Your Questions Answered: Optimizing Cooling Power for AI Workloads: A Reference Design Approach

A July 14, 2026, webcast discussed why customers and engineers should avoid optimizing around a single piece of equipment and instead take a holistic, system level approach. Here are the additional audience questions from the event.

As today’s rack densities and site power demands increase, organizations need cooling strategies that are efficient, scalable and fast to deploy. In a July 14, 2026, webinar, Optimizing Cooling Power for AI Workloads: A Reference Design Approach, presenters from Trane explored the concept of a thermal management reference design for an artificial intelligence (AI) factory — what it is, why it matters and how it helps teams optimize data center thermal performance within site constraints.

In this session, attendees learned how to make informed choices that reduce peak cooling power through a holistic thermal management approach, unlocking available power for generating compute or revenue. They also discovered design options that reduce annual cooling power usage, minimize redundant infrastructure through intelligent design and incorporate best practices that enhance operational stability.

Additional questions were answered by Andrew Jenkins, systems product manager, Trane, and Krista Hubbs, product manager, data center controls, Trane.

Does the Trane controls system support the use of BACnet/SC?      

Answer: Yes. Trane controls systems support BACnet/SC (BACnet Secure Connect).

In what situations would you recommend using an air-cooled chiller plant with dry coolers for water side economizer versus the reference design of a water-cooled chiller with a bypass free cooling heat exchanger paired with dry coolers?       

Answer: The right approach depends largely on the size and scaling strategy of the cooling block. Water-cooled chillers generally provide higher capacity and better efficiency, making them a strong fit for larger cooling blocks such as the 102.4 MW reference design. For smaller cooling blocks, where the size increments of water-cooled chillers may be less practical, air-cooled chillers with integrated free cooling can be a better fit. Air-cooled plants can also offer advantages for phased deployment, particularly when capacity needs to scale as data halls are leased or commissioned over time.

Are there any adiabatic architectures for transcritical applications?

Answer: Yes. Adiabatic cooling can be an effective option for transcritical applications when water is available and when its use meaningfully reduces peak design loads.

Does Trane manufacture energy batteries?  

Answer: Trane offers Thermal Energy Storage (TES) solutions, which are sometimes described as thermal batteries because they store cooling capacity for later use.

Considering the higher temps for the thermal storage, will we be seeing PCMs soon?      

Answer: Trane is actively evaluating a range of thermal storage technologies, including higher-temperature phase change materials (PCMs) as potential solutions to reduce peak electrical demand and overall energy consumption.

Is there a typical tech data sheet for electrical?        

Answer: Yes. Trane equipment is provided with technical data that includes electrical requirements. Future reference designs may also include a dedicated power segment to complement the cooling design.

How are white space humidity levels maintained if the lowest CHW temp for the air-cooled plant is 71 F?          

Answer: In this approach, the fan coil wall supply temperature is approximately 80.6 F, which is appropriate with a 71.6 F chilled water supply temperature. The technical loop operates at a higher temperature — approximately 86 F, or about 3 C above the potential air dew point — to support the intended humidity control strategy.

With 27 C (80.6 F) supply air with up to 40 C entering, you’re running the FCW hot, which is great for chiller efficiency, but it puts you firmly in ASHRAE A2/A3 territory for the air-cooled information technology (IT). Are the switches and storage vendors on board with those inlet temperatures, or is that still a negotiation in every project?

Answer: A successful design requires a holistic approach to ensure that all supporting infrastructure can operate reliably within the planned air conditions. These design temperatures are based on NVIDIA’s DSX guidance, which considers the broader AI factory ecosystem, including compute, power, networking and storage. Final acceptance, however, should always be confirmed with the applicable equipment vendors for each project.

What are the advantages of using the magnetic bearing type chillers?       

Answer: Magnetic bearing chillers do not require oil for compressor lubrication, which can simplify maintenance and reduce complexity associated with oil management systems. In some applications, they can also offer strong efficiency performance and support higher leaving water temperatures. That said, chiller selection should always consider the full system design, as multiple compressor technologies may be suitable depending on project requirements.

What TES tank size and ride-through duration, in minutes, does the Trane system typically recommend for a data center chilled-water system?           

Answer: TES tank sizing and ride-through duration depend on factors such as primary loop flow rate and the required rapid restart time. For example, if the system requires three minutes of ride-through and the primary flow rate is 500 GPM, a 1,500-gallon tank would be required to support the system during a failover event. Final sizing should be based on the specific system architecture and operating requirements.

Whose dry coolers are you using?    

Answer: Trane works with multiple dry cooler partners depending on project requirements, application needs and regional considerations. For details on a specific project, please contact your local Trane account manager.

What is the recommended technology cooling loop fluid with your GigaModular CDU?    

Answer: The GigaModular CDU is compatible with several technical loop fluid chemistries used in the market today. A common example is a 25% propylene glycol solution, though final fluid selection should be based on the specific application and site requirements.

Do the dual loops ever couple considering that the fan coil loops return water temps are close to the supply water temps of the direct to chip? Or do they always stay separate?

Answer: In the current reference designs, the loops are kept separate because the load profiles for direct-to-chip liquid cooling and air-cooling equipment are not uniform. In some cases, there may be opportunities to evaluate coupling medium-temperature return water with high-temperature supply water to improve overall delta T, but that should be assessed on a project-specific basis.

With Trane insights technology/AI, is that able to be an on-prem solution, or does it require outbound communication? If outbound, does it have to be via BMS system, or can you have direct asset outbound without compromising BMS connection?  

Answer: Deployment architecture depends on the specific solution, customer requirements and cybersecurity considerations. In general, Trane can work with customers to evaluate on-premises and connected architectures, including approaches that separate asset-level communications from the building management system where appropriate. The recommended structure should be defined in coordination with the customer’s IT, operational technology (OT) and cybersecurity teams.

For freeze protection, are you implying that the scale makes glycol systems unreasonable?

Answer: No. Glycol remains a common and practical freeze protection strategy, especially in hydronic systems, and it is recommended within the reference designs where freeze protection is required.

Why not use geothermal or cooling ponds? 

Answer: Geothermal systems and cooling ponds can be viable solutions in the right circumstances. However, for very large AI factory applications, the scale of heat rejection required often makes these approaches less common than other heat rejection strategies used in the industry today.

With thermal storage, do you see a higher promotion to ice storage rather than liquid?   

Answer: At the higher loop temperatures being considered in these designs, the market is currently favoring liquid thermal storage rather than ice storage.

What percentage of water is lost in the chilled water loop per hour? What is the overall rate of consumption of water per MW?          

Answer: In these reference designs, no water is evaporated as part of the cooling process. Water is used in a closed-loop system to transfer heat through the thermal chain, so it is not consumed during normal heat transfer operations in the way it would be in an evaporative cooling system.

How does the increased lift offset the higher consumption for the reduced dry coolers and higher delta on that loop?  

Answer: The tradeoff is evaluated at the total system level. A chiller may operate at a coefficient of performance (COP) of approximately six, meaning it delivers about six units of cooling for every unit of electrical input. Dry coolers and pumps typically operate with lower overall efficiency. As a result, shifting more of the cooling work to a high-efficiency chiller and reducing the burden on lower-efficiency components can reduce overall system power consumption, even when chiller lift increases.

What is the lowest approach you ever select on CDUs? Do you go down to 4 F?    

Answer: Yes, some designs have used CDU approach temperatures as low as 2 C (4 F). This can benefit the cooling plant by reducing lift and increasing free-cooling hours, but it also increases flow requirements and associated pump power. As with any design decision, the tradeoffs must be evaluated at the system level.

What is the ROM for a 1 GW system?

Answer: A rough-order-of-magnitude estimate for a 1 GW system depends on many variables, including site conditions, redundancy strategy, cooling architecture, construction approach and project scope. Because of that, a meaningful ROM should be developed based on the specific project assumptions. Please contact your local Trane account manager to discuss the application in more detail.

B&G is suggesting this waste heat could be used for district heating. Any thoughts on that?

Answer: Yes. Heat recovery can be an effective way to improve overall site energy utilization and provide a useful source of low-grade hot water for a district heating application or other adjacent loads. Trane reference design 503 includes an example of this type of strategy.

What is a typical data center room and humidity temperature?      

Answer: Acceptable temperature and humidity ranges depend on the IT equipment and the operating envelope defined by the manufacturer. In practice, these conditions are typically established in coordination with the IT vendors and with reference to ASHRAE guidance for data centers. Your local Trane account manager can help align the cooling design with those requirements.

What is the lead time for a complete 5 MV system?

Answer: Lead time for a complete 5 MW system depends on the selected equipment, project scope, configuration and supply chain conditions. Your local Trane account manager can provide guidance based on the specific equipment selections and project timeline.

How critical is the AI optimization on the impacts of your controls? How do you see that changing and improving in the near future?

Answer: AI optimization is an extension of our controls capability across both hardware and software. Today, we already deliver strong performance through proven engineered control algorithms embedded in our unit and system controls, which can produce meaningful energy and operational savings. AI can build on that foundation by enabling more predictive and adaptive control strategies, creating additional opportunities for efficiency and operational improvement. We also recognize that some operators prefer to adopt AI-based optimization gradually, and our approach can scale to match each facility’s comfort level and operational priorities.

Do you see predictive fault detection features built into assets controls?   

Answer: Yes. We are using the control and sensor infrastructure within equipment to generate data that can be analyzed for early identification of operational issues. At the scale of these mechanical systems, predictive fault detection and diagnostics can play an important role in maintaining uptime, improving reliability and reducing unplanned service events.

What are some of the specific things the AI does that make everything work better?        

Answer: AI is particularly valuable in predictive applications. One example is identifying and mitigating microclimates or heat island effects across a chiller array. By combining operating data with forecasted weather conditions, AI can help predict which units may be more susceptible to elevated thermal stress. That insight can then inform staging and operating strategies that protect equipment, improve resilience and support continuous operation.

Are your air-cooled chillers magnetic lev type?        

Answer: Trane offers an air-cooled magnetic bearing chiller solution. This technology was highlighted in reference design 504 during the webinar.

For the cooling water piping serving the CDUs, would you recommend routing the piping overhead (suspended from the ceiling) or below the raised floor?   

Answer: For liquid-cooled applications, technology cooling system piping is typically routed overhead above the server racks to improve accessibility for service and integration. Final routing, however, should be based on the data hall layout, operational preferences and risk management strategy.

In the block diagram with six chillers, are they in 4% 2B2 (N% 2B2) configuration?

Answer: The six-chiller diagram was intended as a high-level representation of the overall plant concept. In the reference design, each cooling block includes 10 high-temperature chillers and one medium-temperature chiller. Final redundancy configuration should be evaluated in the context of the full system design.

Can you comment on harmonics?     

Answer: Yes. Many thermal management systems can be equipped with harmonic mitigation solutions, depending on the application and electrical design requirements.

Do Trane air-cooled chillers have built-in economizers to “pre-cool” the return water before sending it to the compressor loop?          

Answer: Trane offers air-cooled chiller solutions with direct and indirect free-cooling capabilities. These configurations can reduce the mechanical cooling load and improve overall system efficiency when ambient conditions are favorable.

Are absorption chillers more efficient than electric chillers?

Answer: In most cases, no. Absorption chillers generally operate at a lower COP than electric chillers — often below one, compared with approximately six for a high-efficiency electric chiller in the right application. However, absorption chillers may still be attractive where a low-cost or otherwise available heat source exists, such as waste heat or on-site generation.

Could you tell us which model has the integrated pre-cooling economizer? What are the capacities? Are these options available now, or will they be commercialized soon?      

Answer: These solutions are available. Because model selection and capacity vary by application, configuration and region, the best next step is to work with your local Trane account manager to review the most appropriate options for your project.

Can dry coolers be installed next to each other, or is a minimum clearance required for air circulation. And how much space is needed?         

Answer: Dry coolers do require minimum clearances to limit heat recirculation and to allow proper service access. Exact spacing depends on the equipment design, installation layout and site conditions. Manufacturer recommendations should be followed, and CFD analysis may be appropriate for more sensitive or high-density applications.

Are equipment rack power density threshold levels suitable for air or liquid cooling?       

Answer: The threshold depends on the server design, airflow strategy and overall thermal architecture. Hybrid approaches are also available, including rack-level solutions that combine liquid cooling at the server with air-side heat rejection at the rack. In many cases, once rack densities approach 100 kW, liquid cooling becomes the more practical solution.