Power Systems: From the Substation to the Server

A data center's power system has one job: the servers must never see the lights flicker. Even when a breaker fails at 3 a.m. during a thunderstorm. Everything in this lesson exists to make that promise, and the cost of keeping it explains why data center electrical rooms look like small utilities.

The power path

1. Utility feeds. Big campuses connect at transmission voltage, often 69 kV to 345 kV, sometimes through a substation built just for them. Two feeds from different substations protect against one line going down.
2. Substation. Transformers step the voltage down to medium voltage (often 12 to 35 kV) for distribution around the campus.
3. Switchgear. Rooms of large breakers route power to each hall and isolate faults. Automatic transfer switches move the load to generators when the utility fails.
4. UPS. The uninterruptible power supply cleans up the power and, the instant the utility fails, carries the load from stored energy. Lithium-ion batteries have largely replaced lead-acid in new builds; some sites use spinning flywheels. Battery time is usually sized for minutes, enough to cover generator start-up and a few retries, not hours.
5. Generators. Diesel generators, often 2 to 3 MW each, start within seconds and carry the site until the utility returns. Fuel on site commonly covers a day or two, with refueling contracts for longer. A 300 MW campus can have more than 100 of them.
6. PDUs and server power supplies. Power is stepped down again and divided rack by rack. Each server usually has two power supplies fed from separate paths.

AI is changing the last steps. Racks drawing 100 to 200 kW need so much current at traditional voltages that cables and bus bars get huge. Nvidia and partners are moving toward 800-volt direct current distribution to cut current and conversion losses. Expect new AI halls to look different inside the electrical room.

Redundancy shorthand

N is the amount of equipment needed to carry the full load. Everything else is described relative to it.

Notation	Meaning	Example with a 4 MW load and 1 MW UPS modules	What it survives
N	Exactly enough	4 modules	Nothing. Lose one module and you drop load.
N+1	One spare	5 modules	One component failure or one unit out for service
N+2	Two spares	6 modules	One failure while another unit is down for maintenance
2N	Two complete independent systems, each able to carry everything	8 modules in two separate systems, A and B	Loss of an entire system, including its wiring path
2N+1	Two systems plus a spare	9 or more modules	Loss of one system plus a component in the other

Uptime Institute Tiers

The Uptime Institute's Tier Classification System is the most cited rating for data center infrastructure. Uptime certifies designs and constructed facilities separately; a developer saying "Tier III design" is not the same as a certified Tier III constructed facility. The Tier pages describe what each level must deliver:

Tier	Name	Redundant components?	Redundant distribution paths?	Maintain without shutdown?	Survive an unplanned failure?
I	Basic capacity	Not required	No	No: maintenance needs a full shutdown	No
II	Redundant capacity components	Yes (generators, UPS, chillers)	No	Components yes, distribution path no	No
III	Concurrently maintainable	Yes	Yes	Yes: any part can be shut down without affecting IT	Not guaranteed
IV	Fault tolerant	Yes, in physically isolated systems	Yes, isolated	Yes	Yes: one event should not take out both systems; continuous cooling required

You will often see availability percentages next to Tiers, such as "Tier III equals 99.982 percent." Uptime's current Tier pages do not set numeric availability targets. Those figures come from older material and marketing. If a developer cites one, ask where it comes from. Real-world availability depends heavily on operations and staff, which Uptime rates separately.

Generators and the neighbors

Diesel generators are the part of the power system that neighbors notice. They are tested regularly, often monthly, which means noise and exhaust. Federal rules for emergency compression-ignition engines (40 CFR 60.4211(f)) allow up to 100 hours a year for maintenance and testing, of which up to 50 may be used for limited non-emergency purposes; true emergencies are not capped. States issue air permits, and those permits list the number of engines, their size and the allowed hours. Some campuses now use natural gas generators or large batteries instead, and some sign deals to run generators during grid emergencies, which local air regulators watch closely.

When the backup system surprises the grid

On July 10, 2024, a lightning arrester failed on a 230 kV line in the Eastern Interconnection in an area with many data centers. The line's automatic reclosing made repeated attempts, producing six faults in 82 seconds. Data center control systems that count voltage dips moved about 1,500 MW of load to backup power, and roughly 1,260 MW stayed off the grid for hours. The North American Electric Reliability Corporation (NERC) reviewed the event because a sudden loss of that much load can push grid frequency and voltage too high. The protection worked for the data centers; it surprised the grid. Utilities are now writing rules for how large loads must ride through disturbances.