How many servers fit in a cabinet? U planning, power budget and out-of-band management

The first question from teams moving their own hardware into a data centre is almost always the same: "how many servers can I put in a 21U cabinet?" The answer on paper is twenty-one. The real answer is between twelve and fifteen — and knowing where the gap comes from lets you plan the move correctly from the start.

What is more, height is often not even the real limit.

The unit first: what a U is

Cabinet height is measured in rack units, and 1U is about 4.45 cm. A typical 1U server occupies exactly that; a 2U server takes two. A half cabinet is 21U, a full one 42U.

The 19-inch rack width is standard, but depth is not. A long chassis, or the cable slack that has to be left behind it, can change the layout plan, and you need to state your mounting rail type up front too (fixed shelf or sliding rails). A tower machine cannot be rack-mounted directly at all.

What goes in the rack besides servers

If you plan a cabinet as nothing but a stack of servers, the arithmetic will not hold. Here is the realistic table for a 21U cabinet:

So twelve to fifteen servers if they are 1U, and half that with 2U servers. The same logic gives 28 to 32 servers in a full 42U cabinet.

Cable management is needed even when it is left off the list: a 1U organiser for every four to six servers is a practical rule that keeps cabling out of the airflow. Choosing a vertical PDU buys U directly, because it mounts on the side of the cabinet.

Remaining gaps get blanking panels. They are not there for looks: they stop cold air leaking around to the back of the servers. Without them, the devices at the top of the rack run noticeably hotter.

The real limit: not how many U, but how many watts

Height decides how many servers fit in a cabinet; heat decides how many run without trouble.

Data centres work on a hot aisle / cold aisle layout: servers take cold air at the front, exhaust warm air at the back, and every device has to face the same way. A single unit mounted backwards pulls its neighbour's hot exhaust into its own intake, and the whole rack pays for it. The widely accepted ASHRAE guidance is to keep server inlet air between 18 and 27 °C; going outside that band raises both failure rates and the electricity the fans themselves consume.

A densely packed cabinet draws more power and produces more heat at the same time. So the power budget allocated per cabinet may well be the limit that fills before the U count does. When you plan, ask how many watts are left as often as how many U are free.

Report power as real draw, not the label rating. A power supply marked 750 W does not mean the server pulls 750 W continuously; typical draw on most 1U servers sits well below it. An inflated figure reserves space you do not need, an understated one forces a layout change. Do not forget the plug type either: server power cords are usually C14 while higher-power devices use C20 inlets, and with a dual-PSU server how both cords will be fed has to be settled at the quote stage.

Remote hands: your hardware, your hands elsewhere

The most frequently overlooked part of hosting your own hardware is that the machine is physically far away from you. While the operating system boots, SSH or RDP is enough. But what if the system never comes up, if a BIOS setting has to change, or if you need to reinstall from scratch?

The answer is the out-of-band management controller: iDRAC on Dell, iLO on HPE, IPMI on Supermicro. It runs over a separate (or shared) network port and is reachable even when the machine is powered off; it shows the console, lets you mount virtual media and install from an ISO, and can power-cycle the hardware when needed.

Do two things before delivery: change the controller's default password and bring its firmware up to date. These interfaces are among the first things scanned when left exposed on the internet.

A second set of remote hands is a PDU with individually switchable outlets. When the management controller itself stops answering, that is usually the only way to power-cycle the hardware without sending someone on site.

Colocation or cabinet rental

They are not the same thing. Colocation means taking space per server in a shared cabinet; cabinet rental means the whole rack is allocated to you — you decide the layout, the switch and the cabling, the cabinet is locked and access is logged.

In practice the threshold is crossed when you pass a handful of servers and want to build your own network design. Reaching the same capacity without buying hardware at all is another option; the comparison that puts the two cost structures side by side with current amounts is on the which works out cheaper, a dedicated server or a cabinet page.

If you also hold your own IP range, one more option opens up: the range can be announced through the Datafex network and attack filtering happens at the network layer — the detail is in BGP protection with your own ASN.

Labelling and growth

Changing cabinets means moving: cables come out, servers shut down, and bringing everything back up needs a maintenance window. So the size decision is made against your twelve-month estimate rather than today's server count, with at least a fifth added on top.

When the cabinet fills up there are two ways out: take a second cabinet and link the two over an uplink (non-disruptive), or move into a larger one (needs a maintenance window).

Finally, write the map down. Which U each server occupies, which switch port it is patched into and which PDU outlet feeds it should all be recorded. Without that map, a remote instruction to "restart the third server" turns into guesswork.