Everybody wants to increase their data center power output.

The compulsion toward densification, driven by AI adoption, GPU clusters, cloud scale-out, and HPC, has reached new heights. Operators are seeking to extract more computing capacity – and therefore revenue – from the same physical footprint, but this comes with a corresponding rise in thermal intensity.

As more heat is generated within the same space, both at rack and data hall levels, the challenge of heat rejection is pushing traditional air-based cooling and chiller systems toward their practical and thermodynamic limits.

To support rising rack power densities, operators must evolve their cooling strategies, with a new class of high-density air-cooled infrastructure emerging to address growing thermal constraints. After all, as Patrick Cotton, product management director at Airedale by Modine, puts it:

“Cooling is a critical component of any data center, and there is significant pressure to deliver high-density solutions at scale – you simply can’t operate without it.”

Strain on traditional cooling architectures

As cooling demand for heat rejection increases, operators face an undesirable trade-off, whereby removing more heat from the data hall inevitably means rejecting more heat into the external environment, creating parallel thermal management challenges on both sides of the system.

The scaling effect extends across the entire thermal chain – greater cooling demand drives higher water flow rates, requiring larger pipework and more robust supporting infrastructure. This can compound space and power constraints, creating a cycle in which the infrastructure required to support higher-density IT loads increasingly competes with the IT equipment itself for valuable capacity.

For example, at the heat rejection level, a dry cooler may need to be substantially larger to deliver equivalent cooling capacity under high thermal loads. This has implications not only for physical space, but also for baseline material usage – more heat exchanger surface area, more fans, and more energy to move larger volumes of air.

While rising thermal density would traditionally come at the expense of efficiency, this is an area the industry is unwilling to compromise on. As a result, the pressure falls firmly upon physical footprint density.

Liquid isn’t everything, and air isn’t dead

Traditional air-cooled data center systems indeed face increasing strain as rack densities rise, but that does not automatically render them unfit for purpose. Rather, the challenge lies in how those systems are designed and deployed in these increasingly dense environments.

Cotton, responsible for shaping future product strategy and bringing innovative solutions to market, offers a more grounded perspective on the industry’s direction.

In practice, typical air cooling technologies remain viable at densities of around 30-40kW per rack, and in some cases can be extended toward 80kW. However, as Cotton notes, operating at those upper thresholds typically requires highly specialized equipment and bespoke cooling configurations. This does not represent a hard technical limit so much as a point at which standard approaches must evolve.

“A lot of that comes down to air being a fundamentally less dense fluid with lower specific heat capacity. Beyond a certain thermal load, the required airflow and temperature reduction become increasingly impractical, making the removal of intermediate heat exchange a key opportunity for improving thermodynamic viability,” he says.

Rather than signaling the obsolescence of air-cooled and chiller-based systems, these thresholds highlight the need for a more integrated and expert-led approach that evolves existing infrastructure, blends technologies where appropriate, and applies cooling capacity more precisely where it is needed most.

Densification is undoubtedly accelerating the industry’s shift toward liquid-based and hybrid cooling approaches, but in reality, this is not a wholesale reinvention. In practice, it reflects a gradual evolution of established technologies. At the same time, adoption is uneven, with some operators moving more aggressively toward liquid cooling while others take a more measured approach.

Even in liquid-cooled environments, air-based systems continue to play an important supporting role. Not every workload requires ultra-high-density design at all times, just as not every deployment moves away from standard-density infrastructure.

“The shift from air to liquid cooling, and from standard to high-density environments, is not a binary switch. It is a continuum,” says Cotton.

Sustainability is another area where perception and operational reality often diverge. Data center cooling is often assumed to be inherently resource-intensive, with water usage in particular drawing scrutiny. Yet, as Cotton notes, many modern dry or hybrid systems operate without continuous water consumption while still delivering high efficiency.

Ultimately, cooling evolution is best understood through lifecycle performance rather than isolated metrics, with these combined trends pointing clearly toward a “hybrid-by-design” future for the industry.

Modern capacity for modern density

The drive to deliver more cooling capacity within a smaller footprint is now a defining factor in how modern chiller systems are designed and optimized. A high-capacity chiller requires design emphasis on achieving higher heat rejection density within the same (or reduced) physical and energy envelope.

Providing free-cooling and air-cooled chiller heat rejection for data centers increasingly means designing established systems to operate in new ways, or integrating them more seamlessly across thermal domains. Cotton explains that the development of Airedale’s AI-optimized TurboChill 3MW reflected a deliberate high-density-first design philosophy:

“The product was conceived entirely around high-density requirements. Could it be used for standard density? Yes – but the design process was geared purely towards high-density applications.”

One of the key engineering decisions was to prioritize heat exchanger surface area rather than relying solely on increased airflow. By expanding the available surface area for heat transfer, the system can achieve higher levels of heat rejection without driving fans harder, helping to reduce both energy demand and noise.

Furthermore, the system was designed to be “clean system ready,” enabling – where application requirements allow – the removal of intermediate heat exchange between internal and external loops. With more efficient water treatment, filtration, and material selection, this approach can reduce additional thermal losses while enabling operation at more favorable water temperatures.

The result is not a reinvention of the chiller, but an optimization of how established cooling principles are applied under higher thermal loads. At its core, the system remains recognisably a chiller-based architecture – one adapted to meet the realities of high-density compute, rather than replaced outright.

Customization at scale

When design teams are working with technologies they have not previously deployed, a natural degree of conservatism around design and safety can act as a barrier to deployment. In many cases, this results in infrastructure that is overengineered relative to the actual workload requirements.

For more traditional or converged use cases, suppliers can continue to rely on highly standardized platforms, as customer requirements tend to fall within a relatively predictable operating range. Hyperscalers, however, increasingly span both ends of the spectrum – balancing conventional cloud infrastructure demands alongside rapidly growing requirements for high-density and AI-driven workloads. Expert-led deployments can more effectively combine this standardization with targeted customization, as Cotton explains:

“You can standardize the technologies, have that validated, and then operate with variants of that to the customer. We call it ‘custom at scale’ for tailored designs built on proven platforms and deployed in high volumes. That's quite unique for us.”

By building on established technologies and validated engineering practices, systems can adapt to different deployment requirements without abandoning the efficiencies of scale manufacturing and familiar operational models.

“Those technologies are things our ops teams understand very well today. We already have the technologies and capabilities at our disposal, along with the ability to deliver them within strong timeframes. This allows us to continue delivering at pace and at scale,” Cotton adds.

Looking forward to AI integration

For colocators and neocloud providers offering AI compute as a service, things become a little more complex. These organizations introduce new requirements and operating models, prompting suppliers to adapt validated product platforms into multiple variants aligned to each specific need. This is where hybridization becomes particularly important.

Actually implementing these methods is increasingly shaped by workload specificity. Rather than a single, steady-state hybrid model, Cotton suggests facilities will align more closely with – and in some cases be dedicated to – the compute applications they support.

“AI training, standard-density cloud compute, and inference each place very different demands on infrastructure. At scale, I think we’ll see these workloads move into dedicated facilities. Large operators already have the capacity to do this – building sites purely for AI training, or others focused entirely on traditional cloud,” he says, adding:

“In the past, enterprise data centers didn’t have that scale, and AI wasn’t as prevalent, so there was no need to separate workloads in this way.”

This shift could see hybrid environments give way to more defined, “as a service” models similar to mature cloud platforms. Where operators gain greater ability to tune cooling system parameters, there is scope for more precise and effective control as rack densities rise.

Evolution, not reinvention

Through incremental upgrades and targeted architectural reconfiguration, established technologies are being combined into more capable hybrid systems. Leveraging proven products and engineering them more effectively can accelerate deployment, but, crucially, speed alone does not guarantee better outcomes, as Cotton explains:

“The industry’s desire for speed pushes toward standardization – data center owners can buy chillers and generators en masse with uniform designs. But across those deployments, inevitably, some sites will not receive the best design for their specific conditions because important nuances have been compromised.

“Speed doesn’t necessarily equate to effectiveness, but with more considered design, it is still possible to achieve solutions that work broadly, without losing critical local optimization.”

As densification intensifies, it does not render established technologies like air cooling obsolete. Instead, it extends operating envelopes while creating new opportunities to refine how existing systems are applied. Cotton concludes:

“The industry’s need for reliability and resilience can sometimes drive step changes that look like reinvention, as new approaches are tried and tested. There is always a place for that kind of reinvention, but it needs to run in parallel with continuous evolution – building on what already works, rather than replacing it outright.”

Speak to an Airedale expert about your high-density transition. Find out more at www.airedale.com.