TL;DR — 30 kW AI GPU Rack PDU Spec at a Glance

Specifying a PDU for a 30 kW+ AI GPU rack is a different exercise from specifying a PDU for a 5 kW general-purpose server rack, and most of the catalog advice that works for the latter is dangerously wrong for the former. Because a single H100-class GPU server draws 700–1,200 W continuously and an 8-GPU node pulls close to 10 kW, a rack populated with four such nodes lands at 30–40 kW before networking, storage, and headnode overhead. The PDU spec has to handle that load electrically, thermally, and phase-balanced — all without becoming the bottleneck of the AI training run.

This checklist walks through the six decisions that drive a quoteable PDU BOM for a 30 kW+ AI rack. Each decision is something our engineering team walks buyers through on every OEM inquiry. Browse our full high-density rack PDUs catalog as you read, or use the contact page to spec a GPU-rack PDU with us directly.

What 30 kW Actually Means on a Rack Floor

The first spec confusion is that “30 kW” is not a single number — it is a derived load that depends on the supply voltage. Because watts are voltage times current, halving the supply voltage doubles the current a PDU must carry, and current drives everything else: plug rating, wire gauge, breaker size, and thermal derating. A 30 kW rack loaded exclusively from 120 V single-phase would need at least 250 A per phase — well beyond any standard NEMA plug. A 30 kW rack on 208 V three-phase needs only 83 A per phase, which is well within the range of a NEMA L21-30P or IEC 60309 32A plug.

Why 30 kW Is the New Hyperscale Baseline

Five years ago, the largest AI training racks shipped at 10–15 kW. Today, H100 and B200 generation training racks routinely ship at 30 kW, and the next-generation rack designs published by NVIDIA, Meta, and Microsoft all assume 50–100 kW per rack. Because the rack-level load has tripled in three years, the PDU spec written from a 2022 template is already obsolete. Our engineering team treats 30 kW as the new “mainstream AI rack” baseline and quotes 50 kW+ configurations as the forward-looking baseline for buyers planning a 2027 deployment.

The Plug and Receptacle Map for High-Density GPU Racks

Once the load target is set, the next decision is the plug/receptacle map — what the PDU accepts on its input end, what it exposes on its output end, and how those mating connectors survive 30 kW of continuous thermal stress. Because the wrong plug map means the rack cannot be commissioned at the data center, this decision is made before the rack layout is finalized, not after.

For US-market AI GPU racks, the input plug is typically a NEMA L21-30P (3-phase, 30A, 120/208V) for 30 kW and a NEMA L22-30P (3-phase, 30A, 277/480V) for higher-voltage deployments. The mating PDU input is an IEC 60309 32A connector — the same family of plugs used in industrial settings worldwide. Our 3-phase PDU for American racks product line covers NEMA 5-15R, NEMA 5-20R, and lockable outlets for US deployments, with the same aluminum-alloy housing and 0–60°C operating envelope our 3-phase EU and AU SKUs share.

For EU-market racks, the equivalent input is an IEC 60309 32A or 63A plug at 400V three-phase. The output side on the PDU is where the spec gets interesting.

Outlet Density vs Outlet Diversity: Why Both Numbers Matter

A 30 kW rack needs roughly 24–42 outlets depending on the rack profile. Because outlet density (count per PDU) and outlet diversity (mix of C13, C19, C21) are independent specs, a buyer optimizing for one and forgetting the other ends up with a PDU that has the right count but the wrong connector mix — or the right mix but insufficient outlets.

Outlet Connector Guide for AI GPU Racks

According to the IEC 60320 standard for appliance couplers, the C21 inlet is rated 16A at 250V AC and is the matched mating connector for high-temperature GPU server power supplies — meaning a 30 kW+ rack spec’d in 2026 should not be all C19, but a mix of C19 (for current GPU PSUs) and C21-compatible cabling for the next refresh.

Phase Balancing Math: The 5% Tolerance That Protects GPU Clock

The single most common silent failure in a 30 kW AI rack is phase imbalance. Because GPU training jobs run in lockstep across all 8 GPUs in a node, a transient imbalance event that drops voltage on one phase can throttle the GPU clock on the affected node while the others continue at full speed. The training job doesn’t fail outright, but it stalls — and a 10-minute stall in a 30-day training run is the difference between on-time and late model delivery.

Industry guidance on three-phase distribution recommends keeping phase imbalance below 10%. Hyperscale AI deployments typically target 5% to leave headroom for transient load shifts during gradient-synchronized all-reduce operations. Because the imbalance tolerance shrinks as GPU synchronization tightens, the PDU spec for an AI rack is stricter than the PDU spec for a general-purpose cloud rack.

The phase balancing math works like this. If your rack draws 30 kW total, ideally each phase carries 10 kW. In practice, you can never perfectly balance GPU loads because different GPU SKUs have different power curves, and different training jobs have different phase-dependent power signatures. A 5% imbalance budget means no phase carries more than 10.5 kW when the others carry 10 kW — a tight target that drives both the PDU outlet group arrangement and the upstream breaker balance. The PDU spec must document the outlet group mapping to phase, and the commissioning procedure must verify the imbalance after deployment.

Redundancy Architectures: How N+N vs 2N Changes the PDU Order

Every 30 kW+ AI rack has at least two PDUs, and the redundancy architecture determines how those two PDUs are sized. Because the redundancy decision drives the BOM, the per-PDU upfront cost, and the commissioning complexity, this is the conversation that should happen before the PDU model is selected, not after.

N+N vs 2N vs 1+1: Redundancy Architecture Comparison

For a 30 kW+ AI rack, N+N is the most common configuration: two PDUs per rack, each independently sized for the full 30 kW load, fed from separate UPS sources or PDU busways. Because each PDU runs at 50% load under normal operation, both PDUs sit comfortably in their continuous-load rating with thermal headroom. The 2N configuration doubles the equipment count and is reserved for tier-IV facilities where any single component failure must not affect IT load. The 1+1 configuration shares load across two PDUs and only works if you accept that one PDU failure under full load will trip the rack — usually not acceptable for production AI training runs.

Spec Workflow: From Rack kW Target to a Quoteable PDU BOM

The fastest path from a 30 kW target to a quoteable PDU BOM runs through six questions. Because the answers to these questions directly map to PDU SKU, input plug, outlet mix, and certification scope, walking through them once takes 20 minutes and saves weeks of back-and-forth in the OEM quotation cycle.

1. What is the target per-rack kW and the GPU SKU mix? H100 vs B200 changes the per-rack outlet count and the C19/C21 split.
2. What is the supply voltage and phase? 208V three-phase in the US, 400V three-phase in the EU, 415V three-phase in Australia.
3. What input plug standard does the data center require? NEMA L21-30P / L22-30P for US, IEC 60309 32A/63A for EU/UK/AU.
4. What is the outlet count and mix per rack? Typical 30 kW AI rack: 24–42 outlets, mix of C13 / C19 / C21 with locking variants.
5. What redundancy architecture? N+N is the AI training default; 2N for tier-IV; 1+1 for entry-level redundancy.
6. What certifications are required for the data center market? UL for US, CE for EU, UKCA for UK, EESS for Australia, CCC for China deployments.

The output of these six questions is a one-page PDU BOM: SKU, input plug, outlet count, outlet mix, certification scope, per-PDU quoted cost, and lead time. Because this BOM is what our OEM quotation returns within 24 hours of the spec conversation, having the answers pre-filled by the buyer’s infrastructure team cuts the procurement cycle dramatically. For a complex AI deployment — say, a 200-rack training cluster — the BOM is identical across racks, which means the per-PDU upfront cost drops materially with volume.

To start a 30 kW AI rack spec, share your rack count, GPU SKU, and target supply voltage via the contact form. Our engineering team returns a preliminary BOM with two PDU configurations (typically an N+N pair and a tier-IV 2N option) within 24 hours, plus the certification scope for the target market. From there, the spec conversation is a matter of tightening the outlet mix, confirming the input plug standard with the data center, and verifying the phase balance after the first racks are commissioned.

Frequently Asked Questions

What is the best PDU for an AI GPU rack pulling 30 kW?

The best PDU is a 3-phase vertical unit with an IEC 60309 32A/63A input (or NEMA L21-30P/L22-30P in the US), a mix of C19 and C21 outlets, and N+N redundancy.

How many outlets does a 30 kW AI rack need?

A typical 30 kW AI rack needs 24–42 outlets depending on the GPU node density. Higher-density racks (50+ kW) typically need two vertical PDUs per rack.

What is the maximum allowed phase imbalance for an AI rack?

Industry guidance recommends under 10%. Hyperscale AI deployments typically target 5% to leave headroom for transient load shifts during synchronized all-reduce operations.

What is the difference between C19 and C21 outlets?

The C19 outlet is rated 16A at 250V AC. The C21 inlet is the high-temperature variant matched for newer GPU server power supplies. A modern AI rack spec includes both.

Do I need N+N or 2N redundancy for AI training racks?

N+N is the standard: two PDUs per rack, each independently sized for the full load, fed from separate UPS sources. 2N is reserved for tier-IV facilities.

What input plug does a US-market 30 kW AI rack need?

The standard US input plug for a 30 kW AI rack is a NEMA L21-30P (120/208V three-phase, 30A) or NEMA L22-30P (277/480V three-phase, 30A).

Are Newsunn 3-phase NEMA PDUs UL listed?

Yes. Newsunn’s US-market NEMA PDUs are UL listed since 2018. The PDUs are also CE, GS, EESS, REACH, and ISO 9001 certified.

What is the typical lead time for a 200-rack AI training cluster PDU order?

For OEM orders of 200+ matched PDUs, typical production lead time is 30–45 days, plus 10–20 days for ocean freight.