13 November 2025

Paul Mouton, Afrikamasts

Cooling a data centre has always been a key engineering discipline, but the rise of AI produced heat loads has placed thermal management at the heart of infrastructure strategy.

Traditional server environments, even those with relatively dense configurations, now look modest compared to the high-power, thermally intense AI clusters, powered by PDUs and GPUs. This shift has forced engineers to rethink airflow, liquid cooling systems, redundancy, and long-term sustainability.

In the third quarter of 2019 as build contractor, we were making ready to roll out a state of the art, multi module data centre for a high-end South African client in the data storage market. In 2020 the construction was completed boasting cold aisle containment and the latest CRAC methodology together with a complex air handling system.

In just under five short years, that same DC has become redundant to the latest requirements and will require substantial retrofitting and re-engineering to cater for today’s AI ready rack space. Retrofitting an existing site with water-based cooling is costly. It involves structural changes and careful design to manage the unplanned placing of chillers and piping, and often there are architectural and engineering constraints which will make it almost impossible to achieve.

Older cooling approaches were based on heat exchange systems based on Computer Room Air Conditioners (CRAC) and Computer Room Air Handlers (CRAH) which have long been used in raised floor server rooms. These systems worked adequately when rack densities were under 5 kW per rack. But the landscape changed dramatically with the rise of hyperscale and AI infrastructure. Nowadays, AI ready data centres often need to support power densities between 30-80 kW per rack, especially in high-density CPU GPU configurations. A single AI training cluster can draw the same power as an entire row of traditional servers. This has made older air-based systems increasingly inadequate.

Water is over 3,000 times more efficient at carrying heat than air, making it far better suited for high-density environments. Modern data centres rely on chilled water loops supplied by central cooling plants. The chilled water flows through CRAH units, in-row coolers, or even directly to components via cold plates. In direct to chip cooling, cold plates are mounted directly to CPUs, GPUs, and memory modules, transferring heat to circulating liquid. This allows rack densities to reach 80–100 kW without depending on heavy airflow. A more advanced option is immersion cooling, where entire servers are submerged in special non-conductive fluids. While still not widely adopted, and almost unheard of in Africa, immersion drastically reduces the need for fans and air ducts. However, water-based cooling comes with its own challenges, ensuring leak-proof and redundant piping. Thermal Load requires managing intense heat in a compact footprint is a major challenge. Cooling 80 kW per rack effectively demands hybrid systems that mix air and water, steered by smart controls.

AI platforms are evolving so rapidly that cooling systems need to scale without major redesigns. Modular infrastructure and future-proofing are essential. Cooling can account for up to 40% of a data centre’s electricity bill. Keeping power usage effectiveness (PUE) as close as possible to 1.0 while still ensuring uptime and safety, is an ongoing balancing act. For example, a 500-rack AI centre running at 50 kW per rack would demand 25 MW of IT power. At a PUE of 1.3, the total facility load would reach 32.5 MW.

Cooling is a strategic layer in AI infrastructure. As machine learning and data science scale up, the complexity of cooling follows suit. Whether through chilled water systems, direct to chip liquid loops, or real-time thermal analytics, the future of data centres is being shaped by thermodynamics. On a continent where the power grid is unreliable and load sheading ever frequent, private power supply may be the only answer but only if such power farms can get the green light from failing government institutions.