Beyond air: Rethinking data center cooling for the AI era

Rack power density is no longer climbing - it is leaping. Where 5-8 kW per rack was once the norm, AI training clusters now routinely push beyond 150 kW, with some designs targeting 300 kW. Traditional air cooling, already struggling at 20 kW, has hit a hard physical ceiling. In tropical markets like Singapore and Malaysia, where humidity further suppresses air's heat-carrying capacity, the problem becomes even more acute.

The industry faces a fundamental question: how do we cool the next generation of AI factories without sacrificing energy efficiency, reliability, or time‑to‑market?

The shift from components to systems

For years, data centre cooling focused on individual components - a chiller here, a fan wall there. But at 150 kW per rack, component-level thinking fails. What is needed is a system‑level approach that treats heat as a continuous flow from the server chip to the outside environment.

This means integrating cold sources, distribution units, and terminal cooling into a single, orchestrated chain. It also means rethinking the traditional separation between chiller plants and rack‑level liquid distribution.

Three technology pillars

Based on recent industry demonstrations, several innovations are moving from lab to field.

Midea maglev active CDU - The most radical shift is combining the magnetic bearing chiller and the CDU into one integrated chassis. This can significantly simplify the meidium and high temperature liquid cooling architectures. The result: floor space cut by more than half, tropical PUE below 1.2 (with high‑temperature water and indirect free cooling), and annual energy efficiency ratio above 46. Because the system is largely factory‑prefabricated, on‑site installation time drops by over 50%, directly speeding up AI infrastructure deployment. Thanks to Midea's in-house maglev centrifugal compressor technology, the unit can deliver a wider operation range which also make it compatible with the hybrid air-liquid cooling architecture. What's more the unit not only support AC power supply, also natively supporting high voltage DC power supply.

Midea air-cooled magnetic bearing chillers - In water-scarce or space‑constrained sites, these units deliver a COP of 5.4 even at 35°C ambient. Because they run completely oil‑free, they avoid the 15-25% efficiency degradation seen in conventional chillers after a decade. More importantly, they excel at low load (10-30%), which matches the gradual ramp‑up of most AI data centres. A 100-second restart capability also adds resilience during grid fluctuations.

Midea high-reliability CDUs - As the "heart" of liquid cooling, cooling distribution units must handle massive heat loads without becoming a single point of failure. Today's industrial‑grade CDUs can manage 2,600 kW (at 3K ΔT) with dual‑stage filtration (25 µm + 0.2 µm) and hot‑swappable pumps. N+2 redundancy sensors and A+B power inputs keep maintenance near zero‑downtime.

Beyond PUE: total cost of ownership

While PUE remains a useful benchmark, operators in the tropics must look at total cost of ownership. A solution that over‑provisions for the few hottest hours of the year can cripple operational budgets. Integrated designs with high‑temperature water and free cooling allow the chiller to stay idle for most of the year, slashing both energy bills and mechanical wear.

A manufacturing backbone

None of this works without consistent quality and volume. Building a reliable liquid-cooling supply chain requires dedicated manufacturing capacity for core components like CDUs and cold plates. Investment in automated lines and strict process control - from welding cleanliness to leak testing - ensures that the "cooling certainty" promised in design is delivered on site.

In March this year, Midea's liquid-cooling smart manufacturing base with over RMB 1 billion in investment began construction. It is designed to mass-produce magnetic bearing chillers, CDUs, and other core products at scale, creating an end-to-end capability from R&D and manufacturing to delivery. The goal is to bring core processes in-house, secure supply chain independence, and accelerate the deployment of advanced cooling solutions for AI data centres.

The takeaway

AI's appetite for compute is not slowing down. The cooling industry must respond with reliable, integrated, efficient, and scalable systems - not just better fans or larger chillers. For tropical data centres, the path forward lies in active magnetic bearing technology, highly integrated CDU architectures, and a relentless focus on total cost of ownership. Those who master this system‑level shift will become the true enablers of the AI era.