Data Center Substation Design for AI Loads

A data center substation for AI loads has to convert an IT growth plan into a utility connection, transformer and bus arrangement that remains workable through faults, maintenance and phased expansion. The utility does not see a rack count. It sees real and reactive power, a changing load profile, power-quality behavior, fault contribution and a requested energization schedule.

That distinction matters. A 60 MW IT plan is not automatically a 60 MVA connection, and an installed transformer nameplate is not the same as capacity available to the live halls. The design has to reconcile five quantities: IT load, facility peak demand, apparent power at the point of connection, firm utility capacity and the capacity that is actually energized.

This guide is for developers, energy and site teams, electrical engineers and procurement leaders deciding whether a site can support an AI data center—and how to phase the electrical infrastructure without hiding risk inside a headline megawatt figure.

TL;DR: AI data center substation design

• Convert the IT growth plan into design-day facility MW and connection MVA before selecting transformers or voltage.
• Keep contracted, installed, energized and future capacity separate. Only contracted and energized capacity is available to the live project.
• Choose transformer blocks with one-unit-out operation, fault level, transport, replacement and phased commissioning in view.
• Model AI loads as dynamic power-electronic loads; the utility may need ride-through, recovery and load-step information.
• Reserve land, routes, bays, protection I/O and isolatable interfaces for each phase before the site layout hardens.

Start with the capacity ledger

Before selecting a voltage or transformer size, define the capacity terms used by the project. They should appear in the utility application, single-line diagram, commercial model and commissioning plan with the same meanings.

Installed capacity minus peak demand is installed headroom over peak demand. Installed capacity minus energized capacity is installed but not yet energized capacity. Calling either figure “stranded MVA” without explaining the boundary obscures the decision the number is meant to support.

Convert IT load into a design-day utility case

An annual PUE target can support energy planning, but it is not enough to size a substation. Electrical equipment has to carry a coincident condition: a hot design day, a maintenance state, a cooling restart or a battery-recharge case. The design should build that condition from a load schedule and selected-equipment curves.

A preliminary calculation can be written as:

Facility peak MW = IT MW + design-day cooling MW + conversion losses + other auxiliaries

Connection MVA = facility peak MW / power factor at the point of connection

The example below is illustrative. It is not a ModulEdge product rating or a substitute for a load flow.

The spread is almost 12 MVA before any resilience margin or future phase is added. A PUE-derived 75 MW result would also understate the 84 MW design case by 9 MW. The correct factors depend on climate, cooling architecture, voltage conversions, operating states and the actual rack schedule. The purpose of the sensitivity is to show which assumptions move the connection request.

Connection voltage is a utility decision, not a universal MW rule

There is no worldwide table that assigns one connection voltage to every data center above a given load. Available network voltage, substation topology, local planning rules, protection practice, reliability, land, metering and the network operator's expansion plan all affect the answer.

Published utility requirements illustrate the variation. CenterPoint Energy's customer substation design specification addresses customer-owned substations connected to its transmission system. Dominion Energy's facility interconnection requirements define utility-specific design, protection, metering and operating interfaces. Neither document creates a universal design rule for another territory.

A feasibility request should ask the network operator for:

• firm and non-firm capacity by date;
• proposed point of connection and available voltage;
• maximum and minimum fault level, including X/R ratio;
• earthing and neutral requirements;
• reactive-power, harmonic, voltage-change and flicker limits;
• protection, communications, metering and control interfaces;
• dynamic-model and study requirements for the proposed load; and
• works, land rights, permits, securities and milestones on both sides of the boundary.

The answers can change the site plan. A high-voltage yard, cable sealing ends, fire separation, transformer transport route, harmonic equipment and future bays all consume land that cannot be recovered easily after the campus layout is fixed. The companion data center site-selection guide covers the broader land, fiber, water and permitting decision.

If the network cannot deliver the required capacity or date, on-site generation may become part of the concept rather than a later add-on. The on-site power guide discusses that strategic choice. Its sources, fuel, emissions, controls and islanding behavior still have to be included in the substation studies.

Select transformer blocks around demand, fault level and replacement strategy

Transformer count is not set by demand alone. Larger units can reduce equipment count, while smaller blocks can limit the load exposed to one outage, simplify phased energization and reduce fault contribution per section. Transport constraints, delivery program, spares and maintenance access also matter.

For the 88.4 MVA illustrative case, several nameplate arrangements are possible:

Nameplate arithmetic is only a screen. It does not prove overload capability, voltage regulation, harmonic performance, protection coordination, cable capacity or resilience of the connected bus. The project must also decide whether one-unit-out operation carries the full design case, a controlled reduced load, or only committed halls.

Tier classifications apply to the facility topology and operating outcomes, not to a transformer in isolation. A transformer arrangement should therefore be described by the failure and maintenance cases it supports rather than labelled “Tier III equipment.”

Fault level can set the bus architecture

Adding transformer capacity in parallel can raise prospective short-circuit current faster than the load requires. A first-pass transformer contribution can be estimated from:

Isc = transformer MVA / (√3 × voltage kV × per-unit impedance)

For a 40 MVA transformer at 11 kV with 14% impedance, that simplified contribution is about 15 kA. Four identical units connected to one ideal common bus could contribute about 60 kA before utility, generator, motor and converter effects are considered. Splitting the arrangement into two sections with two transformers per section reduces the simple transformer contribution to about 30 kA per section when the tie is open.

This is not a short-circuit study. It ignores source impedance, conductor impedance, X/R, equipment tolerances and the time-dependent behavior of other sources. It does show why normally open ties, smaller blocks, transformer impedance and switchgear rating have to be considered together.

The follow-on study should test every permitted switching state, including temporary states used during maintenance. It should establish making and breaking duties, short-time withstand, protection selectivity, arc-energy exposure and whether closing a tie creates an unacceptable fault level. The companion data center switchgear guide explains the assembly and procurement decisions downstream of this substation concept.

AI loads add dynamic behavior to the connection study

Large computing loads are power-electronic loads with fast controls. Their response to voltage and frequency disturbances depends on UPS settings, power-supply protection, ride-through logic and recovery sequences. A static peak-MW figure cannot describe that behavior. Our AI data center load-profile guide explains which time-series traces and control assumptions to request before the connection study.

NERC's review of a 2024 large-load-loss event documents the loss of approximately 1,500 MW of data center load following a transmission fault and recommends better modeling, performance requirements and coordination for large loads. It should not be read as proof that every AI facility will behave the same way. It does establish that aggregate load behavior can be material to bulk-system reliability.

The project should provide the network operator with models and settings at the level it requests. At minimum, align these teams before energization:

• utility planning and protection;
• owner's electrical engineer;
• UPS and power-supply vendors;
• cooling controls and building-management teams; and
• commissioning authority and operations.

They need one agreed sequence for disturbance ride-through, load shedding, restart, cooling recovery and battery recharge. Otherwise, locally reasonable controls can combine into an unexpected site-wide step.

Separate the factory-built scope from the site-specific scope

Prefabrication can move assembly, internal wiring, routine verification and simulated control tests into a controlled factory. It does not turn the utility interface, earth grid or final protection settings into standard products.

IEC 62271-202:2022 covers high-voltage/low-voltage prefabricated substations for alternating current voltages above 1 kV up to and including 52 kV. Applicable standards, certification and utility acceptance still depend on the project jurisdiction and equipment scope.

A responsibility matrix should name who owns the utility studies, civil design, earthing, protection model, settings, relay injection, transformer tests, cable tests, site acceptance and integrated commissioning. “By others” is not an interface design.

Phase capacity in blocks that can be isolated and commissioned

A credible phased plan connects commercial milestones to physical assets. It shows how each data hall obtains power, which common systems are required from day one, and what must be interrupted to add the next phase.

The figures are illustrative. The discipline is the point: planned, installed, contracted and live capacity should never be collapsed into one number.

Good phasing also preserves physical options. Reserve the transformer transport path, bay position, cable route, protection I/O, control-system points and commissioning boundary. A spare outline on a drawing is not useful if the future unit cannot be delivered or connected without shutting down the operating campus.

Substation concept checklist

• Is the IT load boundary stated, with a rack or hall schedule by phase?
• Is design-day facility demand calculated separately from annual PUE?
• Are real power, power factor and apparent power visible?
• Has the network operator confirmed voltage, capacity dates and fault data?
• Does the transformer arrangement meet the stated one-unit-out operating case?
• Have all normal and maintenance bus states been checked for fault duty?
• Are reactive power, harmonics and dynamic load behavior included in the study scope?
• Are land, fire, acoustic, oil-containment and transport constraints reserved?
• Does each future phase have an isolatable connection and commissioning plan?
• Does the responsibility matrix name the owner of every utility, protection and testing interface?

Where ModulEdge fits

ModulEdge develops factory-built modular data center platforms and associated MV/LV electrical assemblies. Its modular-substation scope can include incoming or transit medium-voltage arrangements up to 24 kV and 0.4 kV output, with switchboards and control assemblies manufactured by ModulEdge. Transformers and other specialist equipment are selected from qualified OEMs and integrated for the project.

That scope is not a blanket claim to deliver a utility high-voltage substation or secure an interconnection. Connection voltage, transformer selection, protection scheme, civil works, local conformity and commissioning responsibility have to be defined for the site.

For an initial engineering review, send ModulEdge the site, utility correspondence, IT load by phase, target rack schedule, redundancy case and available fault data. The useful first output is a capacity model, one-line concept and interface register—not a transformer count chosen before the inputs are known.

Frequently asked questions

At what size does a data center need its own substation?

There is no universal megawatt threshold. The answer depends on available utility voltage and capacity, network policy, distance to the connection point, reliability, land, protection and the required energization date. A smaller project may need a dedicated substation where the local network is constrained, while a larger one may connect through an existing campus arrangement.

How do you convert AI IT load into substation MVA?

Add design-day cooling, conversion losses and other coincident auxiliaries to the IT load to obtain facility real power. Divide that result by the power factor at the point of connection to obtain apparent power. Then test maintenance, cooling restart, battery recharge, growth and other governing cases before proposing transformer blocks.

How many transformers should an AI data center have?

Transformer count follows the required capacity after one unit is unavailable, bus arrangement, fault level, phase size, transport limits, replacement strategy and maintenance plan. A nameplate sum alone is insufficient. The design must show what load remains supportable in each failure and maintenance state and whether controlled load reduction is acceptable.

Should an AI data center connect at 11 kV, 33 kV or transmission voltage?

The network operator normally determines which voltages and points of connection are feasible. Project demand, local network strength, fault level, land, route length, protection and utility expansion plans all influence the offer. No global rule assigns one voltage solely from the data center's IT megawatts.

Why do AI loads need dynamic grid studies?

AI facilities contain large numbers of power-electronic loads whose ride-through, protection and restart controls can create rapid aggregate changes. Utilities may therefore request dynamic models, recovery sequences, harmonic information and reactive-power behavior in addition to steady-state MW and MVA. The required model and acceptance criteria are utility-specific.

Can a modular substation be fully tested in the factory?

Factory testing can verify the supplied assemblies, internal wiring, interlocks, controls and simulated sequences. It cannot prove the installed earth grid, utility protection interface, field cable joints, final settings or complete site response. Site acceptance and integrated commissioning remain necessary after connection.