RACA Journal July 2026 | Page 16

DATA CENTRE FEATURE SERIES

Data centres

COOLING AT SCALE

Compiled by Eamonn Ryan

Understanding the thermal demands of modern data centres.

Background image by Ismail Enes Ayhan | Unsplash

As digital infrastructure evolves, so too do the cooling requirements that underpin it. Data centres, once designed around relatively modest thermal loads, are now facing rapidly escalating demands driven by artificial intelligence( AI), highperformance computing and cloud services.

At the most fundamental level, data centre cooling is governed by a simple principle: the heat generated by IT equipment must be removed at roughly the same rate. In practical terms, cooling capacity must closely match IT load, making thermal management a core design constraint rather than a secondary consideration.

Rack densities illustrate how quickly this challenge is intensifying. Traditional enterprise environments are typically operated at 5 – 10kW per rack, with more recent standard deployments in the 20 – 30kW range. Today, AI-driven workloads are pushing rack densities beyond 40kW, with some highperformance environments exceeding 100kW per rack.

These increases are not incremental – they are exponential. As compute power rises, so too does the demand for precision cooling, redundancy and infrastructure resilience.

Redundancy is a critical design requirement. Cooling systems are typically engineered with backup pumps, standby units and redundant cooling modules to ensure continuity. Even minor thermal excursions can impact equipment reliability, as elevated temperatures accelerate component degradation and increase the risk of failure.

Physical constraints further complicate the picture. Higherdensity racks are heavier and require reinforced flooring, while cooling infrastructure – such as chillers, dry coolers and coolant distribution systems – adds additional spatial and structural demands. In some cases, existing facilities must be re-engineered to accommodate these changes.

For smaller or decentralised deployments, modular or micro data centres offer a flexible alternative. These systems integrate cooling within a self-contained unit, but are generally limited to lower densities – typically around 20 – 30kW per rack using air-based systems.

IMMERSION VERSUS AIR AND LIQUID COOLING As rack densities rise and AI workloads become more demanding, traditional air cooling is increasingly being supplemented – or replaced – by liquid-based alternatives. The evolution from air to liquid cooling is not abrupt, but it is clearly underway.

Air cooling remains the most widely deployed solution due to its simplicity and familiarity. However, it becomes less effective as densities increase. While advanced airflow management techniques can extend its viability, air cooling is typically best suited to lower-density environments, generally in the range of 20 – 30kW per rack, with higher levels becoming increasingly difficult to manage efficiently.

Liquid cooling addresses this limitation by removing heat closer to its source. In direct-to-chip systems, coolant is circulated directly to high-heat components such as CPUs and GPUs, significantly improving thermal efficiency. These systems are increasingly used in environments exceeding 50kW per rack, particularly in AI and high-performance computing applications.

Immersion cooling takes this a step further by submerging IT hardware in a thermally conductive fluid, that efficiently absorbs heat while remaining electrically non-conductive, allowing components to be cooled safely and far more effectively than with air. Research shows that immersion-cooled systems can achieve substantially higher power densities than air-cooled environments, while also improving energy efficiency metrics such as power usage effectiveness( PUE).

Despite these advantages, immersion cooling introduces operational complexities. Maintenance procedures are more involved, requiring specialised handling of submerged equipment and careful fluid management. Filtration and fluid quality are critical, as contaminants can compromise system performance.

It is also important to note that most liquid cooling implementations are hybrid in nature. Even when heat is removed at the chip or server level, it must still be rejected from the facility, typically via chillers, dry coolers or other heat rejection systems.

The choice of cooling method ultimately depends on application requirements, available space and long-term scalability goals. High-density and space-constrained environments are increasingly favouring liquid-based solutions, while air cooling remains viable for less intensive workloads.

MAINTENANCE BEST PRACTICES FOR DATA CENTRE COOLING SYSTEMS In modern data centres, cooling system reliability is missioncritical. Even brief temperature excursions can result in equipment damage, reduced lifespan and breaches of service level agreements( SLAs). As a result, maintenance strategies must be both robust and carefully managed.

14 www. refrigerationandaircon. co. za RACA Journal I July 2026