800 VDC AI Data Centers: Rethinking UPS, Battery Storage, DC Protection and Cooling Resilience

November 05, 2026 | Blog

Part A — 800 VDC AI Data Centers at a Glance

Artificial intelligence has changed the electrical design of the data center faster than any technology since the server itself. A conventional enterprise rack drew 5 to 15 kW. Today's AI training racks draw 100 to 150 kW, and published industry roadmaps describe racks of several hundred kilowatts and, eventually, about 1 MW. At those densities, the familiar chain of AC power — utility, transformer, double-conversion UPS, power distribution unit, server power supply — starts to break down. The currents become too large, the copper too heavy, the conversion losses too costly and the floor space too valuable.

The industry's answer is high-voltage direct-current distribution to the rack, most often at 800 VDC. That raises an obvious question for owners and designers: if the rack is fed from a DC bus with batteries on it, do we still need a UPS?

Keentel Engineering's answer is yes — but not the UPS you are used to. The functions a UPS performs do not go away. Something must still:

• ride through utility disturbances and outages until the standby generators carry the load;
• isolate the critical load from the grid's voltage sags, swells and harmonics;
• provide redundancy, so a single failure does not drop the compute hall;
• and, new with AI, absorb the fast, synchronized power swings that GPU training jobs create, so the utility and the generators do not see them.

What changes is where those functions live and how they are built. In an 800 VDC architecture the rectifier, the battery and the regulating converter become parts of a DC power system instead of a self-contained AC UPS cabinet. Fewer conversion stages mean higher efficiency. Batteries sized only for "five minutes of runtime" can turn out to be too small in power, too large in energy or wrong in chemistry. Protection, grounding and arc-flash practice must move from AC to DC thinking. And cooling must be powered through the same disturbances as the GPUs, because a liquid-cooled rack can overheat in seconds.

This guide explains each of those changes in US engineering terms, with worked examples, and corrects several simplifications that appear frequently in industry discussions of 800 VDC.

The architecture used in this guide

Keentel Engineering uses one illustrative AI facility throughout this guide.

What Keentel's analysis shows

What Keentel's analysis shows

Figure 1. Left: illustrative end-to-end efficiency of a conventional AC UPS chain (87.2%) and an 800 VDC chain (91.8%), stage by stage. Right: a 1 MW AI rack load swinging ±30% at 0.5 Hz, and the much smoother power drawn from the grid or generator when a DC-bus energy buffer absorbs the swings. The buffer exchanges less than 0.1 kWh per cycle, but it cycles every two seconds.

Five lessons

• The UPS function stays; the UPS cabinet may not. Ride-through, power conditioning, redundancy and transient buffering are still required. In an 800 VDC design they are delivered by a rectifier, DC storage and DC/DC converters working as one system.
• Size the battery for power and cycling, not just minutes. Five minutes at 1 MW is 83.3 kWh at the load, but losses, aging and margin push the nameplate to roughly 137 kWh, and the discharge rate is about 7C. GPU power swings add millions of shallow cycles a year that most batteries are not designed for.
• GPU training loads are a grid problem as well as a facility problem. Synchronized swings of tens of percent at sub-second to few-second periods can excite generator governors and stress the transmission system. A DC-bus energy buffer is one of the most effective places to absorb them.
• DC fault energy is a design driver. Battery-fed DC buses can deliver tens of kiloamperes with no current zero to help interruption. Clearing time, not just voltage, decides arc-flash energy, which is why fast solid-state and hybrid breakers matter.
• Cooling must ride through with the compute. Chillers can stay on generator-backed AC, but coolant distribution unit pumps and controls must be on the uninterruptible system. A liquid-cooled loop may have only tens of seconds of bulk thermal buffer, and the GPUs themselves far less.

Engineering review of common statements about 800 VDC and UPS

Part B — Technical Guide: UPS, Storage, Protection and Cooling for 800 VDC AI Facilities

1. Why AI racks are moving to 800 VDC

1.1 The current problem

Every conductor, busbar, connector and breaker in a power system is sized by current. For a given power, current falls in direct proportion to voltage:

and resistive loss in a conductor of resistance R rises with the square of current:

A 15 kW enterprise rack at 54 V draws about 280 A, which a rack busbar handles easily. A 1 MW rack at 54 V would draw more than 18,000 A. No practical rack busbar, connector or protective device carries that current in the space of a rack, and the copper alone would weigh more than the compute.

Worked example 1 — current and loss for a 1 MW rack

Assume the same busway, with a loop resistance of 0.2 mΩ, feeding a 1 MW rack at three voltages.

Raising the distribution voltage from 54 V to 800 V reduces current by a factor of about 15 and conductor loss, for the same conductor, by a factor of about 220. In practice designers take part of the benefit as lower loss and part as less copper. Either way, the economics at hundreds of kilowatts per rack favor high-voltage DC.

1.2 Why DC rather than higher-voltage AC

AC distribution at 415Y/240 V or 480Y/277 V to the rack is possible and is used today for racks up to roughly 100–200 kW. DC has three further advantages at higher densities:

• Fewer conversion stages. Every GPU ultimately runs on low-voltage DC. An AC chain rectifies, inverts and rectifies again; a DC chain rectifies once.
• Simpler integration of storage. Batteries and capacitors are DC devices. On a DC bus they can be connected through a single converter, or in some designs directly.
• No reactive power or phase balancing at the rack. DC distribution avoids power-factor and three-phase balance considerations inside the hall.

The price is new engineering challenges: interrupting DC fault current, grounding and insulation monitoring, DC arc flash, and a less mature code and product base. Those are covered later in this guide and in Keentel's companion article on 800 VDC rack fault protection.

1.3 What "800 VDC" actually means

Three arrangements are described as "800 V":

A 1 MW rack draws 1,250 A only if 800 V is the pole-to-pole voltage. If the rack input is genuinely ±800 V, the current is 625 A but the system is a 1,600 V DC system with very different equipment ratings, clearances and code requirements. Always state the pole-to-pole voltage, the pole-to-ground voltage and the grounding method on the single-line diagram.

2. What a UPS does — and which functions survive in a DC architecture

A conventional double-conversion UPS performs five functions. All five are still needed in an 800 VDC facility.

The last row is new. A conventional UPS was never designed to absorb repeated swings of hundreds of kilowatts every few seconds. In an AI facility, that duty may set the size of the storage system more than ride-through does.

One function disappears: the static bypass. A conventional UPS can transfer the load directly to the utility if the inverter fails. A DC load cannot be fed from raw AC, so redundancy must come from multiple rectifier and converter paths instead. That makes the redundancy architecture more important, not less.

3. Four 800 VDC power architectures

3.1 Architecture A — AC UPS with rack-level AC/DC conversion

The facility keeps a conventional AC UPS (often at 415Y/240 V or 480Y/277 V), and each rack contains power shelves that convert AC to 800 V or directly to about 50 V. This is the least disruptive path and uses mature, listed equipment. It keeps the losses of double conversion plus rack rectification and puts significant heat and weight in the rack.

3.2 Architecture B — Centralized rectifier, DC-bus battery and regulated 800 V

A medium-voltage or 480 V supply feeds large rectifiers with active front ends. The rectifier output forms a DC bus with lithium-ion storage. A regulated DC/DC stage produces a stiff 800 V distribution bus to DC switchboards and rack feeds. This is the arrangement most often drawn for 800 VDC AI halls. Its strengths are fewer conversion stages, centralized storage and a regulated rack voltage. Its challenges are DC protection at high fault currents, large DC switchboards and a less mature listing landscape.

3.3 Architecture C — Medium-voltage solid-state transformer to 800 VDC

A solid-state transformer (SST) converts directly from 13.8 kV or 34.5 kV AC to 800 VDC, replacing the 60 Hz transformer, the low-voltage switchgear and the rectifier with one power-electronic system. SSTs promise the highest efficiency and smallest footprint. They are emerging technology at data center scale, with limited field history, and they move the medium-voltage interface, harmonic performance and protection into a single vendor's power electronics. Owners should treat them as a pilot technology with careful protection and reliability studies.

3.4 Architecture D — Distributed battery backup units

Battery backup units (BBUs) in each rack or row, as promoted in open hardware specifications, provide short ride-through close to the load, while the facility relies on generators for longer outages. BBUs remove a central UPS but distribute lithium-ion batteries throughout the white space, which affects fire protection, maintenance and code compliance under NFPA 855 and the local fire code.

3.5 Choosing among them

4. Conversion-chain efficiency

4.1 Why stages matter

The efficiency of a power chain is the product of the efficiencies of its stages:

Removing one stage at 96% efficiency is worth as much as improving every other stage by several tenths of a percent. That is the core efficiency argument for 800 VDC.

Worked example 2 — AC UPS chain vs 800 VDC chain for a 50 MW hall

Keentel used illustrative full-load stage efficiencies. Actual values depend on equipment, loading and operating mode, and must come from manufacturer data and tests.

For 50 MW delivered to the nodes:

The 800 VDC chain saves about 2.9 MW of input power. At continuous operation:

At $70/MWh that is about $1.8 million per year. Every megawatt of conversion loss also becomes heat that the cooling plant must remove, so the total saving is larger once cooling energy is included. Two cautions apply. First, modern AC UPS systems offer high-efficiency modes above 98% that narrow the gap, at the cost of less conditioning. Second, part-load efficiency matters: a 2N system often runs each side below 50% load, where converter efficiency falls.

4.2 The front end and harmonics

Whatever the architecture, the facility presents a large power-electronic load to the utility. Active front-end rectifiers keep current distortion low and can provide reactive power support; multi-pulse rectifiers with phase-shifting transformers are an alternative. The facility must meet IEEE Std 519-2022 at the point of common coupling, and large-load interconnection requirements increasingly ask for documented harmonic and ride-through performance. See Keentel's companion articles on harmonics and IEEE 519 for the method.

5. Redundancy architectures for DC power

Redundancy in a DC power system follows the same logic as in an AC UPS system, with one difference: there is no static bypass to fall back on.

For dual-fed DC racks, the two feeds are usually combined in the rack through diodes or active OR-ing, and each feed must be able to carry the full rack load after a failure. Apply the same loading rules used for A/B AC feeds: each feed carries about half the load normally and the full load after a failure, within the continuous rating of its protective device.

Sectionalize the DC bus. A single, solidly tied 800 V bus with all batteries and rectifiers connected can deliver very high fault current and can be taken down by a single bus fault. Splitting the bus into sections with fast tie devices limits fault current, contains failures and allows maintenance.

6. Sizing energy storage on the DC bus

6.1 Energy is only the starting point

The simplest sizing statement is runtime: "five minutes at full load." For a 1 MW rack that is:

That number is correct, but it is the energy the load consumes, not the battery nameplate. The battery must also cover converter losses, cannot be discharged to zero, loses capacity as it ages, delivers less at low temperature, and needs design margin.

Worked example 3 — battery nameplate and C-rate for a 1 MW rack

The battery must also deliver 1 MW, plus losses, from about 137 kWh of nameplate energy:

A discharge rate above 7C is far beyond typical energy-optimized lithium-ion cells used in grid storage (often 0.25C to 1C). The application needs high-power cells designed for UPS duty, with the thermal design, internal resistance and cycle-life ratings to match. In practice, the power requirement, not the energy requirement, frequently decides the battery.

6.2 Runtime must match the generator plan

Five minutes is a convention, not a requirement. The real requirement is the time needed to start, synchronize and load the generators, plus margin for a failed start and a retry or manual intervention. Section 8 develops a timeline. Many AI designs settle between 2 and 5 minutes at full load, but the decision should come from the generator plan and the owner's risk tolerance.

6.3 Chemistry and listing

Whatever the chemistry, the energy storage system should be listed to UL 9540, its cells and modules to UL 1973, and its fire behavior characterized by UL 9540A testing to support NFPA 855 and the local fire code. UPS equipment is listed to UL 1778.

7. GPU load transients and the grid

7.1 What the grid sees

Large AI training jobs run thousands of GPUs in lockstep. When the GPUs move between compute and communication phases, the whole cluster's power can swing by tens of percent within milliseconds and repeat at periods from fractions of a second to several seconds. A traditional data center load is almost flat; an AI training hall can look like a large, oscillating industrial load.

These swings matter in three places:

• Inside the facility, they stress converters and DC-bus voltage regulation.
• At the generators, during an outage, they can exceed the step-load capability of the engines and excite governor oscillations.
• At the grid, multi-hundred-megawatt campuses swinging together can create forced oscillations, voltage flicker and frequency disturbances. NERC's large-load work, ERCOT and SPP have all raised concerns about large computational loads, including their ride-through behavior and their fluctuations.

7.2 Buffering on the DC bus

A DC power system is a natural place to absorb these swings. A fast energy buffer — capacitors, supercapacitors or a high-power battery under a dedicated controller — supplies the rapid component of the load, while the rectifier draws a smoothed average from the AC system.

Figure 1 (right) shows a 1 MW rack swinging ±30% at 0.5 Hz. With a 4-second smoothing time constant on the rectifier, the AC-side swing falls from 600 kW peak-to-peak to about 80 kW, and the buffer exchanges well under 0.1 kWh per cycle.

Worked example 4 — why battery cycle life is the hidden constraint

Assume the buffer exchanges about 0.05 kWh every 2-second cycle and the battery is the 137 kWh system from Worked example 3.

Five to six thousand equivalent full cycles a year would consume the rated life of many lithium-ion products in one to two years. That is why designs pair batteries for ride-through with capacitive or supercapacitor buffers for transients, or select chemistries rated for very high cycle counts. The control system decides which device carries which frequency band.

8. Generator coordination

8.1 The transfer timeline

The ride-through storage must cover the full time from loss of utility to stable generator operation, with margin.

Worked example 5 — outage timeline for a 13.8 kV paralleled generator plant

On this timeline, two to five minutes of storage provides a reasonable margin. Less than one minute leaves no room for a failed start. Far more than five minutes adds cost and fire load without much benefit if the generator plant is reliable.

8.2 Step loading and rectifier walk-in

Diesel generators commonly accept large single steps; natural-gas engines often accept much smaller steps. A hall of power-electronic loads cannot be dropped onto generators all at once without large frequency and voltage dips. Rectifiers should "walk in," ramping their input power over several seconds while the DC storage carries the balance. The ramp rate must be coordinated with each generator's step-load acceptance and governor response.

Generators also see the GPU power swings described in Section 7. During extended generator operation, the DC-bus buffer must keep those swings away from the engines, or the governor and voltage regulator must be shown, by test or simulation, to tolerate them.

8.3 Returning to utility

Retransfer to utility after an outage should be a closed transition where the utility allows it, or a controlled open transition carried by the DC storage. Either way, the storage state of charge must be sufficient at the moment of retransfer, and recharge current must be limited so the recharging batteries do not add a large step load to the generators or the utility.

9. DC bus voltage regulation and transients

9.1 Steady-state voltage drop

DC voltage drop is simply current times resistance:

At 1,250 A, a loop resistance of 0.2 mΩ causes a 0.25 V drop — negligible at 800 V. At 54 V and 18,500 A, the same conductor would drop 3.7 V, almost 7% of the supply. Again, the higher voltage makes the problem disappear.

9.2 Holding the bus through a load step



When the load steps faster than the regulating converter can respond, the bus capacitance supplies the difference. The energy drawn from the capacitor during the step must keep the voltage within its limit:

For small deviations this simplifies to:

Worked example 6 — capacitance for a 300 kW step

A 1 MW rack steps up 30% (300 kW). The regulating converter needs 50 ms to respond. The bus must stay within 2% of 800 V (16 V).

The exact energy calculation gives about 1.18 F. That capacitance stores only about 0.1 kWh at 800 V — trivial as energy, but a large bank in film or electrolytic capacitors and a significant source of fault current. Supercapacitor modules behind a DC/DC converter can provide the same effect with smaller physical size, at the cost of another converter. The converter control bandwidth is the real lever: halving the response time halves the capacitance.

10. DC protection, grounding and arc flash (summary)

Keentel's companion article on 800 VDC rack fault protection covers solid-state circuit breakers, selectivity and fault detection in depth. This section summarizes what changes when the UPS becomes a DC system.

10.1 Fault current from the storage

A battery-fed DC bus can deliver very large fault currents, limited only by internal resistance and cable resistance:

Worked example 7 — bolted fault current on an 800 V bus

Four battery strings, each with 25 mΩ internal resistance, are connected in parallel to the bus through cables totaling 5 mΩ.

Seventy kiloamperes of DC, with no current zero to help interruption, exceeds the interrupting rating of many low-voltage DC devices. The bus capacitance adds a very fast discharge peak on top. Fault current, not just load current, must set the interrupting ratings of every device on the bus — and is the strongest argument for sectionalizing the bus and for current-limiting fuses or solid-state breakers near the storage.

10.2 Grounding and ground-fault detection

The most common 800 VDC arrangement is ±400 V with the midpoint grounded through a high resistance. A first ground fault on either pole then causes only a small current, the system keeps running, and an insulation monitoring device (IMD) or ground-fault detector raises an alarm. A second ground fault on the opposite pole is a pole-to-pole fault, so the first fault must be located and cleared promptly. Ungrounded and solidly grounded DC systems are also possible; each changes fault current, touch voltage and detection. Document the choice and follow NEC Article 250 provisions for DC systems.

10.3 DC arc flash

IEEE Std 1584-2018 covers three-phase AC systems only. For DC, NFPA 70E-2024 Informative Annex D provides calculation methods, including the maximum-power method, which assumes the arc voltage is half the system voltage:

where IE is in cal/cm², V(sys) in volts, I(bf) is the bolted fault current in amperes, T(arc) is the arcing time in seconds and D is the working distance in centimeters. Annex D notes that arcs in enclosures can produce up to about three times the open-air value.

Worked example 8 — DC incident energy at 800 V

Using the 71 kA bolted fault current and a working distance of 45.7 cm (18 in.):

The maximum-power method is deliberately conservative, but the lesson is robust: on a battery-fed DC bus, clearing time decides the hazard. Fast interruption near the storage turns a potentially lethal arc flash into a manageable one. Labels, PPE and energized-work procedures must follow NFPA 70E.

11. Cooling resilience

11.1 What must stay on

A liquid-cooled AI rack has very little thermal buffer. If coolant stops flowing, GPU junction temperatures can reach throttling limits in seconds. The cooling system must therefore be split by how fast each component matters:

11.2 Thermal ride-through



The bulk thermal buffer of a coolant loop is:

Worked example 9 — coolant-loop ride-through for a 1 MW rack

Assume the secondary loop for one rack holds about 400 kg of 25% propylene glycol (c(p) ≈ 3,900 J/kg·K), the allowable coolant temperature rise is 10 K, and 85% of the rack's 1 MW goes into the liquid.

Eighteen seconds is the bulk average, and only if the coolant keeps circulating. If the pumps stop, the coolant at the cold plates heats almost immediately and GPUs can throttle within a few seconds. Pumps on the uninterruptible supply are not optional. Chillers can ride through on generators only if chilled-water or coolant thermal storage covers their restart time, which for large chillers can be several minutes. Keentel's companion articles on CDU liquid cooling and digital-twin pump-failure analysis develop these transients in detail.

12. Codes, standards and listings

DC provisions in US codes are evolving quickly. Code editions after 2023 expand coverage of higher-voltage DC systems, and local adoption varies. Confirm the edition adopted by the authority having jurisdiction and engage the AHJ early, because many 800 VDC products are new and some listings are still in progress.

13. A practical design checklist

• Define the voltage system on the single-line diagram: pole-to-pole voltage, pole-to-ground voltage, grounding method and insulation monitoring.
• Choose the architecture (A, B, C or D) based on rack density, maturity, footprint and the owner's risk tolerance.
• Calculate conversion-chain efficiency at realistic part loads, not just at full load.
• Set the redundancy architecture and verify each feed and module can carry its post-failure load.
• Size storage for power, energy and cycling, separately for ride-through and for transient buffering.
• Build the generator timeline, including detection, start, synchronization, walk-in and retry margin.
• Study GPU load transients with the facility's actual workload profile, and specify what the utility and generators may see.
• Calculate DC fault currents from storage, capacitance and converters, and set interrupting ratings and sectionalizing accordingly.
• Perform DC arc-flash analysis using NFPA 70E methods and specify clearing times that keep incident energy manageable.
• Put CDU pumps and controls on the uninterruptible supply, and verify thermal ride-through for chillers on generators.
• Coordinate with the utility and the ISO on harmonics, ride-through and fluctuation requirements for large loads.
• Engage the AHJ early on listings, battery siting and DC code provisions.
  

Part C — Three Anonymous Case Studies

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

Case study 1 — A "five-minute" battery that could not deliver the power

Situation. A developer in the Southeast was converting part of an existing data center into a 12 MW AI training hall with 200 kW racks. The design used a centralized rectifier, lithium-ion storage on the DC bus and regulated DC distribution to the racks. The battery had been specified as "five minutes at full load," and the vendor offered an energy-optimized product sized at exactly the load-side energy plus a small margin.

Analysis. Keentel's review recalculated the requirement using converter efficiency, usable depth of discharge, end-of-life capacity and temperature. The required nameplate was about 65% higher than offered. More importantly, the discharge rate at full load was above 6C, while the offered cells were rated for continuous discharge near 1C. The vendor's datasheet showed that, at the required power, the battery would reach its voltage cutoff in under a minute and its cells would overheat. Keentel also found that the training workload profile supplied by the owner showed power swings of about 25% every one to three seconds, which the battery controller had not been asked to handle.

Engineered solution. Keentel specified a high-power UPS-grade lithium iron phosphate system sized for the corrected power and energy, with a separate supercapacitor buffer and controller to absorb the training-load swings. The rectifiers were set to a slow power ramp so the utility and the generators would see a smoothed load. The generator timeline was documented and the storage runtime set at three minutes based on it.

Outcome. Commissioning tests with simulated outages and recorded workload profiles showed the bus held within its voltage limits, the generators carried the hall after a 40-second transfer, and the AC-side power swing was reduced by more than 80%. The owner avoided a storage system that would have failed its first full-load test.

Case study 2 — An 800 VDC bus with arc-flash energy nobody had calculated

Situation. An AI campus in Texas installed 800 V DC switchboards fed by rectifiers and a large battery system, with all battery strings tied to a single bus. The electrical contractor asked for arc-flash labels before energization. The design team had performed an IEEE 1584 study for the AC system but had no DC study.

Analysis. Keentel modeled the DC system, including the battery strings, the DC bus capacitance and the converter contributions. The bolted fault current on the tied bus exceeded 80 kA, and the main DC breakers had a clearing time near 100 ms. Using the NFPA 70E Annex D maximum-power method, incident energy at the main switchboard exceeded 80 cal/cm² for an arc in an enclosure — beyond any practical PPE category. Several branch devices were also applied above their DC interrupting ratings.

Engineered solution. Keentel recommended splitting the battery into four sections, each connected to the bus through current-limiting DC fuses and a fast tie arrangement, and replacing the branch devices with properly rated DC breakers. Maintenance procedures were changed so that energized work on the main bus required the battery sections to be isolated. Insulation monitoring was added to the high-resistance grounded midpoint.

Outcome. Available fault current at the switchboard fell to about a quarter of its original value and clearing times dropped by an order of magnitude. Incident energy fell to levels where labeled PPE and energized-work procedures were practical. The authority having jurisdiction accepted the revised design and the study.

Case study 3 — Chillers on generators, pumps not on anything

Situation. A liquid-cooled AI facility in the Mid-Atlantic tested its emergency transfer. The compute load rode through on the DC storage exactly as designed, but within seconds of the utility outage, dozens of racks throttled and several shut down on high temperature.

Analysis. Keentel reviewed the electrical drawings and event logs. The CDU pumps had been connected to the generator-backed mechanical distribution, not to the uninterruptible system, on the assumption that "cooling is on the generators." During the 30-second generator start and transfer, coolant flow to the racks stopped completely. The bulk coolant-loop buffer was about 20 seconds, but with no flow, the cold plates heated much faster, and GPU temperatures reached their protective limits in under ten seconds.

Engineered solution. Keentel moved the CDU pumps, valve actuators and cooling controls to the uninterruptible DC system through dedicated DC/DC converters, with selective protection so a pump fault could not affect the compute bus. The chiller plant stayed on generator-backed AC, and a chilled-water storage tank was sized to cover the chillers' restart time with margin.



Outcome. A repeat transfer test showed uninterrupted coolant flow, no GPU throttling and stable coolant temperatures through the full generator start and chiller restart. The owner adopted the change as a standard for all subsequent halls.

Part D — 20-Question Technical FAQ

Keentel Services, References and Contact

How Keentel Engineering Helps

Keentel Engineering supports AI data center developers, owners, operators and EPC contractors from concept to commissioning:

• Power architecture studies. Comparison of AC UPS, centralized DC, solid-state transformer and distributed battery architectures, with efficiency, reliability, footprint and cost analysis.
• Single-line diagrams and design. MV service, unit substations, rectifier and DC distribution design, with clearly defined voltage systems and grounding.
• Energy storage sizing. Ride-through and transient-buffer sizing for power, energy and cycle life, with generator timeline analysis.
• Generator integration. Paralleling, step-load and walk-in studies, transfer schemes and retransfer coordination.
• DC protection, grounding and arc flash. DC short-circuit studies, device ratings, selectivity, insulation monitoring and NFPA 70E DC arc-flash studies and labels.
• Cooling power resilience. Classification of cooling loads by time to consequence, uninterruptible supply design for CDUs and thermal ride-through analysis.
• Utility interconnection and large-load studies. Harmonic analysis to IEEE 519-2022, ride-through and fluctuation studies, and EMT modeling of large computational loads for utilities and ISOs.
• Code and AHJ support. NEC, NFPA and UL compliance strategies for new DC equipment.

Why Clients Choose Keentel

• Grid and facility expertise in one team. We design from the 230 kV substation to the 800 V rack feed.
• Power system studies from 4 kV to 765 kV, including EMT modeling of power-electronic loads.
• US codes and standards fluency. NEC, NFPA, UL, IEEE and NERC, applied to emerging DC technology.
• Evidence-based design. Calculations and simulations tied to equipment data and commissioning tests.
• Responsive senior engineers, with direct access to experienced professional engineers throughout the project.

Talk to Keentel Engineering

Planning an AI hall at 200 kW per rack or beyond? The power architecture decisions you make now — voltage, storage, protection and cooling resilience — will set your efficiency, reliability and safety for the life of the facility. Keentel Engineering can help you get them right.

• 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 — Articles 250, 480, 645, 706 and 712.
• NFPA 70E-2024, Standard for Electrical Safety in the Workplace — Informative Annex D, DC arc-flash calculation methods.
• NFPA 70B-2023, Standard for Electrical Equipment Maintenance.
• NFPA 110, Standard for Emergency and Standby Power Systems.
• 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 Applications; UL 1778, Uninterruptible Power Systems.
• IEEE Std 519-2022, IEEE Standard for Harmonic Control in Electric Power Systems.
• IEEE Std 1584-2018, IEEE Guide for Performing Arc-Flash Hazard Calculations (AC systems).
• IEEE Std 946, IEEE Recommended Practice for the Design of DC Power Systems for Stationary Applications.
• IEEE Std 493, IEEE Recommended Practice for the Design of Reliable Industrial and Commercial Power Systems (Gold Book).
• Doan, D. R., "Arc Flash Calculations for Exposures to DC Systems," IEEE Transactions on Industry Applications, 2010.
• Stokes, A. D. and Oppenlander, W. T., "Electric Arcs in Open Air," Journal of Physics D: Applied Physics, 1991.
• NERC, Large Loads Task Force materials and white papers on large computational loads.
• Open Compute Project, specifications for rack power and battery backup units (general reference).

Disclaimer

This article is provided by Keentel Engineering for general technical education. It is not engineering advice for any specific project. Architectures, efficiencies, fault currents, incident energies and other results are illustrative and simplified; actual values depend on equipment, settings, operating conditions and project-specific studies. Codes, standards, listings and utility requirements change over time, and local adoption varies; always consult the current approved documents and the authority having jurisdiction. Keentel Engineering accepts no liability for decisions made on the basis of this article without a project-specific engineering review.

© 2026 Keentel Engineering. All rights reserved. Written by Keentel Engineering.