ASHRAE, the National Electrical Manufacturers Association (NEMA) and Pacific Northwest National Laboratory (PNNL) have developed the AI Data Center Energy Performance Framework, which provides voluntary guidance for the design, commissioning, retrofit and operation of AI data centers.

The framework applies across the facility life cycle and covers hyperscale, edge and retrofit applications. Its objectives extend beyond reducing facility energy use to include water efficiency, grid reliability and resiliency, and compute performance. It does not replace applicable codes, ASHRAE standards (such as thermal guidelines for data centers), manufacturer requirements or owner requirements.

The framework's energy and thermal efficiency recommendations are particularly relevant to data center operators managing the transition from conventional air-cooled IT to higher-density AI infrastructure. They include notes on segmenting high-density AI computing from general-purpose IT loads; scalable direct liquid cooling (DLC) infrastructure; air management; elevated operating temperatures; economization; intelligent cooling controls; and recommissioning.

Individually, many of these measures are established engineering practices. However, the broad-based adoption of AI computing brings added complexity through the combination of much higher rack densities, hybrid air and liquid cooling, variable workloads and potentially different temperature requirements within the same facility. As a result, thermal management is becoming less about maintaining a uniform environment and more about managing several thermal environments concurrently. This Uptime Intelligence report summarizes the thermal management aspects of the framework.

Segment the thermal environment

The framework recommends physically and logically segmenting AI and non-AI loads so that different cooling methods and resiliency approaches (including for power delivery) can be applied.

This is an important distinction for mixed-use facilities. A high-density GPU cluster (typically with an average rack density of 80 kW and above) using direct-to-chip cooling may have very different thermal requirements from adjacent CPU computing servers, storage systems and networking hardware. Applying shared temperature setpoints and flow rates across both environments can lead to inefficiencies or, worse still, overcool the low-density environment (typically under 20 kW by modern standards) without adequately serving the high-density zone.

Segmentation allows the cooling architecture to follow rack density and technology. High-density zones can use cold-plate (also known as direct-to-chip) cooling, rear-door heat exchangers (RDHx) or immersion cooling, while lower-density zones can remain solely air-cooled. Each zone can then use appropriate facility water and approach temperatures, flow rates and control sequences.

The framework also links segmentation with scalability. Rather than building the full cooling system for an uncertain future load mix, the guide recommends that operators make provisions for an expandable, modular DLC infrastructure, including coolant distribution units (CDUs), manifolds, secondary fluid networks and phased heat-rejection capacity.

This is particularly relevant because the efficiency of cooling equipment varies with load. Oversized pumps and heat-rejection equipment may spend years operating far below their design points, while infrastructure without sufficient expansion capability can constrain future rack deployments.

Quantify the heat remaining in air

The framework also notes that a primarily liquid-cooled environment may still need air cooling. Most direct-to-chip systems transfer heat from selected components, such as processors, to a liquid loop, while memory banks, storage drives, networking components and power electronics often continue to reject heat to air. This may amount to 20-30% of a rack's total heat load. The framework therefore treats high-density infrastructure as a hybrid thermal environment that requires concurrent management of the DLC and residual air loads.

The engineering implication is significant. Data center operators need to understand how much rack heat is actually transferred to liquid as opposed to air. This can be problematic because the exact ratio depends on IT hardware design and may be affected by coolant temperature setpoints — specifications that are often unknown at the design stage of the mechanical systems.

For example, an 80% liquid cooling heat capture ratio on a rack with a 120 kW thermal load leaves 24 kW to be removed by air. Even a high direct liquid cooling (DLC) capture ratio can therefore leave a substantial air-cooling load. Room cooling capacity should consequently be based on the expected residual air-side heat load of the installed IT equipment, not simply on the number of racks designated as liquid-cooled.

Changes in server architecture or liquid cooling coverage can also change this requirement over time. Newer IT hardware designs may have higher capture ratios through increased cold-plate coverage of components, pushing them above 90% or even approaching 100%. However, not all racks in the room may be fully liquid-cooled, as development and production servers, storage systems and network racks may share the same environment. The general trend toward higher liquid coolant temperatures (e.g. 45-55°C/113-131°F on the return side) will also means that some additional heat will radiate back into the room, which needs to be taken into account.

If conservative DLC supply temperatures are used (e.g., 25°C/77°F), RDHx can reduce the room-level airflow burden by transferring most of the rack air exhaust heat into a liquid loop before it mixes with the surrounding air. Where RDHx is not used, CRAHs, CRACs, in-row units or other room systems still need sufficient capacity and airflow to manage the residual load and any fully air-cooled equipment.

Higher temperatures, tighter air management

Before raising temperatures, the framework emphasizes conventional air-management fundamentals: hot-aisle/cold-aisle arrangements, containment, minimizing bypass and recirculation, and matching airflow to IT demand.

The important measurement is the condition of air entering the IT equipment, not simply the average room temperature. Room averages can conceal recirculated exhaust, inadequate airflow and localized hot spots. Rack-inlet sensing therefore becomes increasingly important as operators move closer to the upper portion of an approved environmental envelope.

Once airflow distribution is controlled, supply-air temperatures can potentially be raised while maintaining acceptable rack-inlet conditions. The framework favors operation within ASHRAE TC9.9 recommended conditions and, where equipment specifications and risk criteria permit, consideration of allowable Class A1 conditions. X-factor data indicates that controlled, brief excursions to higher inlet temperatures can have a limited effect on annualized hardware failure rates as outdoor temperatures approach cooling-system design conditions. Where climate, cooling architecture and reliability requirements support this strategy, new builds or major refurbishments may benefit from smaller cooling equipment and associated power infrastructure, reducing capital costs without reducing the IT heat load served (see Looking for the x-factor in data center efficiency).

Recommended and allowable conditions should not be treated as interchangeable. Moving into an allowable range by design can increase economizer opportunity, but it also reduces the thermal margin available during equipment failures, rapid load changes or adverse ambient conditions. Data center operators need to consider their facilities' available thermal margin through continuous commissioning to verify that capacity requirements are met under normal and, particularly, abnormal conditions (see Assessing data center operating headroom for extreme weather).

Heat rejection at higher temperatures

The framework extends the temperature discussion beyond room air to the setpoints of the technology cooling system (TCS) used for high-density AI equipment.

A TCS is an integrated thermal system that can include IT-side and facility-side liquid loops, CDUs, pumps, valves, piping, heat exchangers, instrumentation, leak detection, controls and heat-rejection equipment. Different liquid technologies, including direct-to-chip, RDHx and immersion, place different requirements on this system.

Supply and return temperatures are particularly important because they determine which heat-rejection technologies can be used efficiently.

The framework identifies three principal economizer approaches:

• Airside economization uses ambient air to cool down the data center environment. Direct fresh-air systems use filtered outdoor air when temperature, humidity, particulate and corrosion conditions are acceptable. Indirect systems use a heat exchanger to avoid mixing internal and external air. Both approaches can incorporate evaporative or adiabatic assistance to extend economization hours.
• Waterside economization and dry or adiabatic/evaporative coolers reject heat without chiller compressor operation when ambient conditions permit.
• Refrigerant-based economization reduces compressor operation in direct-expansion cooling systems.

The main point about TCS and economization is that the range of viable temperature options is considerably wider than with air cooling. Modern IT hardware can accept warm coolant, with outlet temperatures of 45-55°C (113-131°F), without affecting IT performance. In many climates, this can reduce chiller compressor operation to a low number of hours a year, or even eliminate entirely where water consumption is acceptable.

The appropriate strategy depends on climate, water availability, air quality, redundancy and the required IT operating temperatures. This creates an important engineering trade-off. Evaporative heat rejection can reduce energy consumption but increase water use. Dry cooling can reduce process water consumption, but it is more constrained by outdoor dry-bulb temperatures and will require more compressor and fan power, which can only be mitigated by using higher facility water (and, in turn, coolant) temperatures.

From setpoint to operating envelope

The framework also recommends moving beyond rigid temperature control through wider or floating setpoints that track ambient conditions within limits.

A cooling plant designed to maintain the same supply-air or supply-water temperature regardless of IT load and weather may consume energy to preserve thermal margin that is not always required.

With floating control, temperature is reset within an approved operating envelope. As outdoor temperatures rise, allowing a controlled increase in data hall temperatures within the approved operating envelope can extend economizer operation and reduce compressor energy, particularly at partial load. Conversely, when ambient temperatures fall, supply setpoints can be gradually lowered. This helps offset the accelerated wear of IT hardware at elevated temperatures, neutralizing any negative impact on failure rates. In many climates, floating temperatures may even benefit hardware failure rates.

All of this requires coordinated control of more than just temperature. Variable-speed pumps can respond to actual flow demand. Differential-pressure reset can reduce pumping energy as valve positions change. Fan speed can respond to rack-inlet conditions. Economizer enablement can vary with ambient conditions, while CDUs regulate IT-side flow, pressure and temperature. In effect, the cooling system manages thermal margin rather than a single temperature setpoint.

This distinction becomes increasingly important for AI, because the thermal load itself is more dynamic.

Link cooling to AI workload power

Large training workloads can cause rapid changes in cluster power, but some of these changes are known to the IT system before the cooling system experiences them. The framework recommends intelligent controls that monitor, learn and predict AI workload power patterns to optimize the energy used for cooling, including pumping energy.

This opens up the possibility of moving from constant flow to fast-reactive and, ultimately, to predictive cooling control. If a regularly scheduled workload is expected to increase cluster power, cooling controls could pre-position pump flow, CDU capacity or heat-rejection equipment rather than waiting for temperatures or differential pressure to move outside their normal ranges.

The reverse applies when compute demand falls: pump, fan and cooling capacity can be reduced rather than maintained for peak conditions.

Weather forecasts can provide another input. Expected IT load, ambient conditions and redundancy status could be combined to determine temperature resets, economizer availability and equipment staging.

Predictive operation does not eliminate the need for conventional safeguards. Equipment limits, alarms, redundancy requirements and failure sequences remain necessary. The opportunity is to give the cooling system more information about when and where thermal demand will occur.

Continuous commissioning as IT evolves

More dynamic operation also makes recommissioning more important. AI infrastructure can change substantially after initial commissioning as rack densities increase, liquid loops expand and new server generations alter flow, temperature and residual-air requirements. Particularly with DLC upgrades, shortcuts in commissioning best practices at any level can introduce unknown operational risks that may not surface until months or years later.

The framework calls for continuous commissioning and ongoing optimization. Verification should include rack-inlet conditions, containment, liquid-loop temperature and flow, CDU performance, economizer sequences, sensor accuracy, control resets and response to representative IT-load changes.

This is particularly important for facilities using floating temperatures. Efficiency increasingly depends on controls and sensors accurately identifying how much thermal margin is available. A control strategy commissioned for one generation of AI hardware may no longer be appropriate after the IT configuration changes.