Liquid cooling becomes the practical route
Air can cool high-TDP chips in principle, but only with taller heat sinks and therefore taller servers. The example on page 10 of the white paper shows a 42U rack holding 32 GPUs in air-cooled 10U servers, compared with 168 GPUs in liquid-cooled 2U servers. Direct-to-chip cooling brings its own demands: tighter control of coolant temperature, flow and chemistry, a coolant distribution unit to separate facility water from the server loop, and integration challenges caused by the lack of industry standards. The report recommends a design assessment of the planned liquid-cooled loads and the existing facility before any deployment.
Racks, floors and cabling must be specified for heavier loads
High-density AI racks can weigh over 1,800 kg. The report recommends racks at least 750 mm wide, 1,200 mm deep and 48U high, with a static capacity above 2,270 kg and dynamic capacity above 1,600 kg, validated by an independent third party. Concrete slab floors should be checked for loads over 3,000 kg, and high-temperature C21/C22 connectors are advised where rack temperatures exceed the rating of standard C19/C20 connectors.
Supply chain, software and services reduce risk
With lead times lengthening, the report advises specifying physical infrastructure alongside the IT stack, using standardised and prefabricated sub-assemblies, and working with fewer, well-established suppliers. Before retrofitting, a feasibility study using power monitoring (EPMS) and DCIM should establish actual peak draws and spare capacity, followed by capacity analysis, protection coordination, arc flash and short-circuit studies. Once live, a digital twin of the IT space helps operators avoid tripped breakers, hot spots and stranded capacity as safety margins shrink. The paper ends with a look at emerging technologies, including medium voltage distribution in the IT space, solid state transformers and circuit breakers, and higher-voltage racks.
