LV Distribution: Busway, PDUs, RPPs & Rack Power

Everything upstream of this chapter — the utility tie, the substation, the medium-voltage ring, the transformers, the UPS — exists to deliver clean, redundant low-voltage power to a row of racks. This chapter is about the last leg: how that LV power gets from the floor-level distribution into the rack, into the shelf, onto the busbar, and finally onto the accelerator tray. It is the least glamorous part of the power chain and, in the 2026 density regime, one of the most binding. A GB200 NVL72 rack draws roughly 132 kW; the published roadmap puts Rubin Ultra Kyber-class racks near 600 kW and Google, Meta, and Microsoft are openly standardizing toward 1 MW racks. At those numbers the humble copper bar stops being plumbing and becomes a thermal and economic constraint in its own right.

The forks in LV distribution are unusually unforgiving because so much of this layer is poured into the slab or bolted to the ceiling at fit-out time. Choose a 400A busway and a 208V PDU standard, and a hall that was perfectly sized for 40 kW air-cooled racks in 2023 cannot feed a 132 kW liquid rack in 2026 — not because the cooling fails first, but because the copper is undersized and the breakers trip. This is the I²R wall: every doubling of current at fixed voltage quadruples resistive loss, and that loss shows up as heat, voltage droop, and copper mass you pay for by the tonne.

The first fork sits above the rack. Power reaches a row of racks one of two ways. Overhead busway (busbar trunking, "track") runs a continuous enclosed copper or aluminum bus along the top of the row, and each rack draws from a tap-off box that clamps onto the track wherever you need it. Cable-and-PDU runs discrete feeders — conduit and cable, or underfloor whips — from a floor-standing power distribution unit (a floor PDU or remote power panel) to each rack's branch circuit. Both deliver the same electrons; they differ entirely in how the density ramp is serviced.

Busway's decisive advantage is reconfigurability under load. In an open-channel design, a tap-off can be added, moved, or upsized anywhere along the track without de-energizing the bus — you swap a 60A tap-off for a 100A tap-off as the rack behind it grows from 30 kW to 50 kW, and the row never goes dark. That is precisely the property a density ramp wants: the irreversible substrate (the trunking ampacity) is poured once, and the reversible layer (the tap-offs and whips) tracks each GPU generation. Cable-and-PDU, by contrast, fixes the branch-circuit count and conductor size at install. Growing a rack past its provisioned feeder means a new home run from the panel — a conduit pull through a live hall, a panel-schedule change, an electrician with a hot-work permit. The flexibility you didn't buy at fit-out is the flexibility you pay triple for mid-life.

The counter-case is real. Busway carries a higher ampacity rating in a single fault path, so the arc-flash energy at a tap-off and the consequences of a bus fault are larger; the trunking is a long-lead, vendor-specific item with limited cross-compatibility; and for a hall that will never exceed a known, modest density, discrete feeders are cheaper and simpler. The honest reading: busway is the density-ramp instrument, cable-and-PDU is the fixed-density instrument. If there is any chance the hall hosts a generational density step-up — and in 2026 there almost always is — the busway's option value dominates.

Below the tap-off sits the rack PDU — the vertical power strip the IT gear plugs into. The decision that dominates here is voltage class, and it is a pure I²R argument. Legacy halls distribute 208V three-phase to the rack; modern high-density halls distribute 415V/400V three-phase (line-to-line), delivering 240V/230V line-to-neutral to each outlet. At fixed power, raising the rack-PDU voltage from 208V to 415V halves the current, which quarters resistive loss and lets the same conductor and connector carry roughly twice the load. A 60A 208V three-phase PDU tops out around 17–20 kW usable; a 415V three-phase PDU at the same amperage clears the high-30s. That is why 415V (and its 400V European twin) is the de facto rack standard for liquid-density AI rows — busway trunking supporting 415V/480V now represents the majority of new global deployments. Sticking with 208V in a high-density hall is the quiet way to strand half your rack capacity in copper.

The connector layer follows the voltage. AI racks lean on IEC 60309 pin-and-sleeve connectors (the blue/red industrial "ceeform" plugs) for the feed and high-amperage IEC 60320 C19/C21 or the newer C39/C41 outlets for the rack whips, because the legacy C13/C14 ecosystem cannot carry the per-server current modern accelerator trays demand. The rack whip — the flexible cordset from tap-off to PDU — is the field-serviceable reversible element; its connector and amperage are matched to the PDU class, and getting that match wrong is a re-order, not a re-wire.

A/B dual feeds are the rack-level expression of facility redundancy. Two independent PDUs, fed from two independent busways or panels on different UPS branches, let dual-corded equipment ride through the loss of one path. The discipline that breaks operators is the 50% rule: each feed must be loadable to no more than half its capacity, or the survivor cannot pick up the failed path. Single-corded equipment (increasingly common in cost-optimized inference and in some OCP designs that push redundancy down to the shelf) breaks this model and forces a different reliability story — covered in Chapter 4.5 and reframed as goodput-vs-availability in Chapter 12.2.

Hyperscale and the OCP ecosystem take a different cut at the same problem. Instead of distributing AC to every server and converting per-PSU, the Open Rack V3 (ORV3) design centralizes conversion in a power shelf at the top (or middle) of the rack: AC in, a bank of hot-swappable rectifier modules (commonly 12 × ~3 kW, giving ~33 kW per shelf, ~26.5 kW at N+1), and a vertical 48V DC busbar running the height of the rack. Compute and switch trays blind-mate directly onto that bar — no per-server power cords, no rack whips, no C19 forest. NVIDIA's GB200 NVL72 follows the same philosophy with a rack busbar rated to roughly 1,400A feeding the 72-GPU tray stack.

The 48V choice is, again, an I²R decision — just relocated. Forty-eight volts is high enough to cut busbar current ~4× versus the legacy 12V server-distribution bus (quartering loss in the bar), yet low enough to stay under the 60V SELV touch-safety threshold so technicians can service trays without DC-arc and shock procedures. It is the sweet spot the OCP ecosystem converged on, and it is the launch pad for everything in Chapter 4.7: once you are already running a DC bus inside the rack, the natural next moves are to raise it (±400V Mt. Diablo / Diablo 400, 800V NVIDIA reference) and to disaggregate the conversion into a sidecar power rack beside the compute, which is exactly what >600 kW Kyber-class racks require because the power shelves no longer fit in the compute rack at all.

The fork for an operator: AC rack PDUs are the commodity, multi-vendor, retrofit-friendly path that any colo can deliver; the OCP 48V shelf path raises e2e efficiency (fewer conversion stages, the rectifier bank runs near its efficiency sweet spot) and slashes the cabling labor and copper of a per-server cord plant, at the cost of committing to the OCP rack form factor and a narrower supply base. Below ~50 kW the AC-PDU path is rarely worth abandoning; above ~100 kW the shelf-and-busbar architecture's efficiency and density advantages compound, and by the time you reach the megawatt rack it is effectively mandatory.

Strip the LV layer to physics and one equation governs it: resistive loss is I²R, and that loss is dissipated as heat inside the very copper that is supposed to deliver power. Hold voltage constant and double the rack power, and current doubles — but loss quadruples. To hold loss and voltage droop flat you must either raise the voltage (the entire rationale for 415V PDUs, 48V busbars, and the ±400/800V roadmap in Chapter 4.7) or add copper cross-section. At 600 kW–1 MW per rack, the second option hits a wall the industry now calls the copper-mass problem: the busbar needed to carry the current at a tolerable loss is so massive and so heavy that it strains the rack structure, costs a fortune by the tonne, and — the killer — cannot reject its own self-generated heat by natural convection. An NVL72's 5,184 in-rack NVLink copper cables are the same lesson at the signal layer; the power layer is hitting it at the bar.

This is the decision that converts a passive component into an engineered one. Past roughly 140 kW, and unavoidably at the megawatt rack, the in-rack power bus must be actively cooled — the liquid-cooled busbar. The same warm-water loop that cools the cold plates is tapped to carry heat out of the busbar itself, so the bar can run at a higher current density without exceeding its thermal limit. NVIDIA's Vera Rubin reference and the OCP megawatt-rack work both make the liquid-cooled busbar a first-class element. It is the cleanest example in the power chain of thermal-electrical co-design: you can no longer size the electrical path and the cooling path independently, because the electrical path is a heat source the cooling path must service.

Two cross-cutting practices close the chapter. First, phase balancing is the cheapest capacity you will ever recover. Three-phase racks draw on three poles, and IT load rarely distributes itself evenly across them; an unbalanced rack trips the most-loaded phase while the other two sit half-idle. Intelligent PDUs that report per-phase current let you balance the plug-up and reclaim the headroom that imbalance was stranding — the same telemetry that lets you load to 90% (above) is what lets you balance to it. Second, grounding and bonding of the busway and rack-PDU chassis is not optional housekeeping: at these fault currents the bonding path determines whether a ground fault clears safely or arcs, and the whole earthing, SPD, and EMC treatment lives in Chapter 4.11. The LV layer is where the power chain finally touches the chip, and it is where small, poured-in-place decisions about copper and voltage decide how much of your expensive upstream capacity ever reaches the GPUs.