Can Chiller-Less AI Data Centers be achieved in the Nordics?

The industry’s response to ‘warm-water’ GPU cooling envelopes (S45, up to ~45 °C inlet temperatures) has marked a significant shift in data centre thermal design philosophy, particularly for AI data centers. Here's our technical assessment of the AI data centre landscape and our perspective on whether chiller-less data centers are achievable. 

Our opinion is that until end-user requirements, GPU OEM operating temperatures, and site-specific ambient conditions align to enable full-year free cooling, mechanical refrigeration will continue to play a material role in high-density AI infrastructure. SO, the chiller is here to stay in one form or another.

Polar’s portfolio and geographic positioning are well-suited to leverage increasing opportunities for high ambient operation, reduced chiller operation and maximising sustainability, whilst still achieving rapid deployment.

Call for Cooling

Data center (DC) critical cooling, whether to serve Cloud or AI data centres, has always been a major contributor to PUE. In this article, we consider the reality of reducing this and ultimately repurposing of utility power for more compute.

Unless you have access to a year-round fjord, can seawater discharge, or can utilise extreme latitude direct air-cooling systems, a refrigeration cycle, or a chiller will feature in your critical cooling deployment.

Advancing GPU technologies are continuing to push the envelope in all aspects including kW/rack, Tokens per Watt and now maximum inlet temperatures to the GPU.

Whether the GPU accepts 45°C or less, it still needs that heat removing and rejecting. The ambient environment within which the data centre and critical cooling is located defines the optimum means to reject this heat; the location defines the extent at which we can reduce chiller plant.

Maximising the Free Cooling Opportunity

The data centre industry is incredibly efficient at optimising our energy use. The industry has been lowering PUEs for decades to realise more IT compute, with free cooling being one of the most effective means to do this; maximising the window of compressor free cooling operation.

Maximising the free cooling hours per annum is achieved through maximising the leaving water temperature (LWT) of the chiller serving the critical IT, coupled with operating in an environment that permits direct heat exchange i.e. run the cooling loop warmer, locate the data center somewhere cold.

In this example, evaporative water usage is not considered for heat rejection for well-documented reasons, covered elsewhere. Air-cooled chillers (ACC) and water-cooled chiller / dry air cooler (WCC/DAC) arrangements are the most viable non-evaporative heat rejection solutions. The achievable approach temperatures depend on specific equipment selection and operating conditions but are typically in the order of 10–20°K for ACC systems and 5–12°K for WCC/DAC configurations.

Once heat re-entrainment, prevailing wind effects, ASHRAE extreme weather conditions, and transient scenarios are considered, a further conservative local ambient temperature uplift will need to be accounted for at the heat rejection plant. This additional thermal penalty further reduces the available window for compressor-free (chiller-less) operation.

Raising LWTs to support higher GPU inlet temperatures can assist with chiller-less operation, but the real opportunity lies with the data centre location. At Polar, our pedigree in the Nordics positions us well to realise the potential of high latitudes and cooler climates.

While power, land and AI demand remain key priorities, efficiency is now a competitive battleground. Polar’s AI-ready facilities target PUEs of 1.12 - 1.15 through advanced liquid cooling, renewable power and purpose-built infrastructure, turning operational efficiency into a strategic advantage.

GPU Entering Water Temperature (EWT)

The above assumes the GPU inlet temperature is 45°C. This is not universal to all GPU technologies, nor is it accepted by all end users.

End users and operators need flexibility in future technology changes as well as ‘headroom’ when external ambient conditions are extreme, transient conditions occur, or the resilient plant fails. If the GPU inlet temperature is reduced to meet the end user requirements, then the opportunity for total free cooling (or chiller-less operation) also reduces.

Non-GPU Cooling: Enter the ‘Dual Loop’

Non-GPU elements need consideration, including the GPU air cooling component, the network racks, which will progressively become liquid cooled, UPS battery rooms, as well as other critical spaces such as MMRs.

All these non-GPU cooling terminals require much lower chilled water or facility water, further negating the prospect of a chiller-less operation.

Although these non-GPU systems may represent a relatively modest proportion of total cooling demand, they often dictate the minimum mechanical cooling requirement for the wider facility. Therefore, lending themselves to a dedicated standalone cooling system operating independently of the primary warm-water AI cooling loop. These systems are to be located on a separate lower-temperature loop.

Final thoughts

The elevated GPU inlet temperatures currently being deployed in the most advanced GPU technology need significant end-user adoption before a chiller-less data centre can be achieved. Until then the constrained mechanical power cannot be reutilised through an improved PUE and more token production, nor can chiller critical infrastructure be totally removed.

Even where elevated GPU inlet temperatures are technically permissible, external ambient conditions remain the governing constraint on achievable economiser operation. Consequently, the practical opportunity lies not necessarily in complete refrigeration removal, but in materially reducing compressor operation through hybrid heat rejection strategies.

In this context, water-cooled chillers coupled with dry air coolers (WCC/DAC) present a particularly effective arrangement. The DAC maximises economiser operating hours through lower approach temperatures and direct ambient heat rejection, while the water-cooled chiller (WCC) provides trim cooling during extreme operating conditions.

Achieving an entirely compressor-free operation across all annual conditions would require GPU and facility water temperatures to operate substantially above peak ambient conditions. At these higher temperatures, the reliability and performance of GPUs and other components suffer, the risk of thermal throttling increases, and component ageing accelerates.

Finally, to address the lower temperature constraint of non-GPU cooling demands required elsewhere in the data centre facility, a ‘dual cooling loop’ solution is the AI data centre direction of travel.

Polar works closely with customers to design each facility to meet their specific operational requirements, including cooling. This approach supports current AI deployments through tailored cooling and infrastructure solutions, while preserving the flexibility needed for future generations of high-density AI technology.

With our extensive Nordic experience, access to favourable high-latitude environments, and technical expertise in resilient critical cooling infrastructure, Polar are uniquely positioned to support the next generation of high-efficiency AI data center deployments.