How power actually moves through an AI data center after it enters the site — medium-voltage switchgear, unit substations, UPS systems, PDUs, and busway — and why the distribution architecture is built around redundant, electrically independent A and B paths.
From the Utility Meter to the Rack: The Full Distribution Chain
Power distribution in an AI data center is a series of deliberate step-downs, each one trading a lower voltage for higher current, moving from the utility interconnection all the way to the individual server power supply. Understanding this chain matters because a problem or a design decision at any single stage — undersized medium-voltage switchgear, an inadequately-rated busway run, a UPS sized for yesterday's rack density — can bottleneck the entire facility's usable capacity, regardless of how much power the utility connection itself can deliver. The full path, at a glance: utility service → main substation (steps down to medium voltage) → medium-voltage feeders to unit substations near each hall → 480V distribution → UPS → PDUs or busway → rack. See AI Data Center Electrical Design — Complete Guide for the full picture including generation and backup systems; this article focuses specifically on the distribution path itself.
Medium-Voltage Distribution: Moving Power Across the Site
Once the utility's high transmission voltage is stepped down at the main substation, it's carried across the site at medium voltage — commonly 13.8 kV or 34.5 kV in North America — via underground duct banks or overhead lines to a series of unit substations, each one serving a specific data hall or pod. Medium voltage is used for this leg specifically because it keeps current (and therefore conductor size and resistive loss) manageable across the long distances a hyperscale campus can span; running the full site at 480V directly from the main substation would require impractically large conductors and would waste significant power as heat.
The 480V Step-Down and Why It's the Workhorse Voltage
Each unit substation transforms medium voltage down to 480V three-phase, which is the voltage level nearly all UPS systems, PDUs, and downstream distribution equipment in a North American data center are built around. 480V strikes a practical balance: low enough to work with standard, widely available switchgear and cabling, high enough that current levels stay manageable even at data-hall-scale power draws. This is the voltage the UPS system (see Data Center UPS Systems Explained) receives and outputs before power continues downstream.
PDUs and Busway: The Last Stretch to the Rack
Inside the data hall, 480V power from the UPS is stepped down again — typically to 208V three-phase — by power distribution units (PDUs), which also provide circuit breakers, metering, and monitoring for the branch circuits feeding individual racks. Two physical approaches dominate: traditional whip cabling running from PDU to rack, or overhead busway, a rigid or flexible conductor track that can be tapped at any point to add a new rack connection. Busway has become the preferred choice for many high-density AI deployments — see Data Center Busway Explained — specifically because GPU rack layouts and power requirements tend to change often, and busway lets an operator add or relocate rack power drops without pulling new circuits from the PDU.
Why Everything Runs in Redundant A and B Paths
From the unit substation all the way down to the rack, critical AI data center power distribution is built as two electrically independent paths — conventionally labeled A and B — each capable of carrying the full critical load on its own. A server with dual power supplies draws from both the A and B path simultaneously in normal operation; if either path fails or goes down for maintenance, the other continues supplying full power with zero interruption. This dual-path architecture is what the terms N+1 and 2N actually describe (see N+1 and 2N Redundancy Explained) — it's not a single concept but a distribution philosophy applied consistently at every stage of the chain described in this article.
Sizing for a Moving Target
One of the harder realities of AI data center power distribution is that the load it's built to serve keeps changing — successive GPU generations draw meaningfully more power per rack than their predecessors, often faster than a typical data center's planned refresh cycle. This is a major reason operators increasingly favor distribution architectures (like busway, and oversized medium-voltage switchgear headroom) that can absorb higher future densities without a full electrical retrofit, rather than sizing distribution tightly to today's known GPU power draw.