AI data centre liquid cooling is moving from specialist technology to core infrastructure. The reason is simple: modern AI racks concentrate far more computing power—and therefore far more heat—than conventional server rooms were designed to handle.
New rack-scale systems are being designed together with their cooling, power and networking. That shift changes what a data centre is: less a building filled with interchangeable servers, more a tightly engineered industrial machine.
Why air cooling is reaching its limits
Fans and chilled air work well when heat is spread across many modest-power servers. AI accelerators pack large numbers of transistors and high-bandwidth memory into dense systems. When dozens of accelerators are connected in one rack, moving heat into room air becomes inefficient and noisy.
Liquid carries far more heat than air. Cold plates can sit directly on processors and memory, moving heat into a facility water loop without immersing the electronics. Other designs submerge hardware in dielectric fluid, although immersion remains less common in mainstream hyperscale deployments.
How liquid cooling fits next-generation AI racks
Nvidia’s Vera CPU rack design integrates 256 liquid-cooled processors for large numbers of AI-agent environments. The broader Vera Rubin platform connects CPUs and GPUs with high-bandwidth links, turning the rack into a coordinated computing system.
The company’s DSX infrastructure framework goes further by coordinating compute, cooling and power. Nvidia says one approach uses 45°C cooling water and workload controls to operate more processors within a fixed power budget. Those performance figures are vendor claims and will vary by facility and workload.
Warm-water cooling can be an advantage
Liquid does not always need to be chilled to refrigerator temperatures. If chips can operate with relatively warm inlet water, a facility may reduce or avoid energy-intensive mechanical chillers during suitable weather.
Warmer output water is also easier to reuse for building or district heating. Whether that is practical depends on location, nearby heat demand and infrastructure economics.
Liquid cooling does not eliminate water questions
A closed internal loop can circulate the same fluid repeatedly, but heat must still leave the facility. Cooling towers may consume water through evaporation, while dry coolers use more electricity in hot weather. The best design depends on climate, water availability, electricity prices and local constraints.
Leaks, corrosion, pump reliability and maintenance procedures also become critical. Operators need sensors, redundant loops and technicians who understand both computing and mechanical systems.
Why the cooling race matters
AI performance is increasingly limited by megawatts and heat removal, not only by how many chips can be purchased. A more efficient cooling system can fit more useful computing into the same electrical envelope.
That makes thermal engineering a strategic technology. The next generation of AI data centres will be judged not simply by processor count but by how reliably they turn power, water and cooling capacity into useful computation.
The cooling method is only part of the resource picture. Our guide to AI data centre water use explains how facility design changes what water figures mean.
Sources and further reading
Reporting note: performance figures from Nvidia describe company products and reference designs. This article is informational and does not provide investment advice.
Leave a Reply