Rack density for AI: what changes when racks draw 40 kW or more

Rack density is the power one rack draws, in kW. Ordinary enterprise racks use a few kilowatts; racks of AI and GPU servers often need 40 kW or more. Every kilowatt becomes heat, so density decides how a hall must be powered and cooled. A site with plenty of total capacity can still be unable to host dense racks if its distribution and cooling were designed for lower loads.

Air cooling works well at lower densities and can be stretched with hot and cold aisle containment. As density rises, moving enough air becomes the limit. The usual next steps are:

• Rear-door heat exchangers: water-cooled doors that take heat out of the air as it leaves the rack.
• Direct liquid cooling: coolant piped to cold plates on the chips, now common for current GPU servers.
• Immersion: servers submerged in a non-conductive liquid.

Liquid cooling needs pipework, coolant distribution units and a water loop in the hall, which is hard to add to a building that was not planned for it.

Dense racks need more power delivered to each position: higher-rated feeds, more circuits per rack and busways sized for the load. Ask what each rack position can draw on its own, not only the hall average. A hall rated at a modest average can sometimes host a few dense rows, but only if the distribution allows it.

It can. Fully loaded GPU racks, especially with liquid cooling, are heavy. Raised floors and older buildings have load limits, so ask for the floor loading per square metre and whether racks stand on a slab or a raised floor.

State the density per rack, the number of racks, and the cooling your hardware needs, for example direct liquid cooling with a given supply temperature. On the marketplace, density and cooling are standard fields in every request and listing, and both count toward the fit score. If you are choosing between buying racks and renting GPUs, our guide to GPU compute is a good next read.