As AI and high-performance computing push rack densities beyond 100 kW, liquid cooling is becoming standard. Once operators choose direct-to-chip cooling, one decision shapes the entire facility design: the temperature of the water supplied to the IT equipment. The two most debated options today are 32°C and 45°C. Both avoid traditional chilled water at 12°C to 18°C, but they lead to very different energy, cost, and compatibility outcomes.
In this comparison, the figure is the facility water supply temperature entering the coolant distribution unit (CDU). With a typical 10K temperature rise across the IT, a 32°C supply returns at about 42°C, while a 45°C supply returns at about 55°C. Under ASHRAE TC 9.9 liquid-cooling classes, these correspond broadly to W32 and W45 facilities.
|
Factor |
32°C Water |
45°C Water |
|
ASHRAE class |
W32 |
W45 |
|
Typical return temperature |
42°C |
55°C |
|
Mechanical cooling |
Often required in hot climates |
Rarely required |
|
Free cooling / dry cooling |
Partial, climate-dependent |
Year-round in most climates |
|
IT compatibility |
Widest |
Limited to validated servers |
|
Heat reuse value |
Low without a heat pump |
High, direct reuse possible |
|
Typical facility PUE impact |
Lower than air cooling |
Lowestdit |
Energy Efficiency and PUE
The main argument for 45°C water is energy. Heat can be rejected to outside air whenever the ambient temperature is below the return water temperature. With a 55°C return, a dry cooler can reject heat even on a 40°C day, allowing chiller-free operation in most locations, including much of Southeast Asia, the Middle East, and the southern United States.
A 32°C system cannot do this consistently in hot climates. When ambient temperatures reach 35°C to 40°C, chillers or free-cooling chillers must assist, consuming compressor power. As a result, a well-designed 45°C facility can achieve a partial PUE 0.05 to 0.15 lower than a comparable 32°C facility, and it can also eliminate refrigerant use and reduce maintenance.
In temperate climates such as Northern Europe or the northern United States, the gap narrows. A 32°C system can use free cooling for most of the year, so the additional saving from 45°C is smaller.
IT Performance and Compatibility
The main argument for 32°C water is compatibility. Most current direct-to-chip servers, CDUs, manifolds, and rear-door heat exchangers are designed and warranted for W32 operation. Operators have a wide choice of vendors, and chips run cooler, which supports higher boost frequencies and lower leakage current.
At 45°C, coolant entering the cold plate is 13K warmer. CPU and GPU junction temperatures rise by a similar amount unless flow rates or cold-plate designs are improved. Some accelerators will throttle sooner, age faster, or require specific validation for W45 operation. Seals, hoses, and quick disconnects must also be rated for continuous high-temperature service. For this reason, 45°C is currently best suited to purpose-built AI clusters where the IT and facility are designed together, not to mixed enterprise environments.
Both temperatures sit above the typical data hall dew point, so neither system should cause condensation on the liquid loop under normal operation.
Heat Reuse and Sustainability
Return water at 42°C is usually too cool for direct district heating, industrial processes, or absorption cooling. It normally needs a heat pump to become useful, which adds cost and energy.
Return water at 55°C is different. It can directly supply district heating networks, offices, greenhouses, and some industrial processes. For operators with a nearby heat user, or with ESG reporting requirements, this can turn waste heat from a cost into revenue and significantly improve Energy Reuse Factor (ERF) and total carbon emissions.
A 45°C system paired with dry coolers can also operate with near-zero water consumption, improving WUE compared with evaporative systems.
Which Should You Choose?
Choose 32°C if you need maximum server choice, are retrofitting an existing site, operate a mixed IT fleet, or cannot accept any performance risk on current-generation GPUs. It is the safer, proven default for most colocation and enterprise projects today.
Choose 45°C if you are building a new, single-purpose AI facility, your server vendor validates W45 operation in writing, you are in a hot climate, or you have a valuable use for waste heat. The energy and heat-reuse benefits are real, but only if the IT is designed for them.
In practice, many operators are taking a hybrid path: selecting 32°C for day-one operation, but sizing dry coolers, piping, and CDUs so the facility can migrate to 40°C to 45°C as server technology matures. This protects compatibility today without locking the data center into chillers for the next 15 years.
Containerised AI Solutions by TLS Energy
TLS Energy provides containerised AI data centre solutions that combine the enclosure and cooling system in a prefabricated, transportable package. Its offering includes the data centre enclosure, liquid-cooling and air-cooling modules, coolant distribution units (CDUs), manifolds, and interconnecting pipework, together with integrated power distribution, controls, leak detection, and fire protection. Systems can be engineered for either 32°C or 45°C cooling water, depending on the IT equipment, site climate, and energy targets of each project. By factory-assembling and testing the modules before delivery, TLS Energy helps operators deploy high-density AI capacity faster, with less site work and a scalable path from a single container to multi-megawatt installations.