Inside the white space: servers, storage and racks

A data centre exists to house IT equipment, and nearly every design decision follows from what that equipment needs: how much power each rack draws, how much heat it produces, how heavy it is, how it is cooled, how it connects to the network and how often it is replaced. Engineers who design the building rarely choose the servers, but they must understand them well enough to ask the right questions and to anticipate change, because a building lasts 20 to 30 years while the equipment inside is replaced every 3 to 6. This lesson opens the cabinets: the main kinds of IT equipment, how racks are arranged, the power and heat of conventional and AI workloads, and the numbers that turn a list of equipment into a design load.

The white space

The white space (also called the data hall, computer room or IT space) is the area that holds IT equipment. Its opposite, the grey space, holds the support systems: electrical rooms, UPS and battery rooms, mechanical plant, network entrance rooms, loading docks and offices. In a typical modern facility, the white space takes perhaps 30% to 50% of the gross floor area; high-density AI facilities have proportionally less white space and more grey space, because each square metre of white space needs more power and cooling equipment.

Within the white space, equipment is mounted in racks (or cabinets): standardized steel frames 600 mm or 800 mm wide, about 1,070 mm to 1,200 mm deep, and usually 42U to 52U tall, where one rack unit (U) is 44.45 mm (1.75 in). Racks are placed side by side in rows, with all equipment facing the same way so that cool air enters at the front and warm air leaves at the back. Rows face each other front to front and back to back, creating alternating cold and hot aisles (Lesson 5).

Servers

A server is a computer designed to run continuously in a rack. Its main parts are one or two processors (CPUs), memory (DRAM), storage drives, network interface cards, power supplies, and fans that pull air from front to back. Servers come in several forms:

• Rack servers: 1U or 2U general-purpose machines, the most common building block of enterprise and cloud data centres.
• Blade servers: thin servers that slide into a shared chassis providing power, cooling and networking.
• Multi-node and hyperscale servers: custom designs, often following Open Compute Project specifications, that share power supplies and fans across several nodes to reduce cost and energy.
• Accelerated servers: servers with graphics processing units (GPUs) or other accelerators for AI and high-performance computing, typically 4U to 10U with eight accelerators, or integrated into liquid-cooled rack-scale systems.

The power drawn by a server varies with its workload. Idle servers still draw a significant fraction of peak power (often 30% to 50% for older designs, less for modern ones), so lightly used servers waste energy; consolidating workloads onto fewer, busier servers through virtualization is one of the most effective ways to save energy in IT.

Storage and networking

Data is stored on hard disk drives (HDDs), which are cheap per terabyte but mechanical, and solid-state drives (SSDs), which are faster and use less power per operation. Storage is organized in arrays and distributed systems that keep several copies of data, often in different racks or buildings, so that a single failure loses nothing. Tape libraries remain in use for long-term archives because tape is cheap and draws no power when idle. Storage equipment is heavy: a fully loaded storage rack can exceed 1,500 kg, which matters for floor design (Lesson 9).

Networking connects servers to each other and to the outside world. In a common arrangement, each rack has a top-of-rack switch connected to aggregation or spine switches, which connect to routers and to the carriers and cloud providers outside (Lesson 7). AI training clusters add a separate high-bandwidth fabric connecting thousands of GPUs with very low latency, which requires large amounts of optical fibre and constrains how far apart racks can be.

Rack density

Rack density, the power drawn per rack, is the single most important number for a data centre designer. It determines the cooling method, the size of the electrical distribution, the floor loading and the layout. Densities have risen steadily:

• Traditional enterprise rooms: 3 kW to 6 kW per rack.
• Modern cloud and colocation halls: 8 kW to 20 kW per rack.
• Air-cooled AI and high-performance computing: 20 kW to 50 kW per rack, near the practical limit of air cooling.
• Liquid-cooled AI racks: 100 kW to over 130 kW per rack today, with vendors announcing designs of several hundred kilowatts per rack.

Because density is rising quickly, designers increasingly plan for a mix: lower-density air-cooled racks for storage and networking, and high-density liquid-cooled racks for accelerators, with infrastructure that can be adapted as the mix changes.

Nameplate, design and actual load

Equipment nameplates state the maximum power a power supply can deliver, not what the server actually draws. A server with two 1,600 W power supplies (one redundant) might be labeled 1,600 W but draw 600 W on average and 900 W at peak. Designing from nameplates alone would oversize the facility several times over. Designers therefore distinguish:

• Nameplate power: the rating of the power supply.
• Peak measured power: the maximum observed under heavy workloads, often 50% to 70% of nameplate for conventional servers but close to 100% for AI training.
• Average power: what determines energy use and cost.

The design IT load of a facility is set from the expected peak power of the planned equipment, with an allowance for growth. Because AI training runs keep thousands of GPUs at nearly full power in synchronized patterns, AI halls operate much closer to their design load than traditional halls, which reduces the safety margin that diversity used to provide and can produce rapid swings in power as training jobs start, pause and checkpoint.

Heat from IT equipment

Essentially all electrical power drawn by IT equipment becomes heat in the room (Lesson 1). Each server's fans move air through it, and the temperature rise across the server depends on its power and airflow:

where \(\rho c_p\) for air is about 1.2 kJ/(m³·K). Servers typically raise air temperature by 10 K to 20 K. Fans adjust their speed to keep components cool, so the airflow a server needs changes with inlet temperature and load, and the cooling system must supply at least what the servers draw (Lesson 6).

Power supplies and rack power

Each server converts AC (or, in some designs, DC) to the low voltages its components need using one or more power supply units (PSUs). Most servers have two PSUs fed from separate power paths (an A feed and a B feed), so that either path can fail or be maintained without stopping the server. Rack power distribution units (PDUs), long power strips mounted vertically at the back of the rack, deliver power to each device and often measure current per outlet for monitoring.

In North America, racks are commonly fed at 208 V (from 208Y/120 V panels) or, increasingly, at 415Y/240 V, which lets a single three-phase circuit deliver more power. In Canada, 600 V distribution feeds transformers near the white space that produce these rack voltages.

Hyperscale designs often replace individual server PSUs with a shared power shelf in each rack that produces a DC bus at about 48 V to 54 V, distributed along a busbar to all servers. This reduces conversion stages and copper and suits high-density AI racks.

Refresh cycles and flexibility

Servers are typically replaced every 3 to 6 years, network equipment every 5 to 7, while the building, electrical and mechanical systems last 15 to 30 years. Each generation of IT equipment tends to draw more power per rack, and the arrival of AI accelerators has accelerated this trend sharply. Facility designers therefore build in flexibility: spare capacity in electrical rooms, space and connections for adding liquid cooling, structural floors that can take heavier racks, and modular blocks that can be upgraded independently. Retrofitting an existing air-cooled hall for high-density liquid-cooled racks is possible but costly, and many older facilities cannot support the newest AI systems at all.

Equipment arrives in large volumes: a single AI cluster may require thousands of servers delivered, unpacked, installed, cabled and tested in weeks. Loading docks, freight elevators, staging rooms where packaging is removed (to keep combustible cardboard out of the data hall), and corridors wide and strong enough for loaded racks are part of the design.