Load estimation and capacity planning

Every decision in the electrical design of a data centre, from the size of the utility service in Lesson 1 to the rating of a single rack circuit, starts from an estimate of load. Estimate too low and the facility runs out of power before it is full; estimate too high and millions of dollars of transformers, switchgear and generators sit idle for years, and the utility reserves capacity that could have served someone else. This lesson explains how data centre loads are built up from the IT equipment to the utility meter, why nameplate ratings mislead, how diversity, redundancy and growth enter the calculation, and how the arrival of AI training clusters has changed both the size and the behaviour of the load.

From the server to the meter

The load at the utility meter is built up in layers:

• IT load: servers, storage and network equipment. This is the load the facility exists to serve, and it is usually the "critical load" that is protected by UPS.
• Mechanical load: chillers, pumps, cooling units, fans and cooling towers. It depends on the IT load, the cooling design and the weather.
• Electrical losses: UPS, transformer and distribution losses, typically 5% to 10% of the IT load in older designs and less in modern ones.
• House load: lighting, offices, security, building management, fire systems, heating of generator engines and battery rooms.

The ratio of total facility power to IT power is the power usage effectiveness (PUE, DCE-200). For electrical design, what matters is the peak PUE on the hottest design day, not the annual average, because the electrical system must carry the worst case.

Nameplate, measured and design load

Every server carries a nameplate power rating, usually the rating of its power supply. Real equipment almost never draws that much: a server with a 1,100 W power supply may draw 300 W idle and 650 W at full load. Summing nameplate ratings therefore overstates the load, sometimes by a factor of two, and leads to oversized and underused infrastructure.

Better estimates come from:

• Measured power of the actual equipment running representative workloads, from vendor power calculators or tests.
• Measurements from existing facilities running the same kind of workload.
• The power budget per rack agreed with the IT team or the tenant, which then becomes a contractual limit enforced by the rack circuits.

Diversity and demand factors

Not every load peaks at the same moment. The diversity factor is the sum of the individual peak loads divided by the peak of the combined load; its inverse, the coincidence factor, is the fraction of the summed peaks that occurs at once. Electrical codes use demand factors for many building loads, allowing feeders to be sized below the connected load.

In traditional enterprise and colocation data centres, diversity is significant: different tenants and applications peak at different times, and few racks reach their allocated power. Measured coincident peaks of 60% to 80% of the allocated rack power are common. Designers can use this to place more racks behind a given amount of upstream capacity, a practice called oversubscription, but they must then monitor the load and have a plan for what happens if it grows (Lesson 10).

AI training clusters are different. Thousands of accelerators work on one job in lockstep, so they reach peak power together: the coincidence factor approaches 1, and the diversity that once allowed oversubscription disappears.

Redundancy and usable capacity

Redundancy changes the relationship between installed and usable capacity. In a 2N design each path must carry the whole load, so each is loaded to at most 50% in normal operation; in a distributed redundant design with \(N\) blocks sized so that any \(N-1\) can carry the load, each block is loaded to at most \((N-1)/N\); in N+1 designs the usable capacity is that of \(N\) units. These architectures are covered in DCE-340; for load planning the key quantity is usable critical capacity, the IT load the facility can support with its redundancy intact.

Growth and phasing

Data centres are rarely filled on day one. A campus may be built in phases of 10 MW to 100 MW over several years, and each hall fills over months as customers move in. Electrical infrastructure is therefore designed for the full build-out, with the long-lead and hard-to-expand parts (the utility service, substation space, main duct banks, generator yard) sized for the ultimate load, while modular parts (UPS modules, PDUs, generators) are added as the load arrives. The plan must keep redundancy intact at every intermediate stage, not just at the end.

How AI changes the load

AI clusters change data centre load in three ways that matter to electrical engineers:

• Density: racks of 100 kW or more concentrate load into fewer, heavier circuits, busways and cooling units (Lesson 5).
• Coincidence: synchronized accelerators remove diversity, so design load approaches the sum of rack peaks.
• Dynamics: during training, accelerators alternate between computation and communication phases many times a second to every few seconds, and a whole cluster may drop from full power to a fraction of it at once when a job stops or checkpoints. Swings of tens of megawatts within a second are possible on the largest campuses.

Rapid swings stress generators and UPS systems, which must respond to large load steps, and can disturb the grid: utilities and grid operators such as NERC have raised concerns about large loads that change quickly or disconnect suddenly. In July 2024, a voltage disturbance in northern Virginia caused about 1,500 MW of data centre load to transfer to backup power at once, an event NERC studied as a warning of what large, sensitive loads can do to grid stability. Mitigations include energy storage that smooths the load seen by the grid, software that smooths power draw at the accelerators, and coordination with the utility on ride-through settings (Lesson 8).

Documenting the load

The results of load estimation are recorded in a load schedule (a table of every load with its rating, design demand and the panel or bus that feeds it) and summarized in the design basis, which states the design IT load per hall and per rack, redundancy, design-day conditions, PUE, growth plan and margins. These documents drive the sizing of everything that follows and are updated as the design develops (Lesson 12).

The Canadian Electrical Code permits demand factors for some loads and requires feeders and services to be sized for the calculated load; utility connection requests to provincial utilities such as Hydro-Québec and Hydro One, or through the Alberta Electric System Operator (AESO) in Alberta, are made on the basis of the design load and its ramp.