From Busway to Rack PDU
The last stretch inside the building is the shortest and the most crowded. Power leaves the UPS and switchgear at low voltage and now has to be split thousands of ways, delivered to every rack, and metered along the way. This is the distribution layer, and its components — power distribution units, busways, remote panels, and rack-level power strips — are where the single site-wide feed finally becomes thousands of individual server connections.
The first device in this layer is the power distribution unit, or PDU, in its large floor-standing form. A facility PDU takes the low-voltage feed from the UPS output and subdivides it into many branch circuits, each protected by its own breaker and often individually metered. In many designs the PDU also contains a final transformer that trims the voltage to exactly what the racks expect. Think of the floor PDU as the panel board of the data hall: one input, many protected and monitored outputs, each heading toward a row or a cluster of racks.
How those outputs travel to the racks is itself a design choice, and the industry has largely moved from conduit to busway. Traditional distribution ran individual cables in conduit from the PDU to each rack — reliable but rigid, because adding or moving a rack meant pulling new cable. Overhead busway replaces that with a continuous rigid or semi-rigid bus bar running above the rows, into which powered tap-off boxes can be clipped anywhere along its length. Need power at a new rack position? Clip in a tap box. Busway turns the distribution network into something modular and reconfigurable, which matters enormously in a facility whose rack layout changes as hardware generations turn over.
At the rack itself sits the rack power distribution unit — the vertical strip of outlets, usually mounted in the back of the rack, into which every server plugs. Do not confuse it with the floor-standing PDU; the rack PDU is the last piece of infrastructure before the server's own power supply. Modern rack PDUs are intelligent: they meter current per outlet, report power draw over the network, and can switch individual outlets on and off remotely. That per-outlet telemetry is what lets operators track power usage down to the individual server, balance load across circuits, and catch a rack drifting toward its breaker limit before it trips.
Redundancy persists all the way to this last layer, and it is visible in the cabling. In a resilient design, every rack receives two independent feeds — an A side and a B side — each traced back through separate rack PDUs, separate busways, separate floor PDUs, separate UPS systems, to the two independent power trains established upstream. Each server has two power supplies, one plugged into each side. Lose the entire A side — a UPS, a busway, a breaker — and every server keeps running on B without a flicker. The two-path topology that began at the building transformer is only meaningful if it is carried, unbroken, all the way to the outlet the server plugs into. A single break in that discipline collapses a 2N site to a single point of failure at the rack.
Density is reshaping this layer fast. A rack of general-purpose servers might draw a handful of kilowatts. A rack packed with AI accelerators can draw many tens of kilowatts, and the trend is sharply upward. That density forces higher-current rack PDUs, heavier busway ratings, and careful attention to how many high-draw racks share a branch circuit. The distribution layer that comfortably fed a hall of traditional servers has to be redesigned — thicker bus bars, higher-amperage tap boxes, more circuits per rack — to feed a hall of accelerators pulling several times the power in the same floor space. The physical distribution network is now a binding constraint on how dense a facility can go.
By the end of this layer the transformation is complete in one sense: the single medium-voltage feed that entered the campus has been stepped down, buffered, split, routed, and metered until it arrives as a pair of ordinary power cords at the back of a server. But the descent is not over. Inside the server, that low-voltage alternating current still has to become the low, precise direct-current voltages the silicon actually runs on — the final, and in efficiency terms one of the most consequential, transformations in the entire stack.