Rack System

At its core, a rack system is a tall metal cabinet for computing equipment. The devices inside it are not placed side by side on tables but rather slid in flat, one above the other, like drawers in a chest of drawers. The dimensions are standardized worldwide: the mounting width is 19 inches, or about 48 centimeters. The height is measured in rack units, abbreviated U, with one unit being just under 4.5 centimeters tall. A typical cabinet has 42 such units and is therefore about two meters tall. Because all manufacturers adhere to this standard, equipment from different companies fits into the same cabinet.

From Shelf to the Smallest Building Block of the Data Center

For a long time, a rack was just furniture. You bought individual servers, slid them in, and cabled them up. With AI applications, this has shifted. A large language model, that is, a program that generates text, no longer fits on a single computer. It has to be distributed across many specialized chips that work closely together.

That is why manufacturers today sell entire racks as a single product. Nvidia, for instance, markets its systems with dozens of graphics chips in one cabinet, internally connected so quickly that, from the software’s perspective, they behave like a single enormous chip. The price for such a rack runs into several million dollars. When financial news talks about delivery times or order backlogs, it is often precisely these units that are meant.

For operators, the rack is also the unit of measurement for space, power, and cooling. A data center is not planned in terms of individual devices but in terms of slots and kilowatts per slot. A classic rack with standard servers draws about 5 to 15 kilowatts. AI racks run at 100 kilowatts and more, which is as much as several dozen households.

Rails, Power, and the Limits of Air Cooling

Inside, the rack has perforated rails on both sides. Devices are attached there using brackets or extendable slide rails. At the very bottom usually sit power distribution units and batteries that bridge brief outages. At the top or in the middle sit the networking devices, so that the cables to all servers stay short. Cabling runs in channels at the back, while the front stays clear for access.

More important than the mechanics is the airflow. Racks stand in rows with their front sides facing each other. This creates cold and hot aisles: the devices draw in cool air from the front and blow the waste heat out the back. If the two mix, the temperature rises and the chips throttle themselves to avoid overheating.

For AI racks, air is no longer sufficient. That is why liquid cooling is used: water flows through metal plates that sit directly on the chips and carries the heat away. This explains why many older data centers cannot simply install new AI systems. The cabinets would fit, but the power supply and cooling loop are missing.

Where These Cabinets Actually Stand

Rack systems stand wherever data is processed: in the data centers of Google, Amazon, or Microsoft, in company basements, in universities, and in many schools. When you open a website or stream a video, a device sitting in such a cabinet is responding. Even small variants with six or twelve rack units exist, for instance for recording studios or clubhouses.

In the news, you mostly encounter the term in connection with AI infrastructure expansion. Companies state how many racks they can ship or install per quarter. This figure is considered an indicator of how quickly computing power actually becomes available. Ordered chips are of no use as long as the cabinet, cooling, and power connection are missing.

A common misconception is that a rack is itself a computer. To begin with, it is only a standardized enclosure. What is new, however, is that with AI systems the enclosure, chips, cabling, and cooling are planned as an inseparable package. In this sense, the rack has indeed become the machine.