Thermal Triage: Not Every Watt Needs Intensive Care

An AI data center produces different grades of heat. Serving them all from one facility water loop at one setpoint wastes compressor energy and provisioned power the same way admitting every ER patient to the ICU wastes critical care.

Key Takeaways

• AI data halls generate distinct grades of heat: high-temperature chip heat, residual air-side heat and precision loads in electrical and support spaces.
• When one facility water loop serves every load at one setpoint, the coldest requirement in the building sets the water temperature for all of them.
• That forces heat that could be rejected with dry coolers or outdoor air through compressors, and the power provisioned for that compressor capacity can’t be allocated to compute.
• Matching cooling to each heat grade, with setpoints tested against local climate and water conditions, can shift more of a facility’s provisioned power to compute. Power Compute Effectiveness (PCE) measures that allocation at the design stage.

If you walked into a busy emergency department (providing you could walk) your first point of contact would be a reception desk. Questions on the type of injury or illness would follow. A triage team is there to sort arrivals by severity, directing each person to the specific level of care their condition requires. That crucial sorting process is what keeps the most intensive, costly medical resources free for the severe cases that really need them.

A hospital that admitted every burn, scrape and sprain to the intensive care unit (ICU) would jeopardize critical patients, not to mention blow its financial budget.

High-density AI data centers face a similar challenge when it comes to allocating the right thermal capacity. Some facility designs fail to put a technical triage system in place to right-size cooling according to actual need; instead, they serve every load from a single facility water loop at a single setpoint. The potential result is the coldest requirement in the building can dictate the water temperature for the chip loop, the air-side loop and the electrical rooms alike.

Three Grades of Heat, Three Dedicated Loops

Any facility running direct-to-chip cooling already separates the technology cooling loop from facility water at the coolant distribution unit (CDU). That separation is a given. The real question is what happens on the facility side: whether that water is produced at one temperature for every load, or at temperatures matched to what each load needs.

Many AI data halls generate several distinct thermal loads that differ significantly in temperature, volume and required precision. For example:

• High-Temperature Chip Heat (The Critical Case): High-density AI accelerators such as GPUs run extremely hot. However, the coolant directly touching them can be significantly warmer than the ambient room air while keeping the silicon within safe operating limits. ASHRAE’s thermal guidelines allow facility water supply temperatures up to 27°C (class W27), 32°C (class W32) and beyond for liquid-cooled equipment. This heat rewards direct capture right at the source, through direct-to-chip liquid cooling served by a CDU.
• Residual Air-Side Heat (The Routine Case): Ambient thermal loads from servers and general building spaces represent a large volume of heat, but at relatively undemanding temperatures. Air-side cooling and rear-door heat exchangers are usually sufficient in temperate climates; hot or humid sites may need more.
• Electrical & Support Spaces (The Precision Case): Uninterruptible Power Supply (UPS) areas, battery rooms and electrical switchgear require tight, stable temperature control to remain within strict equipment tolerances. Here, thermal stability is the priority rather than extreme refrigeration, which makes DX CRAC units a natural fit.

Matching cooling to each heat grade is best done via three dedicated cooling loops: a high-temperature loop, a medium-temperature loop and a low-temperature zone. Each loop is custom sized to the specific grade of heat it serves, rather than forcing the entire facility to match the single coldest requirement in the building.

Building an Intelligent Cooling Toolkit

Rather than selecting a single technology, this approach assigns each loop the rejection method best matched to its heat grade.

• The high-temperature loop pairs dry coolers with a CDU serving direct-to-chip liquid cooling. Because the coolant returns warm, heat can be rejected to outdoor air with little or no compressor involvement, but only when ambient conditions sit comfortably below the loop’s supply temperature. Hotter or more humid sites need trim chilling or adiabatic assist, which brings back compressor energy and, in the adiabatic case, water use.
• The medium-temperature loop uses free-cooling chillers for air-side loads and rear-door heat exchangers. How many hours they can run without compressors depends on the local climate.
• The low-temperature zone uses self-contained DX CRAC units, which sit outside the facility water loops, for power rooms, UPS areas and batteries, where stability matters more than volume.

Compare that spread with a conventional plant, where a single facility water pipe may end up feeding CDU, CRAH and CRAC alike at one temperature. In suitable climates, the three-loop design can remove the largest share of heat from compressor duty altogether.

Site Conditions Decide the Split

The three-loop logic holds everywhere, but the loop temperatures and rejection methods may vary. Each setpoint must be set against the site’s design ambient conditions and its local weather data. Dry coolers can carry the high-temperature loop without compressors only for the hours when outdoor conditions sit below the supply temperature minus the heat exchanger’s approach. In cooler climates that covers a large share of the year. In hot or humid climates, it covers far fewer hours, and the design needs trim chillers or adiabatic assist for the rest.

Water is the other variable. Dry coolers avoid evaporative water use but consume more energy as outdoor temperatures rise. Adiabatic assist recovers some of that energy at the cost of water. In regions where water is scarce or expensive, that trade-off can decide the design, and it should be tested alongside the energy case rather than assumed.

The same logic applies to the medium-temperature loop, whose free-cooling hours depend on local climate, and to the residual air-side load, which depends on design ambient. A site-specific analysis of hours above each supply setpoint, run against local weather data, is what turns the three-loop architecture from a principle into a design.

The Hidden Cost of a Single Setpoint

When a facility produces all its cooling water at one uniform temperature, the most sensitive, cold-dependent requirement in the building dictates the operating point for every megawatt in the facility. Heat that could easily be rejected at warmer temperatures via outdoor air or dry coolers is instead forced through energy-intensive compressors.

The single-setpoint design also means pumps and distribution equipment draw power continuously regardless of load grade, simply to move water that in many cases never needed chilling. Because compressors are among the most power-hungry components in data center infrastructure, every kilowatt of compressor capacity provisioned to lift heat that never needed lifting is a kilowatt of provisioned power that can’t be allocated to compute.

That structural trade-off can now be explicitly quantified by a new design metric: Power Compute Effectiveness (PCE).

Understanding Power Compute Effectiveness (PCE)

While PUE measures how efficiently an operating facility uses energy, PCE is a design-stage metric that evaluates what share of a facility’s total provisioned power is provisioned for compute, versus support infrastructure (cooling, electrical losses and auxiliary loads).

Expressed as a value between 0 and 1, PCE is calculated as provisioned compute power divided by total provisioned power.

Example: A facility with 100 MW of total provisioned power and 70 MW provisioned for compute has a PCE of 0.70.

Recognizing its importance for next-generation facilities, the Open Compute Project (OCP) formally accepted a foundational white paper defining the PCE metric in October 2026.

The paper names cooling architecture comparison as one of the metric’s core applications, and positions PCE to “evaluate infrastructure capacity allocation at the design level” as compute density rises and utility interconnection constraints tighten. Its worked hyperscale example makes the stakes concrete: of 100 MW provisioned, cooling and heat rejection claim 16 MW before electrical losses and auxiliary loads take a further 5 MW, leaving 79 MW for compute and a PCE of 0.790.

Where Design Optimization Begins

Ultimately, an emergency department successfully treats patients and saves lives because of the quality of its doctors, nurses and equipment but also because those resources are allocated deliberately where they are needed most.

An AI hall’s thermal loads need the same discipline. Chip heat, residual air-side heat and support spaces are three different cases. Serving them all from one loop at one temperature strands power that ultimately could have been directed to compute.

Frequently Asked Questions

What is Power Compute Effectiveness (PCE)?

PCE is a design-stage metric that shows how a facility’s provisioned power is allocated. It equals provisioned power for compute divided by total provisioned data center power. PUE measures how efficiently an operating facility uses energy, while PCE measures how much of the provisioned power is allocated to compute.

Why does a single cooling setpoint strand power in AI data centers?

When one facility water loop serves every load, the coldest requirement sets the water temperature for all of them. Heat that could be rejected with dry coolers or outdoor air goes through compressors instead, and the power provisioned for that compressor capacity can’t be allocated to compute.

Can direct-to-chip liquid cooling run without compressors?

For part of the year, at many sites. Dry coolers can carry a high-temperature loop without compressors when outdoor conditions sit below the supply temperature minus the heat exchanger’s approach. Hot or humid sites need trim chilling or adiabatic assist for the rest of the year.

Airsys designs and builds cooling for every grade of heat in the data center, from direct-to-chip liquid cooling to precision cooling for power and battery rooms. To explore how your site’s climate and thermal loads would shape a right-sized cooling design, and what that means for your facility’s PCE, connect with the team here.