The answer first: size the data center transformer on the after-diversity design load — not the sum of rack nameplates and not the utility service rating — and let three forces set the real specification: AI-era rack density, the redundancy topology (N+1 vs 2N), and the harmonic content of IT rectifier loads
A data center transformer sits between the medium-voltage utility feed — typically 13.8 kV, 20 kV or 33 kV — and the low-voltage distribution that feeds UPS modules, power distribution units (PDUs) and IT racks at 480 V or 415 V. Its kVA rating follows the design load after diversity, then redundancy (N+1, 2N, 2(N+1)) multiplies the number of units, not each unit beyond full load. Four properties separate a data center transformer from a commercial-building unit: the load runs at a high, flat load factor around the clock so loss grade drives total cost of ownership; IT switch-mode rectifiers inject harmonics that add winding eddy-current loss; the redundancy topology decides how many units and at what rating you buy; and the installation is usually indoor, close to people and equipment, which pushes toward dry-type or ester-filled units and a low noise footprint. Specify those four correctly and the fleet runs cool and efficient for its 20-year life; size on raw connected kW and you either strand capacity or overheat windings.
Key takeaways
- Size on the design load, then apply redundancy. Transformer kVA follows the after-diversity design load (connected load × a diversity factor), not the arithmetic sum of rack nameplates; redundancy then multiplies the number of units.
- Harmonics from IT loads raise winding loss. Servers and GPUs rectify AC with switch-mode supplies; specify a K-factor rating or a documented IEEE C57.110 derate, and consider zigzag or phase-shifting units to cancel triplen and 5th/7th harmonics.
- Load factor is high, so losses dominate TCO. Because the unit sits near full load continuously, a better loss grade and low no-load loss pay back fastest here — the data center is the asset where efficiency specs matter most.
- Code and layout pick the type. Indoor halls and NEC 450/645 rules push toward dry-type (cast-coil) units; outdoor or vaulted substations allow oil-immersed.
- Noise and heat are first-class constraints. Near office space and whitespace, a low-noise design and a defined cooling class belong in the specification, not on the afterthought list.
What does a data center transformer do — and how is it different from a commercial-building unit?
In a data center the transformer is a step-down device, not the step-up unit you would specify for a solar or wind plant. Utility power arrives at medium voltage and the transformer drops it to the low voltage that UPS modules, PDUs and rack PDUs actually use. A large double-conversion UPS is commonly fed at 480 V (North America) or 415 V (Europe/Asia), so the classic chain is: MV switchgear → step-down transformer → LV switchboard → UPS → PDU → rack. Some high-density halls feed the UPS directly at medium voltage and step down at the row, but for most colocation and enterprise halls the workhorse is the MV/LV step-down transformer.
Three properties make it a different animal from an office-building transformer:
- Flat, high load factor. An office transformer spends most hours lightly loaded and idles at night. A data center transformer runs near its design load every hour of the year, so both core (no-load) and winding (load) losses accrue continuously — which changes the economics of efficiency completely.
- Harmonic-rich load. Office loads are largely linear (motors, lighting). IT loads are rectifier-front-end: they draw non-sinusoidal current that heats transformer windings in ways a simple kVA rating does not capture.
- Redundancy is built into the count. A commercial building may buy two transformers and never run both. A data center may buy four or ten in a block-redundant scheme where the number and rating are set by the reliability tier, not by peak demand alone.
How big is an AI data center's electrical load — the load that actually sizes the transformer?
The single largest change in data center electrical design over the past three years is rack power density. A conventional enterprise rack drew roughly 5–10 kW. A mainstream AI/GPU rack today is specified in the 30–50 kW range, and the densest liquid-cooled GPU racks are specified well above 100 kW (industry experience / 行业经验值, 2024–2025 vendor roadmaps). Higher rack density does not change transformer physics — a kVA is a kVA — but it changes two things that matter for sizing:
- The electrical load of a hall can grow by an order of magnitude on the same floor area, so the medium-voltage step-down transformer becomes the gating capacity item for how fast a hall can be built out.
- Dense halls concentrate many large loads in a small footprint, which raises fault current, cooling load and harmonic concentration at the transformer — all of which feed back into the specification.
The practical takeaway: do not size the transformer for today's populated racks alone. Data centers are built to a target density and filled over time, so the transformer must be rated for the design (target) load of the hall, not the load of the hardware installed on day one.
How do you convert IT load in kW to transformer kVA?
The conversion is straightforward once you fix the assumptions. The transformer must supply apparent power, so:
kVA = kW ÷ power factor, and then design kVA = (connected IT kW ÷ power factor) × diversity factor × spare margin.
Three assumptions do the real work:
- Power factor. Modern IT power supplies present a power factor near 0.9–0.95 at the UPS input (industry experience / 行业经验值). Using 0.9 is a conservative default; verify against the UPS and PDU datasheets.
- Diversity factor. Not every rack draws its nameplate at once. A diversity factor — often 0.8–0.9 for a mature hall (industry experience / 行业经验值) — converts connected load to a realistic simultaneous load. Using 1.0 quietly oversizes every unit in the chain.
- Spare margin. A growth/spare allowance is added for future fill and for load imbalance across phases. Then the result is rounded up to a standard rating — common MV/LV distribution sizes run 1000, 1250, 1600, 2000, 2500 and 3150 kVA.
This is the same discipline used to size a whole distribution system; the sizing and budget formula article walks through the arithmetic with worked numbers.
N+1 vs 2N vs 2(N+1): how does redundancy change the transformer count and rating?
Redundancy is the most misunderstood input to transformer sizing. The rule to hold in your head: redundancy multiplies the number of units you install, not the rating of each unit past its block load.
- N (no redundancy). The installed capacity equals the design load. Lose one unit and you lose the block — acceptable only for the least critical spaces.
- N+1 (distributed or block redundant). Install one more unit than needed to carry the load. If a block needs N units to carry the design load, install N+1, each rated at design load ÷ N, so any single unit can fail with the rest still carrying full load. Example: a 3 MVA block at N+1 is four units of about 1000 kVA each.
- 2N. Two completely independent systems (A and B), each sized to carry the full design load on its own. Installed transformer capacity is therefore about twice the design load, and the two systems should be physically and electrically separated.
- 2(N+1). Each of the two independent systems also carries its own spare — the highest redundancy tier and the highest installed capacity.
The practical consequence for the electrical engineer: the redundancy tier (as classified by frameworks such as the Uptime Institute Tier Standard: Topology) is chosen first, and it fixes how many transformers exist and whether each is sized for a share of the load (N+1) or for the entire load (2N). Unit kVA then follows from the block load, not from multiplying by "two."
Why are harmonics the biggest hidden driver of data center transformer sizing?
Data center loads are rectifier-front-end. Servers, GPUs, network gear and UPS modules convert AC to DC with switch-mode supplies, and that conversion draws current in short pulses rather than a smooth sine. The result is a non-sinusoidal current with harmonic content — historically strong in the 3rd and 5th harmonics, and still significant in the 5th, 7th and higher orders with modern three-phase supplies.
Why this matters for the transformer specifically: harmonic currents raise the eddy-current and stray losses in the windings, and those losses grow roughly with the square of the harmonic frequency. A transformer that is fine on a purely sinusoidal load can run hot on a harmonic-rich load at the same RMS current. The fixes are standard and well documented:
- K-factor rating. Specify a transformer built and tested for non-linear loads (K-4, K-9, K-13, K-20 in UL 1561 terms). A K-factor unit carries a stronger core, lower flux density and a larger neutral, so it tolerates the harmonic heating. See K-factor transformer selection.
- Documented derate. Where a K-factor unit is not used, apply the derating method of IEEE C57.110 to establish the transformer's capability on the actual load current spectrum, as covered in the transformer harmonics guide.
- Harmonic cancellation. Vector-group pairing (delta–wye and zigzag combinations) and phase-shifting transformers cancel the 5th and 7th harmonics when two rectifier groups are fed from different phase shifts; zigzag grounding transformers trap triplen (3rd-order) currents.
- Grid-side limits. IEEE 519 caps harmonic current and voltage distortion at the point of common coupling, which bounds how much harmonic current the upstream grid will accept and therefore how much filtering or cancellation the design must provide.
The design lesson: harmonics are not a "nice to have" footnote. On a modern IT load, K-factor, neutral sizing and harmonic cancellation are part of how the transformer is rated, not an accessory added later.
Dry-type or oil-immersed for a data center — what do codes and layout decide?
The transformer type decision in a data center is driven mostly by location and fire codes, not by price alone.
- Dry-type (cast-coil or resin-encapsulated). Preferred indoors because there is no oil to contain, no oil-fire risk and no bunding requirement. Cast-coil units also suit the dusty, humidity-controlled electrical rooms of a data hall. They are the default for the MV/LV step-down position inside a building.
- Oil-immersed. Typical for outdoor pads, vaults or a dedicated on-site substation, and generally more economical at the higher kVA ratings. Maintenance and oil-containment requirements apply.
- Natural ester / silicone-filled. A middle path where fire resistance matters but an oil-immersed unit is desirable; higher fire point and better moisture tolerance than mineral oil, at a higher price.
Two code references govern the installation: NFPA 70 (NEC) Article 450 sets transformer installation and protection rules, and Article 645 provides the allowances for information-technology equipment rooms when specified conditions are met (dedicated space, disconnecting means, etc.). Article 708 applies where a facility is classified as a critical operations power system. Where the project follows IEC practice, IEC 60076 governs the transformer itself and IEC 61000-3-2 / -3-12 bound harmonic emission from the connected equipment.
Efficiency and loss grade: why data centers specify low-loss, and how fast it pays back
Because a data center transformer runs near its design load continuously, both its no-load (core) loss and its load (winding) loss accrue essentially every hour of the year — roughly 8,760 hours. That is the worst possible place to buy a cheap, high-loss transformer, and the best possible place to justify a premium loss grade.
The payback logic is simple: the annual cost of losses equals the total watts lost, times the hours, times the electricity price — plus, in a data center, the cost of removing that heat, because every watt of transformer loss also becomes cooling load. That extra cooling term is why data centers justify low-loss grades that a warehouse would never buy. Compare the grades in the S11 vs S13 vs S20 loss-grade comparison and the global energy-efficiency standards comparison, and run the lifetime arithmetic as described in the total cost of ownership guide.
Where no-load loss dominates (a lightly loaded or growing hall), an amorphous-core (S15/SH15) unit cuts core loss dramatically; where load loss dominates (a near-full-load hall), a lower-resistance winding and a better loss grade matter more. Whichever you choose, decide it on lifetime cost, not first price.
Noise, temperature rise and footprint — the constraints a container or warehouse hall adds
Data centers increasingly sit in converted warehouses and modular/container builds, which adds constraints that do not appear on a datasheet's headline kVA:
- Noise. Transformer hum comes from core magnetostriction and grows with core flux density and size. In an electrical room adjacent to office space or near whitespace, a low-noise specification (lower flux density, better core stacking, acoustic enclosure) is part of the requirement.
- Temperature rise and cooling class. The specified temperature rise and cooling method (AN/FN/AF for dry-type; ONAN/ONAF for oil) set how much heat is dumped into the room and therefore into the cooling system. In a tightly packed hall, the transformer's own heat load matters. See cooling methods and temperature rise.
- Footprint and weight. Rooftop and container deployments are weight- and space-limited; the winding material choice and enclosure type shift both.
- Voltage regulation. Long bus runs and harmonic load can drop the delivered voltage; specify tap range and verify regulation with the Uk% and regulation method.
What should appear on a data center transformer order specification?
To remove ambiguity from quotation and manufacturing, put the following on the order specification (see also how to read a transformer nameplate):
- Type and standard: dry-type cast-coil or oil-immersed, per IEEE C57.12.01 (dry) or IEEE C57.12.00 (liquid), or IEC 60076 / GB/T 1094.
- Rating: rated kVA, primary/secondary voltage ratio, and number of phases.
- Vector group: e.g. Dyn11 or Dyn1 — see the vector-group guide.
- Impedance (Uk%): chosen against the available fault current and the upstream protection.
- Taps: off-load or on-load, and the tap range.
- Cooling class and temperature rise (AN/FN/AF or ONAN/ONAF), plus the associated rating steps.
- K-factor / non-linear rating and neutral sizing for the IT load.
- Efficiency / loss grade (e.g. S13/S20 per GB 20052-2020, or the applicable regional efficiency standard).
- Enclosure, IP rating and noise limit in dB(A) at the reference distance.
- Tests: routine tests plus agreed type/special tests, and the nameplate data set.
Five sizing mistakes that strand data center capacity
- Sizing on connected load instead of after-diversity load. Adding every rack nameplate oversizes the unit and wastes capital and losses for years.
- Ignoring harmonics. Buying a standard transformer for a harmonic-rich load, with no K-factor or documented derate, leads to hot windings and premature insulation ageing.
- Misreading redundancy. Treating 2N as "the same unit twice" instead of "two full-capacity systems" leads either to an oversized single unit or to a system that fails its reliability rating.
- Choosing loss grade on first cost. At a near-constant full load, a high-loss transformer is the most expensive unit in the building over its life — the exact opposite of a low-duty commercial install.
- Leaving cooling and noise off the spec. Heat and sound that the room cannot absorb turn an otherwise correct transformer into an operational problem.
Sources / 资料来源
- Source: IEEE C57.12.00 — General requirements for liquid-immersed distribution, power and regulating transformers
- Source: IEEE C57.12.01 — General requirements for dry-type distribution and power transformers
- Source: IEEE C57.110 — Recommended practice for establishing transformer capability when supplying nonsinusoidal load currents
- Source: IEEE 519 — Recommended practice and requirements for harmonic control in electric power systems
- Source: IEEE 493 (Gold Book) — Recommended practice for the design of reliable industrial and commercial power systems
- Source: UL 1561 — Dry-type general purpose and power transformers (K-factor ratings)
- Source: UL 1562 — Transformers, distribution, dry-type, over 600 volts
- Source: NFPA 70 (NEC) Article 450 — Transformers; Article 645 — Information Technology Equipment; Article 708 — Critical Operations Power Systems
- Source: IEC 60076 series — Power transformers
- Source: IEC 61000-3-2 and IEC 61000-3-12 — Limits for harmonic current emissions
- Source: GB/T 1094 — 电力变压器 (Power transformers)
- Source: GB 20052-2020 — 电力变压器能效限定值及能效等级 (Minimum allowable values of energy efficiency and energy efficiency grades for power transformers)
- Source: Uptime Institute — Tier Standard: Topology (data center infrastructure redundancy classification)
- Source: QDTB Engineering Team — industry experience / 行业经验值 for typical MV/LV step-down ratings, after-diversity and spare-capacity assumptions, K-factor selection and AI-era rack density ranges