ARTIFICIAL INTELLIGENCE
Engineering the data
centres that make AI possible By Darren Watkins, chief revenue officer at VIRTUS Data Centres
W
hen most people talk
about artificial intelligence (AI), the conversation quickly turns to algorithms, datasets or cloud platforms. Engineers tend to see it
differently. Behind every breakthrough model sits a physical environment consuming vast amounts of power and generating equally vast amounts of heat.
That reality is becoming impossible to ignore. Training a large-scale model can mean running thousands of graphics processing units (GPUs) in parallel for weeks, pushing power and cooling systems far beyond what traditional enterprise facilities were designed to handle. Inference - the stage where models are deployed into real-world use - adds another constant layer of demand. This shift is forcing data centres to evolve into something entirely new. What once worked for office IT and web hosting is no longer viable. Meeting the requirements of AI depends on a new generation of facilities designed specifically for density,
16
liquid cooling and resilience. For engineers, that means treating AI not as just another workload, but as a driver of entirely new infrastructure principles
Legacy facilities meet their limits For decades, enterprises ran IT workloads on-premise or in shared colocation spaces designed for predictable tasks such as ERP, payroll and web hosting. These applications had steady and relatively low power profiles. A typical rack would draw 2 – 4 kilowatts, with conventional air cooling providing adequate thermal management.
AI workloads are fundamentally different. Inference, where models are used in production, is continuous and often latency sensitive. Together, these workloads create both sustained high draw and unpredictable spikes, which most legacy environments simply cannot cope with.
Modern AI racks can require 50 – 80 kilowatts or more. Air cooling systems saturate under these conditions and electrical distribution networks built for office IT cannot deliver the required power density. Even structural issues such as floor loading, aisle spacing and airflow containment
DECEMBER/JANUARY 2026 | ELECTRONICS FOR ENGINEERS
become barriers. To facilitate these racks, retrofitting is possible but expensive, disruptive and rarely future-proof. As a result, purpose-built facilities are emerging as the practical answer.
Cooling: from air to liquid Managing heat has become one of the most pressing engineering challenges in AI data centres. Once rack power moves beyond roughly 50 kilowatts, the limits of conventional air cooling are quickly exposed. Fans and raised floors cannot shift the heat fast enough or evenly enough, creating hotspots that put both performance and reliability at risk. Liquid cooling is emerging as the practical answer because it moves heat more efficiently than air. In direct-to-chip systems, coolant is channelled through cold plates that sit directly on processors. Immersion systems go further, submerging entire boards in non-conductive fluid. Both methods have advantages. Direct-to-chip designs are often easier to add into existing environments, while immersion cooling achieves higher efficiency but requires facilities to be planned around tanks, pumps and containment from the outset.
Page 1 |
Page 2 |
Page 3 |
Page 4 |
Page 5 |
Page 6 |
Page 7 |
Page 8 |
Page 9 |
Page 10 |
Page 11 |
Page 12 |
Page 13 |
Page 14 |
Page 15 |
Page 16 |
Page 17 |
Page 18 |
Page 19 |
Page 20 |
Page 21 |
Page 22 |
Page 23 |
Page 24 |
Page 25 |
Page 26 |
Page 27 |
Page 28 |
Page 29 |
Page 30 |
Page 31 |
Page 32 |
Page 33 |
Page 34 |
Page 35 |
Page 36 |
Page 37 |
Page 38 |
Page 39 |
Page 40 |
Page 41 |
Page 42
