LV Distribution: Busway, PDUs, RPPs & Rack Power

Everything upstream of this chapter (the utility tie, the substation, the medium-voltage ring, the transformers and the UPS) exists to deliver clean, redundant low-voltage power to a row of racks. What remains is the last leg: getting that power from floor-level distribution into the rack, the shelf, onto the busbar and finally to the accelerator tray. It is the least glamorous part of the power chain and, at 2026 densities, one of the tightest. GB300 NVL72 shipped in 2025 and is deploying through 2026. Lenovo rates its rack at 135 kW TDP with peaks to 155 kW, while NVIDIA's reference architecture plans for up to 142 kW per rack, so size to your OEM's profile and its transient peaks. Vera Rubin has shipped since June 2026: Pegatron's VR200 NVL72 datasheet rates operating profiles of 188 kW Max Q / 228 kW Max P, and NVIDIA's DSX reference design sets a 330 kW facility design basis. Rubin Ultra / Kyber remains a ~600 kW planning point (NVIDIA targets H2 2027; SemiAnalysis reported 2028 in July 2026). Google, Meta and Microsoft are standardizing toward 1 MW racks; OCP published their Mt. Diablo 1.0.0 power-rack specification in October 2026. At those numbers the copper bar stops being plumbing and becomes a thermal and economic constraint in its own right.

The forks in LV distribution are unforgiving because so much of this layer is poured into the slab or bolted to the ceiling at fit-out. Choose a 400A busway and a 208V PDU standard, and a hall that was well sized for 40 kW air-cooled racks in 2023 cannot feed a 132 kW liquid rack in 2026. Cooling is not what fails first; the copper is undersized. √3 × 208 V × 400 A is ~144 kVA theoretical, and the standard 125%-continuous-load rule cuts that to ~115 kVA at 320 A, or about 114 kW at a 0.99 power factor. That is below the 132 kW input of an HPE GB200 NVL72 rack. Only a 100%-continuous-rated listed assembly (~143 kW at that power factor, before project derating) clears it, with little margin. Behind this sits 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 an 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, in conduit and cable or as 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 in how they serve the density ramp.

Busway's main advantage is reconfigurability across a density ramp, not routine energized modification. The trunk offers multiple listed tap positions, so a planned outage can relocate or resize branches without pulling a new home-run feeder. The default sequence is to de-load downstream equipment, set up the project's electrical safe-work condition and lockout/tagout boundary, install or replace only the listed tap-off under the manufacturer's instructions, check engagement and torque, then re-energize and verify. Some assemblies allow insertion or removal while the bus stays energized, but that is an engineered exception, not normal work. It needs the exact listed assembly and procedure, downstream load isolation, an approved energized-work justification, an incident-energy assessment, qualified personnel, task-selected PPE and insulated tools, controlled boundaries, stored-energy control and an approved method of procedure. Cable-and-PDU fixes the branch-circuit count and conductor size at installation. Growing a rack past its provisioned feeder means a new home run, planned circuit and panel isolation, a panel-schedule change and re-termination. Busway keeps future branch placement open, but avoiding an outage is never what decides the safe way to do the work.

The counter-case is real. Busway can concentrate high available fault current in one trunk, so calculate tap-off incident energy and bus-fault consequences from the project's source, protection, clearing time, task and working distance. The trunking is long-lead and vendor-specific, with limited cross-compatibility. And for a hall that will never exceed a known, modest density, discrete feeders are cheaper and simpler. The deciding test is economic: busway's installed premium pays off when the feeder work and outage cost it avoids in likely reconfigurations exceed that premium, and when growth stays within the trunk, end-feed and upstream ratings. Growth past those ratings needs upstream changes either way. Busway is the density-ramp instrument; cable-and-PDU is the fixed-density instrument.

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 an I²R argument. Legacy halls distribute 208V three-phase to the rack; modern high-density halls distribute 415V/400V three-phase (line-to-line), which gives each outlet 240V/230V line-to-neutral. At fixed power, raising the rack-PDU voltage from 208V to 415V halves the current and quarters resistive loss, so the same feed conductor and inlet connector carry roughly twice the load. A 60A three-phase PDU limited to 48A continuous delivers about 17 kVA at 208V and about 35 kVA at 415V (33 kVA at 400V); multiply by power factor for kW. So 415V, and its 400V European twin, is the de facto rack standard for liquid-density AI rows. Staying on 208V in a high-density hall strands half your rack capacity in copper.

In North America the live choice for liquid-cooled racks is 415 V against 480 V. A 415Y/240 V rack feed is not native to a 480 V hall: it needs a 480:415 step-down, either a transformer in the floor PDU or a 415 V secondary on the unit substation. A 480Y/277 V feed skips that transformer and carries the same power on about 13.5% less current than 415 V, which cuts I²R loss in the same copper by about a quarter. It suits shelves built to take it. The OCP HPR V2 shelf specification rates its PSUs at 230–277 V line-to-neutral, and NVIDIA feeds each Vera Rubin NVL72 shelf from two 100 A feeds at 415–480 V AC (DGX user guide, October 2026). The cost is a narrower PDU and connector catalog, and a conventional server whose supply is rated 200–240 V cannot take 277 V, so a mixed hall keeps a 415 V or 208 V transformer for that load. Read each shelf's input range on its datasheet before you choose; Chapter 4.1 covers the 400 V versus 480 V hall.

The connector layer follows the voltage. AI racks lean on IEC 60309 pin-and-sleeve connectors (the blue and red industrial "ceeform" plugs) for the feed. For the server cords they use high-current IEC 60320 C19 outlets (C21 is the high-temperature variant) or C39 combination outlets that accept either a C14 or a C20 plug, 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 must match the PDU class, and a mismatch means a re-order, not a re-wire.

A/B dual feeds are the rack-level expression of facility redundancy. Two independent PDUs, fed from independent busways or panels on different UPS branches, let dual-corded equipment ride through the loss of one path. What breaks operators is the failover limit: after one path fails, the survivor must carry the whole load on every phase within its allowable continuous rating. With equal A/B feeds, each normally carries no more than half that rating, which is 40% of nameplate on a circuit limited to 80% continuous. 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 versus 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 in each PSU, the Open Rack V3 (ORV3) design centralizes conversion in a power shelf at the top or middle of the rack. AC goes in, a bank of hot-swappable rectifier modules (6 × ~5.5 kW on first-generation HPR shelves) converts it, and a vertical DC busbar of about 48–50 V runs the height of the rack. Advanced Energy's ORV3 high-power-rack shelf, for example, is rated 50 V nominal and delivers ~33 kW at N+0 and ~27.5 kW at N+1. Meta's open HPR V2 specification (1.0.0, 12 June 2026) puts six 12 kW PSUs in one 1 OU shelf for 72 kW, or 60 kW at N+1, on two 230–277 V AC cords (60 A NEC or 63 A IEC breakers upstream). The design must support at least ten shelves in parallel on the 50 V bus, and the same rack load takes fewer shelves and fewer rack units. Compute and switch trays blind-mate directly onto that bar, with no per-server power cords, no rack whips and no C19 forest.

NVIDIA's GB200 NVL72 follows the same philosophy to feed its 72-GPU tray stack. NVIDIA's October 2024 OCP contribution rates its busbar at 1,400 A, which at 48 V is about 67.2 kW (a guide illustration). Supply and return carry the same current, so their ratings do not add. HPE's 132 kW rack input does not give you the DC-bus current either. HPE specifies a 50 V bus, so work out current segment by segment from the power each segment delivers, counting conversion losses and auxiliary loads separately. Injection points, segment currents and redundancy come from your OEM's drawing (see Chapter 7.13).

Vera Rubin NVL72 keeps the AC-fed shelf pattern at higher ratings (as of 2026-10-13). Pegatron's RA4803-72N3 rack has four 110 kW shelves fed at 415 or 480 V AC, with 188 kW and 228 kW operating profiles. Flex also makes a 110 kW Rubin shelf, which its datasheet rates at six 18.4 kW modules and 52 V at full load. NVIDIA's user guide lists six air-cooled 18.3 kW PSUs per shelf feeding a 54 V bus. NVIDIA's MGX 1.2 specification, published by OCP as revision 1.2.0 (effective October 6, 2026), supports rack deployments up to 250 kW on a 5,000 A liquid-cooled busbar. These are separate product profiles. MGX 1.2 keeps a 48 V bus at 250 kW, so that rating does not by itself force a move to higher voltage. NVIDIA's DSX 330 kW figure is a different thing again: the facility basis for power and cooling.

The 48V choice is an I²R decision, relocated. Forty-eight volts is high enough to cut busbar current ~4× versus the legacy 12V server bus, which cuts equal-resistance conduction loss to one-sixteenth. It also stays under 60 V DC, the steady-state ES1 voltage limit in IEC 62368-1. Staying under that limit does not cap the energy a stiff, high-current bus can deliver, so tray service still follows the listed assembly's procedure and a task-specific arc and stored-energy assessment. The ecosystem converged on this voltage, and it launches everything in Chapter 4.7. Once you run a DC bus inside the rack, the natural next moves are to raise it (±400V OCP Mt. Diablo, 800V NVIDIA reference) and to disaggregate conversion into a sidecar power rack beside the compute. The sidecar is one offered route for >600 kW Kyber-class planning; whether the shelves fit and the heat can be removed depends on the product.

The fork for an operator: AC rack PDUs are the commodity, multi-vendor, retrofit-friendly path that any colo can deliver. The OCP 48–50 V shelf path centralizes AC-to-DC conversion, lets the rectifier bank run nearer its efficiency peak, and cuts the 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. It does not by itself remove a conversion stage, so compare measured end-to-end losses at normal and failover load rather than assuming a saving. At ~50 kW, ~100 kW and megawatt planning points, compare the offered AC-PDU and shelf/busbar interfaces on fixed axes: usable surviving-feed power, measured losses, routing, connector temperatures, service access and delivery. Higher density sharpens the tradeoff, but no universal threshold forces the move.

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 delivers the power. Hold voltage constant and double the rack power, and current doubles but loss quadruples. To hold loss and voltage droop flat you must raise the voltage (the whole rationale for 415V PDUs, 48V busbars and the ±400/800V roadmap in Chapter 4.7) or add copper cross-section. On a low-voltage bus the numbers get large. Delivering 600 kW to 1 MW takes 12,000–20,000 A at 50 V, or 12,500–20,833 A at 48 V, before any distribution losses. Those currents multiply conductor and contact heating. This is the copper-mass problem. An NVL72's 5,184 in-rack NVLink copper cables show the same copper-count pressure at the signal layer, though they do not set the power bar's cooling needs.

This decision turns a passive component into an engineered one. Whether the in-rack bus needs active cooling comes out of the thermal design, not a rack-kW threshold. Voltage, bar cross-section and resistance, contact losses, enclosure, ambient, allowable temperature rise, redundancy state and current decide it. When a particular bus exceeds its passive thermal limit, a liquid-cooled busbar can carry out the heat and support a higher current density. NVIDIA's MGX 1.2 specification puts the liquid-cooled busbar in the baseline design: a 5,000 A bar for racks up to 250 kW (specification ceilings, not a shipping rack). The OCP megawatt-rack work goes further. 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 practices cut across everything above. First, phase balancing can recover capacity stranded on the most-loaded pole. Three-phase racks draw on three poles, and IT load rarely spreads evenly across them, so one phase can reach its rated limit while the other two stay below theirs. Intelligent PDUs that report per-phase current let operators rebalance the plug-up until every phase sits within its continuous rating. Second, grounding and bonding of the busway and rack-PDU chassis is not optional housekeeping: at these fault currents the bonding path decides whether a ground fault clears safely or arcs, and the earthing, SPD and EMC treatment lives in Chapter 4.11. The LV layer is where the power chain finally touches the chip, and small, poured-in-place decisions about copper and voltage decide how much of your expensive upstream capacity ever reaches the GPUs.