By Willem Weber, Mechanical Engineering Consultant for Digital Parks Africa

As enterprises accelerate their adoption of Artificial Intelligence (AI) and High‑Performance Computing (HPC), many are discovering that the physics of the data centre are no longer on their side. Rack densities are soaring, heat loads are spiking, and power availability is shrinking. The result is a hard stop for traditional design thinking.

This is particularly true when it comes to cooling, which is no longer a background utility but a critical engineering consideration separating organisations capable of sustaining AI growth from those that fall behind.

Traditionally, data‑centre cooling was relatively straightforward. Cold air was pushed through a raised‑floor plenum, delivered via perforated tiles into the cold aisle, and servers’ internal fans pulled it through racks that typically operated at between 5kW and 15kW. The entire ecosystem was built around this air‑cooled model. Once a facility hit roughly 500-600kW, chilled water became the more efficient option, but the principle stayed the same: cool the air and move the air.

When air becomes impractical

However, AI has changed all of that, with racks now drawing 60-100kW, and some deployments even reaching 150kW in a standard 600-800mm wide, 2.2m‑deep cabinet. At those densities, air cannot carry away enough heat. The airflow and pressure required to cool a 100kW rack are physically and energetically impractical.

This is why the industry is shifting to liquid cooling. Liquids such as water or dielectric fluids have a significantly higher heat-carrying capacity than air, enabling far more efficient heat removal. Because pure water is conductive, modern systems increasingly use dielectric fluids or water‑glycol mixes with engineered containment systems to manage leaks safely.

This shift transforms the entire cooling architecture. Fans give way to pumps, airflow management shifts to coolant distribution, and Cooling Distribution Units (CDUs) now manage secondary liquid loops that feed the racks.

Fan walls still have a place in conventional air‑cooled environments, but the technology landscape has shifted. AI‑era densities demand liquid cooling, and the challenge now is how quickly operators can adapt.

Rethinking the physical design

Shifting from air to liquid cooling does not just change the cooling method; it changes the entire physical design of the data centre. Liquid systems require a CDU to run the secondary loop, which is a heavy unit built around dual pumps and plate‑to‑plate heat exchangers, adding significant structural load.

The racks themselves also get heavier, as the combination of high‑density servers and the coolant circulating through them increase the weight per cabinet well beyond that of traditional air‑cooled systems.

Heat rejection adds another layer of complexity. In cooler climates, dry coolers may be sufficient. In hotter regions, operators may require cooling towers or chillers, both of which are extremely heavy, especially when installed on a roof. The building must be engineered to carry this additional load.

Furthermore, the distribution network is also substantial. At today’s heat loads, data centres need to run roughly 300mm pipework, filled with fluid. That is a massive amount of weight threaded through the structure.

Balancing higher rack densities with energy efficiency starts with choosing the right cooling strategy for your environment. Moving to liquid cooling is the first step, but efficiency depends heavily on how you reject the heat.

Location matters just as much. Ambient temperature directly determines how much mechanical cooling is required. A cold climate gives you free efficiency; while a hot climate forces you into more energy‑intensive options.

Why scalability matters

A common mistake in data‑centre design is installing the full cooling plant for a 1MW facility on day one, even though the site may start at only 10-20% of its full load. Large chillers with screw compressors are efficient at full capacity, but they perform poorly at low loads; exactly when the data centre needs efficiency the most. The same applies to oversized fans that can only throttle down to a fraction of their rated output.

The solution is scalability. Deploy multiple smaller compressors or modular cooling units with true variable‑speed capability, so the system can match early‑stage loads without wasting energy. Apply the same logic to airflow equipment and avoid single, oversized fans that cannot turn down far enough.

Ultimately, the rise of AI leaves organisations with no real choice – cooling systems must evolve, and the only question is how quickly they can adapt to liquid‑based architectures. For legacy facilities, the priority is to eliminate inefficiencies:  minimise leaks, reduce thermal losses, and upgrade pumps and supporting equipment to handle modern loads.

Whether designing new capacity or retrofitting existing infrastructure, the path forward is the same. Enterprises must embrace liquid cooling and optimise every component around it. This is now the baseline for operating in an AI‑driven world.