By Willem Weber, mechanical engineering consultant, Digital Parks Africa
This is part two of a two-part series. The shift towards liquid cooling in AI data centres is not simply a matter of replacing air with water or another coolant. Once rack densities reach the levels associated with modern AI workloads, the cooling system begins to influence the physical design of the entire facility.
The transition will not happen in a single step, and conventional air cooling will remain important. DC Studio | Magnific.com
This has implications for structural loading, pipework, heat rejection, energy efficiency and, increasingly, how data centres are designed for future expansion. A liquid-cooled data centre requires additional infrastructure to circulate and manage the coolant.
Cooling Distribution Units (CDUs), for example, incorporate pumps and heat exchangers to manage the liquid loops supplying the racks. These units add weight and require dedicated space within the facility.
The racks themselves also become heavier as high-density computing equipment is combined with liquid-cooling hardware and coolant.
Heat rejection introduces another consideration. In cooler climates, dry coolers may provide an effective means of rejecting heat. In hotter environments, however, the system may require chillers, cooling towers or a combination of technologies.
For African facilities, this makes the local climate an important part of the cooling strategy. A data centre located in a cooler climate may be able to take advantage of lower ambient temperatures for more of the year, while facilities in hotter regions may require substantially more mechanical cooling. The building must also be designed to accommodate the associated equipment and loads.
The weight of the cooling system
As cooling capacity increases, so does the physical scale of the distribution infrastructure. High-density installations can require large-diameter pipework carrying significant volumes of fluid. This adds another structural consideration, particularly where distribution systems run through or above the data hall.
The cooling plant itself can also be substantial. Chillers, cooling towers, pumps, heat exchangers and associated equipment all contribute to the structural and spatial requirements of the facility. These considerations need to be addressed during the initial design rather than treated as an afterthought.
The result is that mechanical and structural engineering become increasingly interconnected as rack densities rise.
Heat rejection matters
Liquid cooling improves the ability to remove heat from the server, but that heat ultimately has to be rejected from the facility. This is where the overall efficiency of the system becomes important.
Moving from air cooling to liquid cooling does not automatically produce an efficient data centre. Pumps consume electricity, heat exchangers introduce temperature differences and the final heat-rejection system determines how effectively the captured heat can be discharged to the environment.
Ambient conditions therefore matter. In cooler conditions, a dry cooler may be able to reject heat with limited mechanical intervention. In hotter conditions, chillers or evaporative systems may be required, increasing the energy demand of the cooling plant.
For African data centres, the optimal solution therefore needs to account for local temperature profiles, water availability, electricity costs and the required level of resilience.
Designing for scalability
Another important consideration is how the cooling system responds as a data centre is built out. A facility may ultimately be designed for a megawatt-scale IT load but initially operate at only a fraction of that capacity. Installing the entire cooling plant at the outset can result in equipment operating well below its optimum load for an extended period.
Large compressors, fans and other equipment may be highly efficient at their design point but less efficient when repeatedly operated at very low loads.
A modular approach can help. Multiple smaller cooling units, combined with variable-speed compressors, pumps and fans, can allow cooling capacity to follow the actual IT load more closely. As additional computing capacity is installed, additional cooling modules can be brought online.
This approach is particularly relevant to AI facilities, where demand can increase rapidly and where operators need the ability to expand without redesigning the entire thermal infrastructure.
The African consideration
For Africa, these issues take on an additional dimension. Data-centre development is accelerating across markets including South Africa, Kenya, Nigeria and other emerging digital hubs. At the same time, operators face the practical constraints of local electricity infrastructure, climate and water availability.
There is therefore no single liquid-cooling architecture that will suit every African facility. A data centre in a hot, water-stressed environment may have very different requirements from one operating in a cooler climate with abundant water. Similarly, the economics of a hyperscale AI facility will differ from those of a smaller regional data centre.
The engineering challenge is consequently not simply to adopt liquid cooling, but to select the appropriate combination of rack-level cooling, distribution, heat rejection and controls for the specific site.
From cooling system to thermal architecture
AI is changing the role of cooling in the data centre. What was once largely a question of maintaining an acceptable room temperature is becoming a much more complex thermal-management problem, extending from the processor itself through the rack, coolant distribution system and heat-rejection plant.
For new African facilities, this creates an opportunity to design for higher densities from the outset rather than attempting to retrofit increasingly demanding workloads into buildings designed around an earlier generation of computing.
For existing facilities, the challenge is more difficult. Operators may need to address airflow inefficiencies, increase heat-rejection capacity, upgrade pumps and introduce liquid cooling selectively as high-density workloads are deployed.
The transition will not happen in a single step, and conventional air cooling will remain important. But as AI pushes rack densities higher, the industry will increasingly need to think beyond air conditioning and towards integrated thermal architecture.
