Optimizing Cooling Power for AI Workloads: A Reference Design Approach

You will achieve free cooling for 100% of the load only under conditions determined by the specific system design and site operating environment. The portion of load that free cooling can support, the heat exchanger capacity relative to the chiller, the outdoor temperature required to satisfy the full load, and the annualized percentage of free cooling are all project-specific. These depend on factors such as climate, load profile, facility water temperatures, and approach temperatures. In other words, these conditions must be evaluated using the actual design criteria and weather data for the project rather than assumed as fixed values.

Optimization of a thermal management system is focused on both peak energy savings and annualized energy savings. Through sequencing, plant coordination, and decision logic, optimization is designed to improve both power and energy performance. This also reflects the importance of power constraints, since cooling efficiency can directly affect the amount of IT capacity available.

The advantages of having cluster management communicate with the chiller plant include improved energy efficiency, resiliency, flexibility, and a better plant response to changing compute demand. In practice, this communication helps coordinate thermal plant behavior with IT load changes, allowing the overall system to respond more effectively to operational needs.

Yes, there is an estimate of the energy savings achievable with AI-assisted thermal management in practice. In many cases, estimates range from a 10–30% reduction in cooling energy, depending on the site, controls maturity, and thermal load. Trane helps customers achieve these savings through a robust site-level controls infrastructure and by applying the right optimization strategy based on the specific needs of the site.

Yes, we do have a solution for directly optimizing cooling system operation to “protect” token generation capacity while maximizing token generation output. This is done by ensuring there is sufficient cooling to maintain reliable IT operation, assuming the thermal system has been designed to meet that demand. We then optimize the operation of the cooling equipment to minimize power usage, which helps free up capacity to support greater token generation output.

A cloud connection is not necessarily required, and the system can operate without sending factory data to the cloud. A thermal management system can be designed to run locally at the site using on-premises controls and analytics. The best approach depends on the system architecture, cybersecurity requirements, latency needs, and the owner’s data governance preferences. In many cases, local or edge-based operation can still deliver real-time monitoring, predictive control, and system optimization, while cloud connectivity is used only if the customer wants additional remote analytics, fleet-level insights, or centralized management.

For a 1 GW data center, the amount of thermal energy storage (TES) and land area needed depends on the required storage duration, the operating temperature range, and the storage technology selected. At a high level, the TES capacity and land requirement would be extremely large if sized to support the full 1 GW load, with storage volume and footprint increasing roughly in proportion to the number of hours of coverage required. In most cases, however, TES is not designed to carry the full 1 GW load for extended periods. Instead, it is typically sized to manage short transients, reduce peak cooling demand, or provide added operating flexibility.

Reference designs can help ensure code compliance by aligning with the exact requirements of the local jurisdiction and the specific compliance path being used, such as IECC or ASHRAE 90.4. Data centers are generally not exempt from code requirements, but they may be permitted to follow data-center-specific provisions rather than standard commercial building rules. As a result, ensuring code compliance requires evaluating the reference design against the applicable local codes and the chosen compliance framework.

As cooling systems move closer to mission-critical IT loads, liquid leak detection is best handled as a layered strategy across mechanical design, equipment packages, and controls/BMS integration. Mechanical design provides containment and sensor placement, equipment packages provide local detection and response, and the controls system ties everything together through alarms, trending, and coordinated action. For mission-critical IT environments, the best practice is to use all three together rather than relying on only one.