Phase Balancing Across AI Racks: Why It Matters and What Actually Works

A 32 A three-phase feed looks like a lot of headroom on paper. 32 A × 400 V × √3 = ~22 kW. That comfortably fits four eight-GPU K-AI servers at sustained load, with margin. So why do half of self-built AI labs end up tripping a breaker in the first month?

The answer is almost always phase balance — or rather, the absence of it. The total rack load is fine. One phase is on fire.

This article is about how three-phase power actually distributes across a rack of AI servers, why AI workloads stress the balancing problem in a particular way, and what to do at PDU outlet level so the breaker stops surprising you.

The math, briefly

For a balanced wye load with line-to-line voltage Vll and total real power Ptot at power factor PF:

I_line  =  P_tot  /  (V_ll × √3 × PF)

On a 400 V European feed at PF = 0.98 carrying 20 kW balanced: I_line = 20 000 / (400 × 1.732 × 0.98) ≈ 29.5 A. Fits a 32 A breaker. Each phase carries ~29.5 A, neutral near zero.

Distribute the same 20 kW unevenly — 12 kW on L1, 5 kW on L2, 3 kW on L3:

Total is still 20 kW. The 32 A breaker doesn't care about totals — it cares about whichever single phase trips first. L1 at 165% of rated trips. The other two phases sat idle.

Unbalanced load doesn't just waste capacity; it converts headroom you paid for into a trip waiting to happen.

The neutral current problem

Three-phase wye systems carry a fourth conductor — the neutral — that returns the vector sum of the three phase currents. With a perfectly balanced linear load the three currents cancel at 120° intervals and the neutral carries zero. With imbalance, the neutral carries whatever doesn't cancel.

A useful approximation for neutral current under imbalanced linear load:

I_N  =  √( I_R² + I_Y² + I_B² − I_R·I_Y − I_Y·I_B − I_B·I_R )

For our 53/22/13 A example that comes out to roughly 35 A flowing through the neutral. In a panel where the neutral conductor is sized for the rated phase current, a 35 A neutral on a 32 A phase rating is uncomfortable. In a panel where the neutral is undersized (an older installation, retrofit work, or an assumption of balanced load), this is where you find melted insulation.

The honest takeaway: in any AI rack worth the name, the neutral conductor needs to be at least the same gauge as the phase conductors, and ideally oversized. We will come back to this in the K-factor section because non-linear loads make it worse.

Why AI workloads stress this more than traditional IT

A traditional enterprise rack — web servers, databases, a couple of network switches — has load that is statistically uncorrelated across machines. Server A's CPU spikes at one moment, server B's at another, the average across the rack is smooth, and balancing across three phases approximately works because the random fluctuations cancel.

AI workloads break that assumption in two specific ways.

Synchronous load correlation. Distributed training runs the same step on every GPU at the same time. A backward pass starts on all eight GPUs of a server simultaneously. If you are training across four nodes, all 32 GPUs hit the all-reduce phase together, then all hit the next forward pass together. The load is not a smoothed average — it is a square wave at the training-step frequency, perhaps 0.5–5 Hz depending on the model.

If all four servers are on the same phase, that phase sees the full square wave. If servers are distributed across phases, each phase still sees the same square wave because the steps are synchronous. Phase balancing does not smooth correlated load. It only flattens the per-phase average.

Sub-cycle transients. A single RTX 5090 is rated at 575 W but capable of transient excursions to 738 W over 5–10 ms, 824 W over 1–5 ms, and 901 W in spikes under 1 ms. An eight-GPU K-AI server therefore has a rated sustained pull around 4–4.5 kW and a per-millisecond peak that can briefly exceed 7 kW. The PSU and the building wiring absorb most of this — but the current spikes are real, they are non-sinusoidal, and they propagate to the upstream PDU.

Standard thermal-magnetic breakers ignore sub-cycle transients. Modern hydraulic-magnetic and electronic breakers do not always; some trip on instantaneous peak, not RMS. If your breaker keeps tripping during specific training phases and the rack is otherwise within its average rating, you are looking at the transient curve, not the average.

What phase balancing can and cannot do

This is why people who came from traditional datacenter operations sometimes get caught out: in their world, balancing is the whole story. In an AI rack it is necessary but not sufficient.

Outlet-level balancing strategy

The practical work happens at PDU outlet level. A 32 A three-phase 0U rack PDU presents banks of outlets per phase. The typical layout for a 30-outlet PDU is outlets 1–10 on L1, 11–20 on L2, 21–30 on L3. Some PDUs use a striped or alternating layout for better cable management — outlets 1, 4, 7 on L1; 2, 5, 8 on L2; 3, 6, 9 on L3. Production-grade PDUs colour-code per phase.

The rule is simple and gets violated constantly: assign servers such that the sustained-load sum on each phase is within ~10% of the others. For identical units it's round-robin; for mixed loads it's bin-packing.

Worked example, 32 A 3-phase PDU (~22 kW rated, 17.6 kW continuous at 80%). How many eight-GPU K-AI servers (sustained ~4 kW each) fit?

The realistic budget: a single 32 A 3-phase rack feed supports about three eight-GPU AI servers, not eight. A four-GPU server at ~2 kW sustained roughly doubles that to six per circuit.

This catches most lab planners off guard. The paper headroom of a 22 kW circuit evaporates once you account for breaker curves, transient peaks, and the 80% continuous-load de-rating in most electrical codes.

Measuring it — outlet-metered PDU

A basic PDU gives you nothing. A metered PDU gives you per-phase RMS current at the inlet. An outlet-metered PDU gives you per-outlet current, which is the unit you actually want.

What to watch:

• Per-phase RMS current, trended over time. Most outlet-metered PDUs expose SNMP or a REST API; scrape into Prometheus, graph in Grafana, alarm at 80% of rated.
• Peak current per phase in a sampling window. If the PDU exposes peak hold, log it; if not, sample at 1 Hz minimum.
• Phase imbalance ratio — (max − min) / max across the three phases. Under 10% is good. Over 25% means you are leaving capacity on the floor.
• Neutral current, if the PDU reports it. If neutral is more than ~30% of any phase current under load, something is wrong with either the balance or the harmonic content.

A 32 A PDU is one piece of equipment. Two redundant 32 A PDUs (A and B feeds) double the data you need to track but also double the failure tolerance — if either feed drops, dual-PSU servers stay up, and the surviving PDU has to absorb everything. Size for that case: a dual-feed rack should run each PDU at ~40% nominal so a single-feed failover stays inside breaker.

Harmonics and the K-factor

Every AI server PSU is a switching power supply, which is a non-linear load: it draws current in pulses around the peak of each half-cycle rather than as a smooth sine wave. This injects harmonic currents back into the supply, predominantly 3rd, 5th, 7th in single-phase systems and 5th, 7th, 11th, 13th in three-phase wye.

The 3rd harmonic hurts most in three-phase: it is a triplen harmonic, meaning the 3rd-harmonic currents from all three phases arrive at the neutral in-phase rather than 120° apart. They add arithmetically. A balanced linear load has near-zero neutral current. A balanced non-linear load with significant 3rd-harmonic content can have neutral current approaching or exceeding the phase current.

This is why K-factor transformers exist. K-1 is a normal linear-load transformer; K-13 is rated for substantial harmonic content; K-20 is what serious AI data centres now specify, because AI clusters present essentially 100% non-linear load with no motors or heaters in the mix to dilute the harmonic content.

For a single-rack lab this is rarely your problem — the building transformer is upstream and sized for a mixed load. For a multi-rack AI room added to an existing building, ask the electrician: what is the upstream transformer's K-rating, is the neutral conductor at least the same cross-section as the phase conductors (oversized is better), and is the panel rated for harmonic content? If the answers are "K-1, same as phase, generic," you are fine for one rack but should revisit before scaling to four.

The half-empty-rack honesty problem

Phase balancing assumes you have enough load to distribute. A rack with one server and twenty empty U cannot be balanced — that one server is on whichever phase it's plugged into. The other two phases sit idle, the neutral carries the full single-server current, and the 32 A 3-phase feed is at maybe 8% capacity but 100% imbalanced.

Three honest options:

1. Accept the imbalance. It's one server. The panel can take it.
2. Drop to single-phase. A 32 A single-phase 230 V feed is ~7.3 kW — fine for one or two 4-GPU servers. Simpler, no balancing problem, no triplen-harmonic neutral issue. The right answer for a small lab with no growth plan in the next 12 months.
3. Plan ahead, wire ahead. If you are growing into 6 or 8 servers in the next year, install the 3-phase feed now and balance properly as you add servers. The cost delta against single-phase is small compared to ripping it out later.

Real-world scenario walkthrough

A realistic mid-size lab: two racks, 8× eight-GPU K-AI servers across both, plus networking, storage, and an Intel-Arc test bench. Total sustained ~35.5 kW (32 kW for the K-AI fleet, 2.5 kW for the Arc bench, 1 kW for switches and storage).

Circuit plan: split four K-AI servers per rack, each rack on a 63 A 3-phase feed (~43.5 kW rated, 34.8 kW continuous — comfortable for four 4 kW servers plus ancillaries).

Rack A outlet assignment:

Per-phase: L1 = 4.6 kW (20 A), L2 = 4.2 kW (18 A), L3 = 8 kW (35 A). Imbalanced — two K-AI units share L3. With only four servers per rack you cannot perfectly balance across three phases; one phase always takes the extra. 35 A on a 63 A feed is 56% of rated, well inside breaker. Accept it. Rack B mirrors the layout for the remaining four K-AI servers and the Arc bench.

Dual-feed redundancy is handled by giving each server an A-feed and a B-feed PSU connection on two physically separate PDUs — see W04 for how this interacts with split-not-redundant delivery on AI server PSUs.

Monitoring: Both PDUs export per-phase and per-outlet current via SNMP. Prometheus scrapes every 15 s, Grafana shows 24 h trend, alarm fires if any phase exceeds 80% of rated for >5 min, or if phase imbalance ratio exceeds 25% for >30 min. The smoothing matters — training jobs starting and stopping trigger short-lived imbalances that are not actionable.

---

This is part of the Kentino Wiki, a reference series on AI compute, robotics, and the systems that connect them. Comments and corrections welcome at info@kentino.com.