Low-Voltage DC (LVDC) Power Distribution for Data Centers: Architectures, Protection and a Practical Path from AC to DC

November 11, 2026 | Blog

Part A — LVDC for Data Centers at a Glance

Data center power systems have been built on alternating current for decades. Power arrives from the utility as AC, passes through transformers and an AC uninterruptible power supply (UPS), and reaches each server, where a power supply finally converts it to the direct current (DC) that every processor actually uses. As rack power climbs from tens of kilowatts toward hundreds of kilowatts, the cost of all those conversion stages becomes harder to ignore: more heat, more copper, more floor space and more components that can fail.

Low-voltage DC (LVDC) distribution moves the AC-to-DC conversion upstream, to the facility or the row, and distributes DC to the racks. The idea is not new: telecom central offices have run on −48 V DC for a century, and 380 V DC data center demonstrations in the United States date back almost twenty years. What is new is the scale. AI and accelerated-computing racks now need so much power that the industry is working on common, vendor-neutral frameworks so that LVDC can be designed, protected, listed, operated and maintained the same way across sites and suppliers.

This guide is Keentel Engineering's engineering view of that framework. It explains the voltage classes in use, the architectures on the path from AC to DC, how DC systems are grounded and protected, what "interoperable" has to mean in practice, how LVDC is operated safely, and how to decide whether and how far a US facility should move to DC.

In plain terms: LVDC means doing the AC-to-DC conversion once, in a few large efficient converters, instead of thousands of times inside individual servers. It can save a few percent of total energy, free space at the rack and simplify storage integration but only if protection, grounding, codes and operating practices are designed for DC from the start.

Scope and companion guides

This guide covers the facility-wide framework: voltage classes, architectures, grounding, protection philosophy, interoperability, operations and migration. Four companion Keentel guides go deeper on specific topics and are referenced rather than repeated here:

• 800 VDC rack fault protection — solid-state circuit breakers (SSCBs) and fault clearing at the rack.
• 800 VDC AI data centers and the UPS — energy storage sizing, GPU load transients, generator coordination and DC arc flash.
• 800 VDC chillers, CDUs and racks — the mechanical–electrical interface and powering cooling equipment from DC.
• Inside the data center rack — rack power, redundancy and cabling.

The illustrative facility used in this guide

The path from AC to DC

What Keentel's analysis shows

Keentel modeled five architectures delivering 20 MW to the 54 V rack busbar, using representative stage efficiencies for modern equipment (assumptions listed in Part B, Section 6).

Energy priced at $70/MWh. Every megawatt of loss also has to be removed by the cooling plant, which typically adds another 15–30% to the savings.

Figure 1. Left: Keentel's stage-by-stage loss comparison for 20 MW delivered at the 54 V rack busbar (illustrative stage efficiencies). Right: the layered protection philosophy for an LVDC system, with the fastest devices nearest the load and time-delayed devices at the source.

Five lessons

• The efficiency gain is real but depends on the baseline. Against a double-conversion AC UPS, LVDC reduces input energy by about 5.0–6.2% in this model. Against an AC UPS running in eco-mode, the gap shrinks to about 1.5–2.7%. Claims of "much higher efficiency" need a stated baseline.
• Voltage matters more than "DC" itself. At equal copper, a 380 Vdc two-wire feeder has higher conductor losses than a 415 V three-phase AC feeder. An 800 Vdc feeder has about one-third of the AC losses. The density benefit comes from higher voltage and fewer conversion boxes, not from DC as such.
• Protection is the hardest part. DC has no current zero to help a breaker interrupt, batteries and capacitors can deliver very fast, very large fault currents, and ground faults are harder to find. LVDC needs a defined protection hierarchy from source to rack.
• Grounding must be chosen deliberately. Midpoint high-resistance grounding, solidly grounded and ungrounded systems with insulation monitoring each change fault behavior, touch voltage and equipment ratings.
• Codes and listings set the pace. In the United States, LVDC "at scale" depends on equipment that is listed for the DC voltage and fault duty, on NEC requirements for DC systems and storage, and on NFPA 70E work practices for DC. Plan the listing path before the single-line diagram.

Engineering review of common LVDC claims

Part B — The LVDC Framework: An Engineering Guide

1. What "low-voltage DC" means in a data center

1.1 Voltage classes in use



"LVDC" is used loosely in the industry to describe any DC distribution below the medium-voltage range. In practice, four voltage classes appear in data center designs, each with a different purpose and a different set of codes, components and safety practices.

A key point for US projects: the National Electrical Code (NEC) does not use the international "low voltage directive" definition (up to 1,500 Vdc). Since the 2023 edition, the NEC generally draws its main threshold at 1,000 V, with separate articles and equipment requirements above that. A DC system at or below 1,000 V is still subject to the NEC's general wiring rules, its articles for DC microgrids, storage and batteries, and the listing requirements for every piece of equipment.

1.2 Voltage quantities that must be specified



"800 V" or "380 V" alone is not a specification. A complete LVDC voltage definition includes:

• Nominal voltage and whether it is measured pole-to-pole or pole-to-ground. A bipolar ±400 V system is 800 V pole-to-pole and 400 V pole-to-ground.
• Normal operating band — the steady-state range the bus must hold, for example 760–820 V or 360–400 V.
• Battery float, equalize and end-of-discharge voltages, if storage is on the bus. These often define the extremes of the operating band.
• Transient limits — the overvoltage and undervoltage levels and durations that loads must ride through without tripping.
• Insulation coordination — the maximum continuous voltage to ground, the impulse withstand of components, and surge protection levels.

Practical guidance: write the voltage specification as a table with nominal, steady-state band, battery-defined extremes, ride-through envelope and maximum voltage to ground. Every rectifier, converter, breaker, fuse, busway and connector must be rated against that table, not against the nominal value.

1.3 Why the conversation has shifted to higher DC voltages

At 54 V, a 200 kW rack draws about 3,700 A. At 800 V the same rack draws 250 A. Rack busbars, cords, connectors and the conversion equipment inside the rack all scale with current. That is why recent industry work has focused on 800 V-class rack inputs, while 380 Vdc remains attractive at the facility and row level where its components are more mature.

2. Why move from AC to DC at all

2.1 The conversion chain today

In a conventional US facility, power passes through these stages before it reaches a processor:

1. Utility transformer (for example 230 kV to 34.5 kV) and a facility MV/LV transformer (34.5 kV to 480Y/277 V).

2. A double-conversion UPS: rectifier (AC to DC), battery bus, inverter (DC back to AC).

3. A PDU transformer (480 V to 415Y/240 V), if racks use 415/240 V.

4. The rack or server power supply: AC to DC, with power-factor correction, then DC/DC to 54 V or 12 V.

5. Board-level voltage regulators (12 V or 54 V down to below 1 V for processors).

Stages 2 and 4 each convert AC to DC. Stage 2 converts DC back to AC in between. An LVDC architecture removes the inverter in stage 2, removes the PDU transformer, and replaces many small AC power supplies with fewer, larger rectifiers.

2.2 What LVDC can deliver

• Efficiency — fewer conversion stages and larger, more efficient converters.
• Space — removing the AC power supplies from racks frees space for compute; removing PDU transformers frees white space.
• Storage integration — batteries connect to the DC bus without an inverter, and fast storage can buffer large load swings.
• Simpler harmonics management — the facility rectifier becomes the single point where harmonic performance at the utility interface is controlled (IEEE 519).
• A path to on-site DC sources — solar, fuel cells and storage can feed the DC bus directly.

2.3 What LVDC does not deliver automatically

• It does not reduce the heat generated by the IT equipment itself, which is most of the load.
• It does not remove the need for redundancy, generators or utility-grade protection.
• It does not, by itself, reduce copper: that depends on voltage level and conductor count (Worked example 1).
• It does not simplify protection — it changes it.

Worked example 1 — Current and copper for a 500 kW row

A 500 kW row of racks is fed three ways: 415 V three-phase AC at 0.98 power factor, a 380 Vdc two-wire bus, and an 800 Vdc two-wire bus.

For a fixed total copper cross-section shared equally among n current-carrying conductors, each carrying current I, the conductor loss is proportional to n²I²:

Interpretation. A 380 Vdc bus needs about 50% more copper than 415 V three-phase AC to have the same losses the opposite of what many presentations imply. Three-phase AC is very efficient at delivering power through copper. An 800 Vdc bus needs about one-third of the copper of 415 V AC for the same losses. The copper and density benefit of LVDC comes from moving to higher DC voltage.

3. Architectures on the path from AC to DC

3.1 Architecture A — AC with double-conversion UPS (today's baseline)

MV transformer → 480 V switchboard → double-conversion UPS → PDU → 415/240 V rack → rack power shelf (AC to 54 V). This is the most widely deployed design. It isolates IT from utility disturbances well, but each kilowatt passes through two full conversions plus the rack power supply.

3.2 Architecture B — AC with eco-mode or multi-mode UPS

The UPS normally passes utility power through a bypass path and transfers to inverter operation only when the supply is disturbed. UPS efficiency rises from about 95–96% to about 99%. This is the fairest efficiency baseline when evaluating LVDC, because it is what an efficiency-focused AC design would actually use. Its trade-offs are a short transfer on disturbance and less filtering of utility events; IT power supplies and rack storage must ride through.

3.3 Architecture C — Hybrid: AC to the row, DC at the rack

AC distribution continues to the row. A row- or rack-level "sidecar" rectifier converts 415 V or 480 V AC to an 800 V (or 400 V) DC rack bus, and DC/DC converters in the rack produce 54 V. Battery backup units may sit on the rack DC bus. This architecture is the usual transition step: it changes the rack without rebuilding the facility, and it can be deployed hall by hall.

3.4 Architecture D — 380 Vdc facility distribution

A central rectifier plant (N+1 modules) converts 480 V AC to a 380 Vdc bus, with batteries connected directly to the bus. DC switchgear and busway distribute to rows, and rack DC/DC converters produce 54 V. This architecture removes the UPS inverter and all AC power supplies and has the longest operating history among facility DC designs.

3.5 Architecture E — 800 Vdc facility distribution

Similar to Architecture D but at ±400 V. Current is halved, so busway and cable sizes fall and the rack can accept several hundred kilowatts. The rectifier may be fed from 480 V AC or, in emerging designs, from a medium-voltage solid-state converter connected directly at 13.8 kV or 34.5 kV. Medium-voltage solid-state conversion removes the 60 Hz MV/LV transformer, but it is a newer technology whose reliability, protection and listing must be evaluated carefully.

3.6 Comparing the architectures

3.7 Where the storage sits

• Central storage on the facility DC bus (Architectures D and E): fewest conversion stages, but the full battery fault current is available on the bus.
• Distributed storage in each rack (battery backup units, Architecture C): the fault energy at any point is smaller and storage scales with the racks, but there are many more battery systems to monitor and maintain.
• Both: a central plant for ride-through to generator start, plus rack-level fast storage to smooth workload power swings.

4. Grounding the DC system

4.1 Why grounding is a first decision

Grounding decides what voltage appears between any conductor and earth, what happens at the first ground fault, how large the fault current is, how faults are detected, and which equipment ratings apply. It must be fixed early because it affects every component.



4.2 The main options

For facility-scale data center DC, midpoint high-resistance grounding with continuous insulation and ground-fault monitoring is the most common recommendation, because it limits the voltage to ground and lets the system ride through a single ground fault while it is located. Whatever is chosen, the scheme must match the listings of the equipment, the requirements of the NEC for DC systems, and the assumptions in the arc-flash study.

4.3 Ground-fault detection on DC

AC residual-current devices rely on the alternating current to induce a signal; they do not detect smooth DC residual current. DC systems use one or more of the following:

• Insulation monitoring devices (IMDs) that inject a low-frequency test signal and measure insulation resistance between the system and ground.
• Midpoint-resistor current measurement in high-resistance grounded systems, with the polarity of the current indicating which pole is faulted.
• DC-capable residual current sensors at feeders, to locate the faulted circuit.
• Pulse-injection location systems that trace a test signal to the faulted branch.

Worked example 2 — Insulation monitoring thresholds

A common practice is to set the insulation alarm at about 100 Ω per volt of system voltage, with an earlier warning at about 500 Ω/V. These are typical values; follow the equipment manufacturer and the applicable product standard.

Interpretation. A large DC system has many parallel paths to ground cables, busway, converters, EMI filter capacitors — so its normal insulation resistance falls as the system grows. A threshold set for one rack can nuisance-alarm for a whole hall. Insulation monitoring should therefore be zoned (per bus section or per row), and EMI-filter capacitance to ground should be limited by specification so that monitoring remains effective.

5. Holding the DC bus voltage

5.1 Regulation and droop sharing

Several rectifier modules usually feed one DC bus. To share load without fighting each other, each module's output voltage drops slightly as its current rises a "droop" characteristic. Load then divides according to how closely the modules' setpoints match.

Worked example 3 — Droop sharing between two rectifiers

Two 500 kW rectifiers feed a 380 V bus carrying 1,316 A in total. Their setpoints differ by 0.5 V (0.13%) because of calibration tolerance.

With 2% droop at rated current (1,316 A per rectifier):

Interpretation. More droop gives better sharing but a larger voltage swing with load. Tight setpoint calibration, digital current sharing or a master–slave control can reduce imbalance without widening the voltage band. The bus voltage band, droop and sharing method must be specified together, because the rack DC/DC converters must accept the full resulting voltage range.

5.2 Load steps and stability

Accelerated-computing workloads can change rack power by tens of percent within milliseconds. The DC bus must hold voltage through those steps using converter control bandwidth, bus capacitance and fast storage. Large numbers of constant-power DC/DC converters present a negative incremental resistance to the bus, which can cause oscillation if the source impedance and control are not designed together. A DC bus stability study impedance-based or EMT simulation — belongs in every LVDC design. Keentel's 800 VDC UPS guide gives a worked capacitance example for a rack load step.

6. Efficiency: counting every stage

6.1 The method



End-to-end efficiency is the product of the efficiency of each stage in series. Small differences in each stage multiply.

6.2 Stage assumptions used in this guide



These are representative values for modern equipment near rated load. Real values vary with load level, temperature, input voltage and product; use manufacturer curves at the expected operating point.

Worked example 4 — Annual savings for a 20 MW IT hall

For 20 MW delivered at the 54 V rack busbar, operating continuously:

Interpretation. Against an efficiency-optimized AC design, LVDC is worth a few hundred thousand dollars a year per 20 MW hall, not millions. The decision therefore rests as much on density, storage integration and future rack power as on efficiency. The figures also show that a well-applied eco-mode AC UPS captures more than half of the gain available from DC.



6.3 Part-load behavior

Most data center equipment runs well below its rating because of redundancy: in a 2N design, each path carries at most half its rated load. Converter efficiency curves typically peak at 40–70% load and fall at low load. A design that looks efficient at rated load may lose much of its advantage at 25–35% load. Compare architectures at the expected operating point, including redundancy.

7. Protecting an LVDC system

7.1 What makes DC protection different

• No natural current zero. An AC arc extinguishes at each current zero (120 times per second at 60 Hz); a DC arc does not. DC breakers must build an arc voltage higher than the system voltage, or use semiconductors to force the current to zero. Breakers, fuses and switches must carry DC ratings at the actual system voltage and time constant.
• Two very different fault sources. Rectifiers are current-limited — typically to 1.1–1.5 times rated current within milliseconds. Batteries, supercapacitors and DC-link capacitors can deliver very large currents with very fast rates of rise.
• Fast rate of rise. With little inductance, DC fault current can reach damaging levels in microseconds near the source of stored energy, before mechanical breakers can move.
• Selectivity is harder. Time-current coordination between rectifier-fed and battery-fed faults behaves very differently, and protective devices must coordinate across both.

Worked example 5 — Battery-fed versus rectifier-fed fault current

A 380 V lithium-ion battery string with total internal resistance 30 mΩ is connected to the bus, with 5 mΩ of cable and bus resistance to the fault. A 500 kW rectifier feeds the same bus, current-limited to 1.5 times rated.

With 10 µH of loop inductance, the initial rate of rise and the circuit time constant are:

Interpretation. The battery delivers more than five times the rectifier's limited current, reaching most of it within about a millisecond. A protective device sized and coordinated for rectifier-fed faults may never see enough current from the rectifier alone to trip quickly and may be unable to interrupt the battery's contribution. Every protective device must be rated for the maximum prospective current (battery plus capacitors plus rectifiers) at the actual time constant, and must also trip reliably for the minimum (rectifier-only, battery disconnected) case.

7.2 A layered protection hierarchy

A practical LVDC framework defines protection levels from the source to the rack, each with a clear job and coordination rules.

The design principle is simple: the device closest to the fault clears it first and fastest, and every upstream device waits. Because DC fault current can rise so quickly, the lower levels increasingly use solid-state or hybrid devices that act in microseconds, while upstream mechanical devices provide time-delayed backup.

7.3 Selectivity tools for DC



• Current magnitude — useful where fault current falls with distance, less useful on short, stiff buses.
• Time delay — upstream devices wait for downstream ones; limited by the energy that can flow while waiting.
• Rate of rise (di/dt) — fast-acting devices can detect a fault by its rate of rise before current becomes large.
• Direction — on buses with storage at several points, directional detection identifies which side the fault is on.
• Communication — zone-selective interlocking or fast communication between devices lets the device nearest the fault act, while others block.
• Current limiting — fuses and solid-state devices that limit let-through make upstream coordination easier.

7.4 Precharge and inrush

Connecting a discharged DC-link capacitor to a live bus draws a very large inrush current. Every converter and every bus section with significant capacitance needs a precharge path — a resistor and bypass contactor, or an electronically controlled soft-start — and the protection must distinguish precharge inrush from a fault. Interoperability rules (Section 9) should define precharge behavior for every connectable device.

7.5 Arc flash on DC

The IEEE 1584 arc-flash model applies to three-phase AC systems only. DC incident energy is estimated with the methods in NFPA 70E Informative Annex D, such as the maximum-power method and arc-resistance-based models. DC arcs can be sustained at lower voltages and longer gaps than many engineers expect, and battery-fed faults can last until the battery is disconnected. Keentel's 800 VDC UPS guide includes a worked DC incident-energy example.

8. Reliability and availability

8.1 Fewer stages, fewer failures

For components in series, availability is the product of component availabilities. Each component's steady-state availability depends on its mean time between failures (MTBF) and mean time to repair (MTTR):

Worked example 6 — Single-path availability, AC versus DC

Illustrative values for a single, non-redundant path (real designs are redundant):

Removing the PDU transformer from the path cuts the expected unavailability of a single path by about a quarter. With N+1 redundancy for example seven rectifier modules where six carry the load the unavailability attributable to rectifier modules drops to a few seconds per year, and the transformer and distribution elements dominate.

Interpretation. Fewer series stages help, but redundancy design matters far more than AC versus DC. An LVDC plant must still be designed to the facility's redundancy target (N+1, 2N, distributed redundant), and the DC switchgear, busway and protection become the new single points of failure to manage.

9. Interoperability: what must be common

An LVDC framework only works at industry scale if equipment from different suppliers can be connected, protected and operated together. The areas that must be specified include:

Practical guidance: put these items into the purchase specification for every DC device — rectifier, busway, breaker, rack power shelf and battery and require type-test evidence. Interoperability is proved at the interface, not assumed from a nameplate.

10. Integrating storage, on-site sources and cooling

10.1 Energy storage on the DC bus

Connecting batteries directly to the DC bus removes the UPS inverter and gives the fastest possible response. The design questions are:

• Voltage window. The battery's voltage swings from float to end of discharge. Either the bus accepts that whole range, or a DC/DC converter regulates between battery and bus (adding a stage but decoupling the two).
• Fault contribution. A directly connected battery sets the maximum fault current on the bus (Worked example 5). String fuses and fast disconnects are essential.
• Listing and siting. Stationary lithium-ion systems in the United States are listed to UL 9540, with large-scale fire testing to UL 9540A, and installed under NFPA 855 and NEC Article 706. These requirements apply whether the battery sits in a central battery room or inside a rack.
• Function. Storage may provide ride-through to generator start, smoothing of fast workload swings, or both. These functions have very different power, energy and cycle-life requirements; Keentel's 800 VDC UPS guide sizes each.

10.2 On-site DC sources

Solar, fuel cells and some storage technologies are natively DC. On a DC bus they can connect through a DC/DC converter instead of an inverter and a rectifier. The benefits are real but bounded: on-site solar is usually small relative to data center demand, and fuel cells need fuel supply, permits and their own protection. Any source that can export to the utility, or that is paralleled with the utility through the facility rectifiers, must be evaluated under the utility's interconnection requirements (IEEE 1547 for distribution-connected resources, or transmission-level requirements for larger systems).

10.3 Cooling equipment

Pumps, fans and coolant distribution units (CDUs) must stay running through utility events, or the IT equipment will throttle within seconds. In an LVDC facility they can be fed from the DC bus through drives listed for DC input, or kept on AC with their own ride-through. Chillers and dry coolers usually remain on AC, backed by generators and thermal storage. Keentel's 800 VDC chillers, CDUs and racks guide covers drive DC-link voltages, precharge and discharge, and thermal ride-through.

11. Operations and safety

11.1 Monitoring

An LVDC system should be monitored at every protection level:

• Bus voltage, current and power at sources, buses and feeders.
• Rectifier module status, sharing and temperature.
• Battery state of charge, state of health, cell voltages and temperatures through the battery management system.
• Insulation resistance and ground-fault location by zone.
• Protective-device status and trip events, time-stamped for sequence-of-events analysis.
• Power-quality data at the AC interface (harmonics, voltage events).

11.2 Working safely on DC

• NFPA 70E applies to DC systems. It includes DC shock-protection boundaries and a DC arc-flash PPE category table, as well as Annex D methods for DC incident energy. An electrical safety program must address DC explicitly.
• Stored energy. Capacitors in converters and on the bus can stay charged after the source is removed. Specify maximum discharge times and require verification of absence of voltage before work. Batteries cannot be "switched off" — only disconnected — and their terminals remain live.
• Lockout/tagout procedures must identify every DC source: rectifiers, batteries, rack-level storage and any backfeed through converters.
• Test instruments must be rated for DC at the system voltage and the appropriate measurement category.
• Labeling. DC equipment needs labels for voltage, polarity, available fault current, arc-flash incident energy and stored-energy hazards.
• Training. Most data center technicians have years of AC experience and little DC experience. Training must cover DC arc behavior, polarity, grounding schemes and insulation-monitoring alarms.

11.3 Maintenance

• Thermographic inspection of DC busway joints and connections, which can loosen and overheat like AC joints.
• Functional testing of DC breakers and solid-state protection, including trip settings and communication-based schemes.
• Battery maintenance per the manufacturer, NFPA 855 and, where applicable, IEEE battery maintenance guides.
• Periodic verification of insulation-monitoring alarms and ground-fault location systems.
• Firmware management for rectifiers, converters and solid-state breakers, with change control.

12. Codes, standards and listing in the United States

Equipment for an LVDC facility must be listed for its application. The key US documents are:

Two practical points matter on every project:



• Listing gaps. A device listed for AC, or for DC at a lower voltage or lower fault current, is not listed for the LVDC application. For 800 V-class distribution with large battery fault currents, verify DC voltage rating, interrupting rating, time constant and listing for every protective device. Where listed equipment does not exist, field evaluation by a nationally recognized testing laboratory and early engagement with the authority having jurisdiction (AHJ) are needed.
• Code editions. NEC requirements for DC systems have changed in recent editions and will continue to change. Confirm which edition the AHJ has adopted and check for local amendments.

International standards such as the IEC voltage definitions, telecom DC interface standards and IEC UPS standards are useful references for terminology and interface design, but US installations are governed by the documents above.

13. Cost and adoption barriers

Capital cost comparisons vary widely by site. LVDC removes UPS inverters, PDU transformers and server power supplies, but adds DC switchgear, busway, protection and monitoring. In most current projects the business case is driven by rack power density and storage integration, with efficiency as a secondary benefit.

14. A practical migration roadmap



14.1 Stages

14.2 Decision matrix

15. LVDC design checklist



• Voltage specification table: nominal, band, battery extremes, ride-through envelope, maximum voltage to ground.
• Grounding scheme selected and coordinated with equipment listings, insulation monitoring and the arc-flash study.
• Fault-current study covering battery, capacitor and rectifier contributions; maximum and minimum fault cases.
• Protection hierarchy defined for all five levels, with selectivity method for each.
• Every protective device verified for DC voltage, interrupting rating, time constant and listing.
• Precharge and inrush limits defined for all connectable equipment.
• DC bus stability study for converter interaction and load steps.
• Efficiency comparison at the expected operating point, including redundancy and part load.
• Storage sized for each function (ride-through, smoothing) and listed to UL 9540 / 9540A.
• Utility interface study (IEEE 519 harmonics, ride-through and reconnection), plus monitoring, communications and time synchronization.
• DC safety program (NFPA 70E DC requirements, lockout/tagout for all sources, stored-energy discharge, training), with the AHJ, insurer and utility engaged at concept stage.
  

Part C — Three Anonymous Case Studies

These composite case studies reflect the kinds of problems Keentel Engineering sees on US data center projects. Names, locations and details have been changed or combined.

Case study 1 — An existing hall that needed 150 kW racks

Situation

An operator in the Mid-Atlantic had a 12 MW data hall built around 480 V switchgear, double-conversion AC UPS and 415/240 V rack distribution, designed for about 15 kW per rack. A tenant wanted to install accelerated-computing racks at about 150 kW each in part of the hall. The operator's first idea was to convert the whole hall to a 380 Vdc facility bus.

Analysis

Keentel reviewed the existing switchgear, UPS, busway, cooling and structure. Converting the whole hall to facility DC would have required new DC switchgear, new busway, a new battery system and an outage the tenant could not accept. Keentel's efficiency model showed that the operator could capture about 3.3 points of efficiency simply by enabling the multi-mode function on its existing UPS systems, after ride-through testing with the new rack power equipment. The new racks themselves needed a higher-voltage rack bus to keep busbar current manageable.

Engineered solution

The hall kept its AC infrastructure. The UPS systems were reconfigured for multi-mode operation after transfer testing. The high-density rows received hybrid power: 415 V AC busway to each row, sidecar rectifiers producing an 800 V rack bus, and rack-level battery backup units for ride-through and load smoothing. Keentel specified the rack bus grounding (midpoint high-resistance), insulation monitoring per row, precharge limits for each rack, and solid-state rack-input protection coordinated with the sidecar rectifier current limit and the upstream AC breakers.

Outcome

The high-density rows were energized in phases without a hall-wide outage. The operator avoided a full DC conversion, captured most of the available efficiency gain through UPS settings, and established a rack-level DC standard it could repeat in other halls.

Case study 2 — A battery that out-muscled the protection

Situation

A developer in the Southwest designed a new hall with a 380 Vdc distribution bus, central rectifiers and a lithium-ion battery plant connected directly to the bus. The protective devices on the DC feeders had been selected using the rectifier current limit as the fault level, because the design team assumed the rectifiers would "limit the fault."

Analysis

Keentel performed a DC fault study that included the battery strings, the DC-link capacitance of the rack converters and the bus inductance. With the battery connected, the prospective fault current at the main bus was more than five times the rectifier-limited current, with a rate of rise of tens of amperes per microsecond. Several feeder devices were not rated to interrupt that current at the system time constant, and the main bus devices would not have coordinated with the feeders for battery-fed faults. In the opposite case battery disconnected for maintenance some downstream devices would not have tripped quickly on rectifier current alone.

Engineered solution

The design was revised with current-limiting DC fuses on each battery string, a fast battery disconnect, feeder devices rated for the full prospective current, and solid-state devices at the busway taps to limit let-through. The protection settings were coordinated for both the maximum (battery connected) and minimum (rectifier only) cases, and the arc-flash study was redone with DC methods.

Outcome

The revised design passed the authority having jurisdiction's review with listed devices at every protection level. Arc-flash labels were reduced at the feeder level because the faster branch devices limited fault duration.

Case study 3 — Insulation alarms that would not stop

Situation

A campus operator in the Midwest commissioned an ungrounded DC distribution system with a single insulation monitoring device for an entire data hall. During commissioning, the insulation alarm triggered repeatedly as racks were added, but no ground fault could be found.

Analysis

Keentel measured insulation resistance and capacitance to ground section by section. Each rack's DC/DC converters had EMI-filter capacitors connected to the chassis, and the long busway runs added capacitance and leakage. As more racks came online, the total insulation resistance of the hall dropped below the alarm threshold, and the large capacitance to ground also disturbed the monitoring device's measurement. There was no fault the system had simply outgrown a single monitoring zone.

Engineered solution

The hall was divided into insulation-monitoring zones by row, each with its own device and DC-capable residual current sensors on the feeders for fault location. A limit on capacitance and leakage to ground was added to the rack power equipment specification, and alarm thresholds were set per zone using the warning and alarm levels in Worked example 2. The operating procedures were updated to explain how to respond to a first ground-fault alarm.

Outcome

Nuisance alarms stopped, and the operator gained the ability to locate a real ground fault to a single row within minutes. The capacitance limit became part of the operator's standard purchase specification.

Part D — 20-Question Technical FAQ

Keentel Services, References and Contact

How Keentel Engineering Helps

Keentel Engineering supports data center owners, developers, operators and EPC contractors through every stage of the move from AC to DC power.

• Architecture and concept studies. Side-by-side comparison of optimized AC, hybrid and LVDC designs at the expected operating load: efficiency, copper, space, storage and cost.
• DC protection and coordination. Fault-current studies including batteries and capacitors, a five-level protection hierarchy, device rating and listing verification, and coordination for maximum and minimum fault cases.
• Grounding and insulation monitoring. Selection of the grounding scheme, insulation-monitoring zoning, ground-fault location and specification limits for equipment leakage and capacitance.
• DC arc-flash studies. Incident-energy analysis using NFPA 70E Annex D methods, labeling and safe work practice support.
• Bus stability and power quality. EMT and impedance-based studies of converter interaction, load steps and bus voltage regulation; IEEE 519 harmonic studies at the utility interface.
• Storage and on-site source integration. Battery sizing for ride-through and smoothing, UL 9540 / NFPA 855 coordination, and interconnection studies for on-site generation.
• Specifications and interoperability. Voltage, grounding, fault, precharge, mechanical and communication requirements written into purchase specifications.
• Utility interconnection and substation design. Large-load interconnection studies, 138 kV to 345 kV substations, and 34.5 kV and 13.8 kV distribution.
• Commissioning support. Protection testing, insulation-monitoring verification and integrated systems testing for DC systems.

Why Clients Choose Keentel

• AC and DC under one roof. We design utility interconnections, substations and facility power systems, so the DC design fits the whole system.
• Power system studies from 4 kV to 765 kV, including EMT modeling of power-electronic systems.
• US code and standards fluency — NEC, NFPA 70E, UL, IEEE, NESC and utility requirements.
• Independent, vendor-neutral advice focused on what the facility actually needs.
• Responsive senior engineers who stay with the project from concept through energization.

Talk to Keentel Engineering

Considering DC power for a new hall, planning high-density racks in an existing facility or reviewing a DC design? Keentel Engineering can help you choose the right architecture and design it safely.

• Phone: (813) 389-7871
• Email: contact@keentelengineering.com
• Book a 15-minute call: calendly.com/keentel-engineering/15min
• Web: keentelengineering.com

References

• NFPA 70, National Electrical Code, 2023 edition (including Articles 480, 645, 706 and 712).
• NFPA 70E, Standard for Electrical Safety in the Workplace, 2024 edition (including Informative Annex D).
• NFPA 855, Standard for the Installation of Stationary Energy Storage Systems.
• UL 9540, Energy Storage Systems and Equipment; UL 9540A, Test Method for Evaluating Thermal Runaway Fire Propagation in Battery Energy Storage Systems.
• UL 1973, Batteries for Use in Stationary and Motive Auxiliary Power Applications.
• UL 1778, Uninterruptible Power Systems.
• UL 62368-1, Audio/Video, Information and Communication Technology Equipment.
• UL 1741, Inverters, Converters, Controllers and Interconnection System Equipment for Use With Distributed Energy Resources.
• UL 891, Switchboards; UL 1558, Metal-Enclosed Low-Voltage Power Circuit Breaker Switchgear.
• IEEE C37.14, Standard for DC (3200 V and below) Power Circuit Breakers Used in Enclosures.
• IEEE C37.20.1, Standard for Metal-Enclosed Low-Voltage Power Circuit Breaker Switchgear.
• UL 489, Molded-Case Circuit Breakers, Molded-Case Switches and Circuit-Breaker Enclosures; UL 489B, for use with photovoltaic systems.
• UL 248 series, Low-Voltage Fuses; UL 2579, Low-Voltage Fuses — Fuses for Photovoltaic Systems.
• IEEE Std 519, Harmonic Control in Electric Power Systems.
• IEEE Std 1547, Interconnection and Interoperability of Distributed Energy Resources with Associated Electric Power Systems Interfaces.
• IEEE Std 1584, Guide for Performing Arc-Flash Hazard Calculations (AC only; cited for scope).
• ASHRAE TC 9.9, Thermal Guidelines for Data Processing Environments.
• Lawrence Berkeley National Laboratory, DC power distribution demonstration reports for data centers (mid-2000s).
• Industry alliance specifications for 380 Vdc distribution in data centers and telecom facilities (current editions).
• Open hardware industry group publications on rack-level and facility-level DC power distribution (current editions).

Disclaimer

This article is provided by Keentel Engineering for general technical education. It is not engineering advice for any specific project. Stage efficiencies, MTBF and MTTR values, fault parameters, thresholds and costs are illustrative; real projects must use manufacturer data, site-specific studies and the code editions adopted by the authority having jurisdiction. DC power distribution technology, equipment listings and code requirements are changing quickly; confirm current requirements before design. Case studies are composite and anonymized. Keentel Engineering accepts no liability for decisions made on the basis of this article without a project-specific review. © 2026 Keentel Engineering. All rights reserved. Written by Keentel Engineering.