The AI compute revolution has fundamentally broken the assumptions behind traditional data center power distribution. Facilities designed around 5–8 kW per rack — the standard that shaped a generation of RPP-based infrastructure — now face racks drawing 30, 60, and in the case of NVIDIA's GB200 NVL72, over 120 kW each. The electrical distribution systems that served enterprise IT workloads for two decades simply cannot scale to meet this demand efficiently.
This guide compares two primary distribution strategies — track busway systems (specifically Starline's plug-in track busway) and remote power panels (RPPs) — from an electrical engineering standpoint, with direct application to high-density AI and GPU compute environments.
The AI Density Inflection Point
The chart below illustrates the trajectory of rack-level power density from traditional enterprise IT through projected next-generation AI workloads. The inflection after 2022 represents the GPU cluster era — a density regime that existing RPP-based infrastructure was never designed to support.
Fig. 1 — Rack power density progression. The dashed line marks the practical ceiling above which RPP-based distribution becomes architecturally constrained. Busway scales beyond this boundary without redesign.
At densities above approximately 30–40 kW per rack — which encompasses any meaningful H100, H200, or Blackwell GPU cluster — the physical and electrical limitations of RPP-based distribution become acute. Understanding why requires examining both systems at the circuit level.
System Architecture: Electrical One-Line Comparison
The following one-line diagrams illustrate the fundamental difference in power distribution topology between a busway-fed system and a conventional RPP-fed system serving the same row of high-density racks. Both diagrams assume a 480V 3-phase source and a target density of 30–60 kW per rack.
Fig. 2A — Redundant A+B busway feeds from independent sources. Each tap-off serves one rack with locally-fused branch circuit protection. Tap-offs relocatable without power outage.
Fig. 2B — Conventional RPP topology. Panel capacity is fixed at manufacture. Load changes require conduit/wire modifications. Single-source architecture typical; redundancy requires full duplication of panel infrastructure.
Aisle Elevation: Installation & Clearance Comparison
The elevation view below illustrates how the two systems occupy physical space within a data center aisle. For AI deployments with 30–120 kW racks, the spatial and thermal implications of distribution architecture are significant.
Fig. 3 — Hot-aisle elevation comparison. Busway overhead routing preserves white space and allows tap-off relocation without outage. RPPs consume valuable aisle-end real estate and require fixed, static wiring.
Engineering Comparison Matrix
| Criteria | Starline Track Busway | Remote Power Panel (RPP) |
|---|---|---|
| Electrical Capacity & Performance | ||
| Max backbone current | Up to 6000A continuous (NEC 368) No practical limit on row density |
Typically 225–600A MCB Panel-limited; parallel panels add complexity |
| Per-rack capacity | 30–400A tap-off units available Supports 120 kW+ per rack with proper transformer |
Typically 20–100A per circuit Multiple parallel circuits required above ~25 kW |
| Voltage drop | Low impedance bus bar; superior voltage regulation over long runs | Conduit/wire impedance increases with density; parallel runs required |
| Short circuit rating | Rated to withstand and interrupt high fault currents; tap-offs provide zone protection | Panel-rated; coordination depends on upstream device sizing |
| Arc flash incident energy | Tap-off integral fusing limits AFIE at point of utilization; supports zone-selective coordination | AFIE determined by upstream OCPD; can be optimized with proper relay coordination |
| Flexibility & Scalability | ||
| Load reassignment | Tap-off units relocate to any point along busway — no outage required (plug-in design) | Requires conduit/wire rework; outage required for circuit changes |
| Capacity expansion | Add tap-off units to existing bus; bus supports incremental load growth without infrastructure change | Panel is full at manufacture; additional panels require new feeder, breaker, and conduit infrastructure |
| Rack density changes | Plug different ampacity tap-off; no electrical work required beyond tap-off swap | New branch circuit, breaker, conduit, and wire required; timeline measured in weeks |
| Row reconfiguration | Busway sections add or remove; entire row lengths adjustable | Fixed panel location; full redesign for major reconfigurations |
| Physical Installation | ||
| White space footprint | Zero white space consumed — fully overhead; no aisle-end panels | 55–110 sq ft per RPP location; aisle-end real estate consumed |
| Installation speed | Modular sections bolt together; typical row installed in 1–2 days | Panel installation is fast; conduit/wire infrastructure takes weeks per row |
| Future-proofing | Busway rated capacity independent of rack count; upgrade tap-offs only | Panel sizing must anticipate future load; over-sizing wastes capital; under-sizing forces replacement |
| Redundancy & Reliability | ||
| A+B feed redundancy | Dual-bus architecture straightforward; each rack served by two independent tap-offs from separate buses | Requires two independent panels per zone; doubles panel count and wiring infrastructure |
| Fault isolation | Integral tap-off fusing isolates rack-level faults; bus remains energized | Branch breaker at panel provides isolation; fault affects multiple circuits depending on panel layout |
| Economics | ||
| Initial material cost | Higher busway material cost than conduit/wire | Lower material cost for simple, static layouts |
| Total installed cost (high density) | Lower at density > ~20 kW/rack due to reduced conduit, wire, labor, and panel count | Higher at density > ~20 kW/rack when accounting for parallel conductors and added panels |
| Moves / adds / changes (MAC) | Near-zero electrical MAC cost; tap-off swap is field operation | High MAC cost; each change requires licensed electrician, conduit work, and potential outage |
| Best fit density | ≥ 20 kW/rack; mandatory above 40 kW/rack | ≤ 15 kW/rack in static, predictable environments |
Critical Design Considerations for AI Deployments
1. The Parallel Conductor Problem
At 480V 3-phase with a power factor of 0.95, a 30 kW rack draws approximately 38A per phase. Manageable. But a 120 kW rack (GB200 NVL72) draws approximately 152A per phase — requiring 2/0 AWG minimum, and at a 42-rack row that means 6,384A of total 3-phase current. At that scale, conduit-based branch circuits become physically impractical: you're running parallel 350 kcmil conductors in multiple conduits per rack, each requiring individual pull, termination, and label management.
Busway eliminates this entirely. A single 2500A busway backbone serves the entire row with a single feeder from the switchgear. Each rack gets one tap-off unit. The math is simple, the installation is fast, and the architecture scales cleanly to any density the transformers and switchgear can support.
Rule of thumb: If your design calls for more than two parallel conductors to any single rack, you have exceeded the practical envelope of RPP-based distribution. Busway is the correct architecture.
2. Tap-Off Unit Selection for High-Density Racks
Starline tap-off units are available from 15A to 400A in both fused and circuit-breaker configurations. For high-density AI racks, the engineering selection criteria are:
- Ampacity: Size tap-off at 125% of continuous load per NEC 210.20(A). A 60 kW rack at 480V draws ~76A per phase; specify 100A tap-off minimum.
- Protection type: Fused tap-offs offer lower let-through energy and better selectivity. Breaker-equipped tap-offs allow field reset without fuse replacement.
- Voltage: Tap-off at 480V to rack-level PDU with internal step-down, or direct 480V to server PSUs (increasing common in hyperscale). Verify GPU server PSU voltage range before specifying.
- Monitoring: Specify metered tap-offs where per-rack power monitoring is required for capacity planning or billing — Starline offers integrated metering with DCIM-compatible output.
3. Transformer Sizing for Busway-Fed AI Rows
The busway backbone capacity is limited by the upstream transformer and switchgear, not the busway itself. For a 24-rack AI row at 60 kW per rack (1.44 MW total), assuming 50% diversity factor (common in batch GPU workloads), plan for:
- Transformer: 2 × 1500–2000 kVA for A+B redundant feeds (N+1 capacity each)
- LV Switchgear: 3000–4000A frame for each bus feed
- Busway: 2500A or 4000A rated bus per row (size to 125% of maximum anticipated load, not 50% diversity)
- Coordinate with mechanical: At these densities, rear-door heat exchangers or in-row cooling CDUs likely; PDU piping coordination required
4. NEC and Code Compliance
Busway systems are governed by NEC Article 368 (Busways). Key requirements for AI data center installations:
- Support spacing: Maximum 5 ft for horizontal runs unless manufacturer lists otherwise
- Overcurrent protection: Required at each point of supply and at each tap-off per 368.17
- Reduction in size: Where bus reduces in ampacity, OCPD at the reduction point unless exception conditions met
- Wet/damp locations: Confirm busway listing for the installation environment (most data center busway is listed for dry locations)
- Working space: Maintain NEC 110.26 clearances around tap-off access points
Engineering note: The NEC tap rule (240.21) allows tap conductors without overcurrent protection at the source if length and termination conditions are met. This does not apply to busway tap-offs — each Starline tap-off includes its own OCPD as an integral part of the assembly, simplifying code compliance and reducing arc flash hazard at the utilization point.
5. Liquid Cooling Integration
Direct liquid cooling (DLC) and immersion systems do not eliminate the need for robust electrical distribution — they change the load density and often increase it. A CDU-cooled GB200 node still draws 120+ kW of electrical power; only the thermal output method changes. Busway is more compatible with liquid-cooled deployments because overhead routing avoids conflicts with underfloor CDU piping, coolant distribution units (CDUs) can be positioned without regard to RPP panel locations, and per-rack electrical capacity can scale independently of the cooling infrastructure changes.
Design Decision Framework
Use this framework to select the appropriate distribution architecture for your project. For most AI and high-performance compute facilities, the answer is busway — the question is usually bus rating and tap-off configuration, not whether to use busway at all.
Specify Starline Track Busway When:
- Rack density exceeds 20 kW average or any rack exceeds 40 kW
- Facility will serve GPU clusters, AI inference, or HPC workloads
- Row configuration or tenant mix is expected to change over time
- A+B redundancy is required at the rack level
- Installation timeline is compressed (busway installs faster than conduit at high density)
- Floor space is at a premium (hyperscale, colocation)
- Per-rack power monitoring at the circuit level is required
- Facility lifecycle exceeds 10 years with technology refresh anticipated
RPP May Be Appropriate When:
- Average density is below 10–12 kW/rack with no growth planned
- Workloads are static and predictable (storage, networking infrastructure)
- Small footprint (<20 racks) with simple layout
- Budget is highly constrained at initial build
- Facility life is short (<5 years) with planned decommission
- Conduit infrastructure already exists from prior buildout
Hybrid approach: Many modern facilities use busway for primary compute rows (where density and change are anticipated) and RPPs for ancillary infrastructure — network equipment rooms, storage rows, and support systems — where density is low and static. This optimizes first cost while ensuring compute infrastructure is future-ready.
Starline Track Busway — Technical Specifications Reference
| Parameter | Starline Track Busway Specification | Notes |
|---|---|---|
| Voltage ratings | 480V, 600V (UL listed); 415V (IEC) | Verify for international projects |
| Current ratings | 100A – 6000A | Select per row load + 25% growth margin |
| Tap-off ampacity range | 15A – 400A | Match to rack PDU or server PSU input |
| Short circuit rating | Up to 100 kAIC | Coordinate with upstream OCPD |
| Bus bar material | Copper (standard); aluminum available | Copper preferred for high-density/high-cycle applications |
| IP rating | IP40 standard; IP54 available | IP54 for raised floor or airflow-intensive environments |
| Listing | UL 857 (Busways) | NEC Article 368 compliant |
| Tap-off access | Continuous plug-in; no tools required | Can add/remove tap-offs under load where UL listed for energized operation |
| Section lengths | Modular; custom lengths available | Order-specific to row length; allow 6–8 week lead time |
| Monitoring options | Per-tap-off metering; DCIM/BMS integration | MODBUS, BACnet, SNMP output available |
| Finish/housing | Steel enclosure, powder coat | Color options available for tenant identification |
Ready to Specify Busway for Your AI Data Center?
Arkhon Power Group is an authorized Starline representative with deep experience in high-density data center power distribution. We provide system design support, submittals, equipment selection, and engineering coordination for projects from design through commissioning.
Contact Our Engineering Team