How To Choose Rack PDU For GPU Servers

Quick Answer

Choosing a rack PDU for GPU servers comes down to four numbers: per-rack kW, input voltage and phase, the C13/C19 outlet mix, and the metering level you actually need. Start from the load. An air-cooled 8-GPU node draws 6–12 kW at the wall under training load, so a four-node rack lands at 25–45 kW — far beyond what single-phase power can deliver. That single fact drives most of the GPU server rack power distribution decisions that follow.

In practice that means: three-phase input, two PDUs per rack on independent A/B feeds, outlets dominated by C19, and at least input-level metering so you can balance phases during install. Single phase vs three phase rack PDU is rarely a real debate above 8 kW per rack — three phase wins on capacity, conductor size, and phase balancing.

When This Problem Appears

Teams hit this decision when they choose PDU for AI rack deployments for the first time: a retrofit of CPU cabinets for GPU nodes, a first H100/H200-class cluster in colo space, or an edge build where the power contract was signed before the IT load was known. The classic discovery moment is realizing the existing 32 A single-phase PDUs top out around 7 kW — less than a single training node draws — or that the colo's "per-rack power" figure is a contractual cap, not the breaker rating.

It also appears at quoting time. A rack PDU for GPU servers is not a commodity line like a patch cable; input plug, outlet map, breaker layout, and metering level all have to be specified, and a vague RFQ produces quotes that cannot be compared against each other.

Key Factors to Check

The datasheet tells you part of the story; the rest you verify with evidence before the unit ships. When sourcing PDUs from China, supplier matching narrows the field to factories that build to the spec you need — and pre-shipment checking confirms the unit on the pallet matches the unit on the quote.

If a supplier cannot produce these, that is the answer. Every PDU order we place can include a checked shipment — see pre-shipment checking for what the inspection covers.

Recommended Accessories or Inputs

A PDU never ships alone. The lines that belong on the same RFQ:

• The second PDU. GPU nodes have redundant PSUs for a reason — spec two PDUs per rack on independent A/B feeds, sized so either side carries the full load.
• C19-to-C20 jumpers for every GPU node PSU, and C13-to-C14 jumpers for switches and management gear. Counts come straight from the outlet map; the C13 C19 outlets for GPU servers question is covered in detail in our C13 vs C19 power cord guide.
• Input cord or hardwire kit, with length measured from the busway or floor drop to the PDU inlet.
• Blanking panels and brush strips to keep the hot rear exhaust away from the cold aisle — cheap insurance for the PDU's own temperature rating.
• Vertical cable managers so 30+ jumpers per rack do not block the PDU breakers.

The AI Rack Power Kit bundles these as one quote, matched to your rack count. On metering, this is the trade-off table we walk customers through:

For most GPU racks, metered vs switched PDU GPU rack debates resolve to monitored-by-outlet: you want to see per-outlet draw, but remotely power-cycling a multi-PSU training node one outlet at a time is rarely useful.

Already have your rack numbers?

Send rack count, per-rack kW, voltage, and target country as-is. We quote the PDU and cord accessory lines within 24–48 hours — no GPUs, servers, or UPS units, just the rack power hardware.

Submit RFQ

Common Mistakes

• Sizing from GPU TDP alone. Eight GPUs at 700 W is 5.6 kW, but the node draws 9–11 kW at the wall once CPUs, fans, NICs, and PSU losses are counted. Size from measured or vendor wall-draw figures.
• Unclear rack power. Quoting a PDU before the colo confirms the per-rack kW cap and receptacle type guarantees a re-quote. Get the contract numbers first.
• Wrong plug type. An IEC 60309 532P6 plug does not fit a 563P6 socket. Input plug standard belongs on line one of the RFQ, tied to the target country.
• Missing cable length on the input side — busway drops and underfloor runs vary per rack, and a short input cord stalls the whole install.
• Incomplete BoM. The most common omission is the entire B side: second PDU, second set of jumpers. The second most common is spares.
• Temperature rating ignored. A 45 °C-rated PDU mounted in the exhaust path of a 40 kW rack is operating outside spec from day one. For GPU racks, specify 60 °C-rated units.
• Unverified certificates. UL/CE/CCC marks on the housing are claims, not proof. Verify the certificate holder before shipment.

Scope boundary, stated plainly: we quote PDUs, cords, and rack power accessories. UPS systems, transformers, switchgear, and other high-voltage equipment are outside our lines, and GPU or AI accelerators are not supplied. If a quote you receive elsewhere bundles those, compare the accessory lines separately — that is where spec mismatches hide.

RFQ Checklist or Next Step

Ten lines make a PDU RFQ quotable in one pass:

1. Rack count and PDUs per rack (A/B feeds?)
2. Rack power — kW per rack under load
3. Voltage and phase available at the rack
4. Input plug type or hardwire requirement, per target country
5. Outlet type counts — C19 and C13 separately, per PDU
6. Metering level — basic, metered, monitored, or switched
7. Cable length for input cords and jumpers
8. Certification requirement for the destination market (UL, CE, CCC)
9. Quantities including spares
10. MOQ flexibility and lead time target

Send what you have — even rack count plus node model is enough to start, and quotes on accessory lines come back within 24–48 hours.

Related Kits and Guides

• AI Rack Power Kit — PDUs, cords, and power accessories quoted as one bundle
• C13 vs C19 Power Cord — the cord-side companion to this guide
• GPU Rack Deployment Checklist — the full accessory picture beyond power
• Solutions for GPU server builders — how integrators source rack power at volume
• Pre-Shipment Checking — what we inspect before a PDU leaves the factory