and we don’t have an application specific training in place, with standards, protocols, operations processes to handle the TCS loops right now.

We’re working quickly through groups, like OCP [Open Compute Project] and ASHRAE, to solve for all of that. But it’s moving incredibly fast. By the time we publish something, it’s all changed again because now we’re also in a state of divergence on criteria. We’re widening the acceptable operating envelopes, introducing new fluids and changing the criteria for all the refrigerants. We’ve decided that we can’t use water now and we don’t really have a system established. I’ll call it a ‘publicity approach’ to the market for how to get back to responsible water use and replenishment. So, when you have no standard of care for what you’re doing, and you’re trying to move fast, the criticality of the process is high, meaning there’s no safety factor. Changes must literally be at the speed of light, and you’re now introducing lots of new technology and governance, and regulations are closing in on the walls all at once. Plus, the newness of it all is incredibly challenging to anyone who has started working in the data center industry.

           How are Building Management System (BMS) platforms evolving to meet the demands for hyperscale[AD3] and edge data centers?

DS:     BMS platforms are evolving in multiple ways. Once upon a time, we tried to be all DDC [direct digital control] controls. We started with pneumatic controls, which was something that mechanical engineers understood well. Then we went electronic, and then to DDC. Now we’ve gone to Operational Technology (OT), very heavily software driven and very network oriented. We’ve gone to a completely different skills set than what the mechanical engineer knows and was trained in.

Data centers came along, and [owners] wanted a higher standard of operator, as any mistake can cost a lot of money. Then, we layered on top of that a lot of data that they need, fast. So, we started pushing the limits of DDC system networks, their protocols and point counts within some of the devices.

So, you layer all these pieces, and what we’re seeing today in the data center industry is that most clients have shifted to a Programmable Logic Controller (PLC) platform so they could get a faster speed of reaction on the software and networks. They’ve shifted to something like a Supervisory Control and Data Acquisition (SCADA) platform to get high point counts and interoperability of the GUI platforms and databases.

Moving forward, the increased density has exaggerated that problem to another level. A data center is more like a factory now and you need controls platforms that can respond to that. It’s not that we can’t still use DDC to do portions of that, because we do; and DDC can’t evolve to solve some of the problems, but the industry needs to come to the mindset that those are absolute must-haves now. You must do things in milliseconds across the board on everything in a data center at this point.

As you think across the data center, you have a lot of different points, and you need to be able to react quickly. The changes are rapid and you must have the right platform that can take all the information in, react quickly and translate it. That’s where you look at the person who is installing it and the person who is driving the engineering and putting all those sequences of operations together, ensuring that it fits and tests. You go through five to six levels of commissioning to ensure it all fits and works the way it needs to, because downtime is not an option.

The other part is, with some of that advancement and adaption, is the CDU. Some people think a CDU is just a plate-and-frame heat exchanger with a couple of pumps. And when you think about it from a control standpoint, that becomes your demarcation into the rack, into the white space, into the factory that you’re building for.

That’s where the criticality of having the right controller, the right system and the right information that you’re getting from the TCS loop comes into play; and ensuring you can react and handle it and drive it forward in a true automation across the entire thermal chain. That you can provide assurance that you know what’s going on and you can react and can ensure that the data center does not have an option to go down. That’s where it is going to continue to evolve.

            How do you balance cost redundancy and operational simplicity?

DQ:     You don’t. While all of this chaotic change is happening in the industry, one thing has taken precedence over everything else, and has driven a lot of bad behavior that we see out there, and it is the schedule. The person who signs a wholesale lease agreement and the developer salesperson, who is not a technical person, will ask, ‘Can you guys get this thing online in 14 months? Okay, cool.’ 

So, in 14 months you must get it online or the deal starts to fall apart. The basic economics of how much rent you can charge per kilowatt, or token, or however they do the deals, start to fall apart. And it’s not optional. You must be done by then, and that ends up trumping a whole lot of stuff and allows project teams to try to rationalize their way out of oversimplifying something that’s really complicated. So, if you don’t have your trusted advisor team in place who’s able to pump the brakes and do the right thing by taking a little extra time, you will find time to do it twice if you can’t find time to do it once the right way. The speed-to-market is a huge thing.

What I also see is the complexity of the non-disclosure agreement. Nobody can talk about anything. So, it’s just a matter of understanding what you’re getting into and taking the time to figure it out up-front. There isn’t time to redo anything in a data center. You’ve got to get it right the first time.

Politics aside, data center development in the US is expanding beyond expectations. The cooling systems they require are themselves a specialized service, with a big learning curve.

By Terry McIver, Editor, RSES Journal

Data center development has rapidly accelerated with no signs of slowing down. This sustained growth presents substantial opportunities for consulting specifying engineers and commercial HVAC&R technicians to apply their expertise across increasingly specialized applications. As computing technologies continue to evolve, data center cooling solutions are also advancing in parallel to meet the high-performance demands necessary to support next-generation infrastructure.

RSES Journal asked Davin Sandhu, Global Portfolio Director for Data Centers, Johnson Controls, to describe some of the basic systems related to data center cooling, and where technicians can start to learn more about it.

“Many existing data centers were intended to store information such as what is used by Netflix or Spotify,” Sandhu said. “These facilities were typically designed to carry power loads of 5–10 kilowatts per rack. Traditionally, two types of air-cooling systems were most often used within these facilities: CRAC [computer room air conditioning] and CRAH [computer room air handling]. CRAHs are referred to as fan coil walls or vertically focused air handling units that have a large, cold-water coil fed from the chilled water plant. CRACs are used in electrical rooms. The larger computer rooms are using CRAHs more often.”

Sandhu explains how newer data centers are being designed for 50–125 kilowatts per rack to meet AI-powered workloads. As these IT densities increase, so do ambient thermal conditions, meaning there is more heat to get rid of. As a result, many data centers are moving toward single-phase direct-to-chip cooling.

Understanding holistic goals and operating requirements of the data center is also necessary when designing cooling systems. Water management, pump maintenance, valve calibration and air distribution must all be considered. 

“In IT terms, this is considered ‘liquid cooling,’ but no liquid ever touches the chips,” Sandhu said. “Instead, the liquid is contained in a fully enclosed loop. It cools the general components touching the chip, but not the chip itself, using a cold plate that is integrated directly where the heat is generated. This allows it to quickly pull the heat away.”

As data center leaders consider direct-to-chip cooling, Coolant Distribution Units (CDU) can be integrated, creating a “chip-to-atmosphere” hybrid solution that can be scaled alongside growing thermal demands. The CDU is integrated within the rack row to maintain precise temperatures around IT equipment. The rejected heat load is then transferred to the facility’s main cooling system, such as a chiller.

Chillers are often central to data center cooling infrastructures. This can be an open system, like evaporative cooling that uses a cooling tower or a closed loop system, which leverage dry coolers or air-cooled chillers. For many data center developers, air-cooled chillers provide essential benefits by conserving water because they do not require water evaporation or consumption. Water is pumped into a closed loop one time and then the closed loop recirculates the cooling fluid and heat is rejected into the air without evaporation.

“Some advanced air-cooled chillers are now being engineered with exceptionally wide operating ranges to meet growing cooling loads—as much as 20°F–up to 130°F ambient temperature while producing chilled water temperatures as high as 85°F,” Sandhu said. “This provides the flexibility necessary to ensure the cooling equipment installed today can still meet the demands of the future.”

Essential Learning Resources

Sandhu emphasized that understanding holistic goals and operating requirements of the data center is also necessary when designing cooling systems. Water management, pump maintenance, valve calibration and air distribution must all be considered. Training can be obtained through participation in ASHRAE Technical Committee 9.9 (Mission Critical Facilities, Data Centers, Technology Spaces, and Electronic Equipment) as the industry's foremost unbiased authority on data center HVAC&R and environmental standards. It establishes the foundational guidelines for temperature, humidity and airflow that major IT equipment manufacturers use to validate hardware warranties.

Technical committees (TCs) develop and sponsor technical sessions at the winter and annual conferences. Information about their future technical program is discussed at each TC meeting and at TC’s Program Subcommittee meetings. ASHRAE publishes papers and transactions from presentations at its conference events. In addition, ASHRAE records most of the seminar sessions from its conferences on DVD. These DVDs are ideal for use at chapter meetings, in university courses or company lunch-and-learns.

Data Center Dynamics is a UK-based events and media company that provides insight and networking opportunities for professionals at work in the data center scale IT infrastructure. The platform’s Academy series provides resources related to data center language and systems and how they relate to each other,” Sandhu said. [AD1] “Data Center Dynamics is not a full certification but it’s great knowledge as far as providing viewpoints of critical systems and components.”

Uptime Institute serves all stakeholders responsible for IT service availability through industry leading standards, education, network, consulting, and award programs delivered to enterprise organizations and third-party operators, manufacturers and providers. It has created and now administers Tier Standards & Certifications for Data Center Design, Construction and Operational Sustainability, and offers accredited infrastructure training programs. “Tiers” refers to the four levels of data center complexity, each with its own set of standards for design and sustainability and “redundancy,” which is basically the level of backup systems in place for a data center.

“At Johnson Controls, the global reach of our Data Center Solutions team allows us to partner with a network of contractors and consulting engineers to examine equipment and participate in hands-on training opportunities. Together, we can develop smarter, healthier and more resilient data centers,” Sandhu added.