Transformers, Harmonics & the AI Non-Linear-Load Problem

Data-center power-quality engineering was built around the 6-pulse rectifier. Double-conversion UPSs and variable-frequency drives drew current in distorted gulps, rich in the 5th, 7th, 11th and 13th harmonics that a three-phase 6-pulse front end produces. Older single-phase server supplies added their own distortion, including strong third-harmonic content. Harmonic currents do no useful work. They raise conductor and transformer losses, and triplen (third and other zero-sequence) currents add up in a four-wire neutral, where balanced 5th and 7th currents do not. Inside the transformer, winding eddy-current loss grows with the square of harmonic order, while other stray losses rise more slowly. Take equal RMS current. Winding eddy-current loss from a 5th harmonic (300 Hz on a 60 Hz system) is about 25x the fundamental's; the 7th gives ~49x and the 13th ~169x (the guide's own arithmetic, using the IEEE C57.110 h² weighting). The industry's answer was a vocabulary: the K-factor transformer, derating per IEEE C57.110, and IEEE 519 limits at the point of common coupling.

The AI hall upends that instinct. A modern accelerator power supply is not a 6-pulse load. That covers the 5.5 kW supplies in OCP power shelves for GB200- and GB300-class racks, the 12 kW unit in OCP's HPR V2 shelf specification, and the 18.3 kW supplies inside Vera Rubin NVL72's 110 kW shelves. Each is an active rectifier with active power-factor correction. It draws near-sinusoidal current at near-unity power factor, with a current THD that must be specified by load point and source conditions. ABB’s 2017 drive comparison, p.24 reports <5% for its low-harmonic drive at nominal load. That is not a figure for every accelerator PSU or for the PCC, and an AFE label is not proof of IEEE 519-2022 compliance.

The harmonic villain the K-rating was invented to fight has largely been engineered out of the load. What is left is a time-domain problem the harmonic textbooks never contemplated: AI training can create large, correlated power swings, with thousands of clean rectifiers moving from idle to full power in milliseconds. Spend on the wrong problem and you buy a heavy K-20 transformer to tolerate harmonic heat that the measured AFE spectrum does not produce. Meanwhile the synchronized load step goes unmitigated, and it shows up as voltage flicker, ramp-rate violations and protective-relay trips toward the grid.

Every campus has a chain of step-downs: utility HV (69–500 kV; 230 kV and above for gigawatt campuses, per Chapter 4.2) to MV (typically 13.8–34.5 kV) at the customer substation, MV to LV (415/480 V) at the pod or block transformer, and, in the legacy world, further conversion inside the rack. Three insulation and cooling families compete for each position. The choice trades loss, fire and containment exposure, footprint and serviceability, and in 2026 harmonics are rarely what decides it.

Liquid-filled (mineral-oil or natural-ester). This is the lowest-loss, highest-density, longest-lived option, and the default for HV/MV power transformers (50–100 MVA substation units) and larger outdoor pad-mount MV-to-LV blocks. Mineral oil cools and insulates well but is flammable and a containment liability. Natural-ester (vegetable-oil) fluids, now common on new AI builds, have a fire point above 300 °C, are biodegradable and ease spill-containment and setback requirements, at a modest cost premium. Indoors, the large mineral-oil units discussed here normally go in a transformer vault with containment under NEC 450.26, subject to that section's exceptions. NEC 450.23 lets a listed less-flammable liquid with a fire point of 300 °C or more go in with no vault at 35 kV and below, in a Type I or II building, if no combustible materials are stored in the area, the liquid is confined and the liquid listing's restrictions are met. A second route at 35 kV and below combines automatic fire extinguishing with liquid confinement. Above 35 kV the ordinary route is a vault. The NEC edition adopted where you build governs. Outdoors at the substation, fire separation and liquid containment are easier to provide; close to the hall, they are awkward.

Dry-type (VPI / ventilated). Air-cooled, with no liquid to contain and a lower fire load, dry-type is the workhorse for indoor MV-to-LV and LV-to-LV positions inside or near the electrical room. The penalties are a larger footprint per MVA, lower overload tolerance and a shorter thermal life under sustained high loading. Losses still matter where every megawatt is precious. For a covered three-phase 1,000 kVA distribution transformer, the federal efficiency floor at 50% load is 99.43% when liquid-immersed and 99.20% when MV dry-type with a 46–95 kV BIL (10 CFR 431.196, checked 13 October 2026). Those are legal minimums, not measured differences between offers. Ask each bidder for guaranteed no-load and load-loss watts at the same kVA, BIL and load profile, then price the annual losses. What dry-type buys is a unit you can site against the wall without a vault.

Cast-resin (cast-coil epoxy). This is the premium dry-type: windings encapsulated in epoxy, with excellent moisture, dust and short-circuit resistance, very low flammability (self-extinguishing), and the ability to sit next to occupied or critical space with minimal containment. Choose it where fire safety and environmental robustness dominate: close to the white space, in seismic zones, in humid or contaminated environments. It is the most expensive of the three and shares dry-type's loss and footprint tradeoff. It is still often the right close-in MV-to-LV block transformer in a dense AI hall, because it removes the oil-containment problem exactly where you want the transformer near the load.

For thermal capacity on your measured spectrum, use the published IEEE C57.110-2018 method and the manufacturer's calculation. IEEE approved a project to revise it (PC57.110) in February 2026, but until a new edition is published the 2018 text is the one to use; Appendix A tracks both.

Three 80 MVA units miss the N−1 requirement. Required planning apparent power is 150 MW × 1.10 / 0.95, about 174 MVA. Each survivor needs 150 × 1.10/(0.95 × 2 × 0.95), about 91 MVA. Two 80 MVA survivors supply 2 × 80 × 0.95 × 0.95, about 144 MW, and two 100 MVA survivors supply about 180 MW. Against the 150 × 1.10 MW requirement, 80 MVA fails and 100 MVA passes.

Carry three 100 MVA units into tender, conditional on the actual spectrum, temperature-rise guarantee, impedance, inrush, transport and protection studies. The load crossover is 2 × 100 × 0.95 × 0.95 / 1.10 ≈ 164 MW facility demand. At the assumed 170 MW flip load, 170 × 1.10 exceeds 2 × 100 × 0.95 × 0.95 and the same candidate fails. An offer close to the crossover needs guaranteed bounds before you accept it; otherwise reduce the accepted load, obtain a larger verified site rating, or add reachable capacity. Equal nameplates without a surviving bus path are not redundancy.

The designer hands the load and PF range to the transformer supplier, and the supplier returns its IEEE C57.110 thermal calculation and guarantees. The screening above is arithmetic, not a heat-run result; Chapter 4.12 covers acceptance of the delivered units' test reports and settings.

The K-factor is one number that summarizes how much extra eddy-current heating a transformer must tolerate from a given harmonic spectrum. It weights each harmonic's current by the square of its order. A purely linear load is K-1. A hall of legacy switch-mode supplies can demand K-13, and a dense bank of older 6-pulse front-ends K-20. Two paths buy the same thermal headroom. You can specify a K-rated transformer (oversized neutral, transposed or sub-divided conductors, extra cooling margin built in), or apply a harmonic derating factor to a standard unit per IEEE C57.110. If the manufacturer's calculation for your spectrum yields an illustrative 0.85–0.90 factor, a 1000 kVA unit supplies ~850–900 kVA for that spectrum and environment. THD alone does not select the factor.

If the offered thermal calculation requires 10–15% derating, compare the stranded capacity and larger pad of the standard unit with the K-rated unit's price, losses, footprint and delivery. The K-rated offer wins when its installed and lifetime costs beat that lost capacity. Keep the derated standard unit when its offer is cheaper and the extra space is acceptable.

“The AI hall is 100% non-linear” has to be read carefully, because the phrase is true in two different senses that point to opposite mitigations. In the frequency domain, the harmonic-spectrum sense the K-factor measures, a modern AFE-fed AI hall is not badly non-linear: the per-unit current THD is low, power factor is near unity, and the 5th/7th/11th content the K-rating fights is modest. In the time domain, the AI hall is a large, fast and correlated load. Every PSU is a switching converter whose aggregate draw can move with the synchronous workload. Within one training job the participating GPUs work through the same step, so their swings correlate, although job partitioning and power smoothing change the aggregate. Take idle power, step amplitude, duration and recovery from a measured rack waveform before you assign a millisecond duty. A transformer winding does not solve this. It is a transient and grid-interaction problem, solved by capacitance, energy storage and software power-capping. Conflating the two senses is a common and costly error: you over-spend on K-rating and under-spend on transient absorption.

IEEE 519 is the rulebook between the facility and the grid. It does not regulate any single piece of equipment. It caps distortion at the point of common coupling (PCC), the boundary where your loads meet the utility and other customers, and the limits depend on the PCC's voltage. IEEE 519-2022 Table 1 caps voltage THD at 8% (5% per individual harmonic) at 1 kV and below, 5% (3%) above 1 kV through 69 kV, 2.5% (1.5%) above 69 kV through 161 kV, and 1.5% (1%) above 161 kV. Those are weekly 95th-percentile values of 10-minute measurements, and the daily 99th-percentile limit on 3-second values is 1.5 times the table value.

Current is limited as total demand distortion (TDD). Table 2 sets 5–20% TDD for PCCs from 120 V through 69 kV, depending on the short-circuit-current-to-load ratio Isc/IL: a stiffer grid tolerates more current distortion because it produces less voltage distortion. Tables 3 and 4 hold tighter limits above 69 kV and above 161 kV, so a campus with a 115 kV PCC uses Table 3 and one at 230 kV uses Table 4. Find your actual PCC, pick its Isc/IL row, and read the individual-harmonic current limits alongside the TDD. IEEE 519-2014 already based current limits on TDD, which divides by maximum demand load current. Current THD instead divides by the fundamental current. IEEE 519-2022 keeps both the voltage and the current limits at the PCC.

The mitigation hierarchy runs cheapest and most effective first. Design the harmonics out at the source. If the loads are AFE/active-PFC PSUs, the specified load-point spectrum may be clean, but background distortion, parallel equipment, source impedance and resonances still decide IEEE 519 compliance at the PCC. That is the best reason the AI hall's harmonic reputation is overstated. If legacy 6-pulse loads exist (older UPSs, mechanical VFDs on chillers and pumps), the choices, in rough order of rising cost and effectiveness, are:

• Passive tuned filters. They are cheap, but they detune and risk resonance with PFC capacitance and cable charging.
• 12- or 18-pulse rectifier configurations. Phase-shifting transformers cancel low-order harmonics, which is effective but bulky and lossy.
• Active harmonic filters. They inject cancelling current across several harmonic orders, but only up to their current rating, bandwidth and control limits. They are expensive. Compare residual distortion at each load point and source state: ABB's brochure (p.9) shows THDi of 12–14% at 50% load for one active-filter comparison, and that is not a PCC result.

The transformer K-rating sits underneath this hierarchy as the thermal backstop, not the primary tool. A facility that K-rates everything and skips the active filter will still trip the PCC limit if a chiller VFD farm is the real offender, because the transformer absorbs heat but does nothing to reduce the current injected toward the grid.

The active front end (AFE) resolved the harmonic half of the AI power-quality problem. A passive 6-pulse rectifier conducts only at the peaks of the voltage waveform, drawing the distorted, harmonic-rich current that the K-factor was invented to survive. An AFE replaces the diode bridge with actively switched devices (IGBTs, increasingly SiC) under closed-loop control. At its rated operating point it draws near-sinusoidal current at near-unity power factor and can reduce distortion at its terminals. That equipment-level current THD is separate from the IEEE 519 PCC TDD test. The AFE regulates the boosted DC bus and, in regenerative variants, can return power to the grid. Ask the supplier of an OCP power-shelf rectifier or a GB200/GB300/Rubin-class PSU for its actual active-PFC topology and spectrum. “AFE” does not certify every modern unit or every load point.

Specify the delivered AFE PSU/UPS spectrum and you can cut harmonics at the source, before they reach transformers, neutrals and the PCC. Verify the remaining mixed-load spectrum and source impedance before you remove any mitigation. Where that study passes, you may avoid a K-20 premium, an oversized neutral or an active filter. Where it does not, their thermal or distortion duty remains, and passive correction still needs a resonance check. The cost sits in the PSU or UPS itself. Against a passive diode front end, active switching adds complexity and gives up a little efficiency at part load. Murata's MWOCP67-5500-B-RM is a 5.5 kW PSU for Open Rack V3 (ORV3) HPR power shelves. Its datasheet (revision B, accessed 13 October 2026) gives fan-inclusive efficiency of 94.0% at 10–30% load on 208–277 VAC. At 30–100% load it gives above 96.5% on 230–277 VAC and above 95.5% on 208 VAC. Its 97.5% is a peak figure at 230–277 VAC, not an efficiency across the load range. AFE units are also sensitive to their own control stability and to weak or islanded sources. In a 2026 greenfield build, the front-end conversion is where the harmonic argument is won or lost. Win it there and the transformer family decision reverts to a loss, fire and footprint optimization instead of a harmonic-survival exercise.

Everything above optimizes the conventional chain: a 50/60 Hz iron-and-copper transformer steps MV to LV, an AC UPS provides ride-through, PDUs distribute, server PSUs rectify to DC, and VRMs step down to the chip. The solid-state transformer proposes to delete most of that. An SST is a power-electronic converter, a medium-frequency transformer plus SiC/GaN switching stages, that converts MV AC to regulated 800 VDC in one converter package containing internal switching and isolation stages, at the facility perimeter. The efficiency reference point is about 98% for a US DOE project prototype (2018–2021) rated 400 kW, 13.2 kVAC to 800 VDC. Kolar, Huber and Drofenik of ETH Zurich and TU Wien cited it in their INTELEC keynote of 13 October 2025, and the same keynote put the state of the art at 98–98.5% and the next target at 99%. That is a 2025 assessment of a prototype, not a product rating. For a purchase, get the offered converter's own test report, with its input and output boundaries, loading, auxiliaries and isolation. Chapter 4.1’s assumptions serve for screening.

Chapter 4.1’s reference AC chain stacks four functional conversion packages (transformer, UPS, PSU, final VRM; the PDU provides distribution) and loses ~13–20% end-to-end. The SST-fed 800 VDC chain collapses to three (SST, rack DC-DC, final VRM), and SemiAnalysis models ~87% utility-to-VRM for its Phase 4 architecture (May 26, 2026). That is an analyst model, separate from Chapter 4.1’s assumed stage losses. The difference of about five points from the guide’s AC reference case needs matched load and loss boundaries before it counts as a procurement benefit.

This chapter treats the SST as a transformer, even though its payoff is felt downstream in the DC architecture. It changes which conversion, protection and service equipment occupies the electrical room, but it does not remove the need for isolation or outage energy. For harmonics, the SST's grid side may be an active rectifier, a diode bridge plus an active filter, or a direct matrix converter. Grid-current quality therefore depends on the topology and controls you are offered, and you should verify it at the PCC across load, voltage and failure range. The K-factor question disappears for the line-frequency transformer the SST replaces, and isolation moves to a medium-frequency magnetic inside the package. Any upstream transformers you keep still need a thermal check against their actual spectrum. The SST also natively provides the regulated DC bus that the disaggregated 800 VDC sidecar architecture wants.

The transformer is a schedule choice as much as an engineering one. Ben Boucher of Wood Mackenzie says US load-serving substation transformers now take more than 160 weeks to deliver, against about 140 weeks in 2023 (Data Center Knowledge, May 4, 2026). Siemens Energy reported lead times of up to five years for US large power transformers in its February 14, 2024 factory announcement. Read that as the market's upper tail on that date, not a current survey or a promised delivery. Generator step-up transformers (GSUs) are a separate category with their own dated register in Chapter 2.3. Transformer delivery is often the longest pole in energizing an AI campus, so the family and rating decision moves to the front of the project, before the load is fully characterized.

Country of origin is now a gate as well as a schedule risk. Executive Order 14421 (August 26, 2026) lets the US Department of Energy prohibit, condition or unwind purchases of bulk-power equipment from covered foreign entities (keyed to the ITAR §126.1 country list, which in practice means China), and order installed units monitored, isolated or removed. It reaches transmission at 69 kV and above, substations and their control systems, so the HV substation's power transformers and breakers are inside it. DOE issued a request for information on September 9, 2026, and its rules are due by December 24, 2026. Before you sign, confirm country of origin and supplier ownership for every item at 69 kV and above and for the BESS, and write who pays for a forced replacement into the supply contract (Chapter 2.3).

A late change can be costly, but the purchase date alone does not freeze the design. Separate the date you reserve the slot from the contract's design-freeze and change-control milestones, and settle the load and harmonic range before the affected rating and thermal design freeze. Get the supplier's written redesign cost and slot terms, and price a later change against them. Chapter 2.3 and Chapter 2.4 cover reservation and contract mechanics. As everywhere in this guide, sort the forks by reversibility and spend the option premium to match.

US efficiency minimums for distribution transformers with input voltage of 34.5 kV or less (10 CFR 431.192 and 431.196) depend on when the unit is made, so tender against the manufacture date. Under the DOE rule of April 22, 2024, a three-phase 1,000 kVA unit made from April 23, 2029 must reach 99.46% (non-submersible liquid-immersed) or 99.28% (MV dry-type, 46–95 kV BIL) at 50% load, up from 99.43% and 99.20%. Ask the supplier for a compliant design and price for your production slot.

Hard to reverse (decide at scoping): the substation transformer family, rating and MVA cushion (apply planning margin, power factor and site rating independently, as in the 150 MW case); the fundamental architecture fork, conventional MV-to-LV-to-DC chain versus an MV-to-800-VDC SST path, because it dictates the entire electrical room, the footprint and the downstream 4.5/4.6/4.7 design basis; and the close-in block-transformer family (cast-resin vs liquid-filled), which sets fire and containment and how near the load you can site it. Reversible (defer, re-decide cheaply): the K-rating margin and harmonic-derating factor on replaceable distribution transformers; active-filter sizing and placement; and the specific AFE PSU/UPS generation within a fixed conversion architecture. The strategic move is to convert irreversibility into optionality where the premium is cheap: provision MV capacity and perimeter space so an SST can be retrofitted, and standardize on AFE front ends so the harmonic posture never depends on a transformer you cannot re-order.