UPS and Electrical Distribution in Data Centers: Medium Voltage, UPS Types and Redundancy

Every data center electrical system follows the same basic chain. The utility delivers power at transmission or distribution voltage to an on-site substation or service entrance. Medium-voltage switchgear distributes it around the campus, and transformers near each building or data hall step it down to low voltage, about 480 V in the U.S.¹ From there, power passes through the UPS, then through power distribution units (PDUs), remote power panels or overhead busway to the rack power strips that feed servers.

Generators sit in parallel with the utility, connected through transfer switches or paralleling switchgear. When the utility fails, the UPS carries the critical IT load from its batteries or flywheel while the generators start, synchronize and take over. Under NFPA 110, a Type 10 emergency power system must restore power within 10 seconds, including the generator start, the transfer and the generator accepting load.² Our guide to backup generators and fuel sizing covers the generator side.

The stakes are concrete. Uptime Institute’s 2025 outage analysis found that power remained the dominant cause of impactful outages, at 54% of cases, even as overall outage frequency declined for a fourth straight year.⁷ Its May 2026 analysis found power still the leading cause, with failures of UPS systems, transfer switches and generators dominant, while outage frequency per site fell for a fifth consecutive year.⁴

Low-voltage distribution at 480 V works for small buildings, but at tens of megawatts it means very large conductors, long cable runs and big switchboards. Medium-voltage distribution moves the same power at a fraction of the current, so cables are smaller and losses lower. A Consulting-Specifying Engineer case study of a multistory data center found medium voltage reduced cost compared with a 480 V design, using a 13,200 V system in a dual-ended main-tie-tie-main arrangement and placing substations close to the loads they serve to cut copper and installation cost.¹

On a large campus the medium-voltage system typically starts at the utility or customer-owned substation, where transmission voltage is stepped down. Bus arrangements and ownership options at that substation are covered in our guide to substation configurations. Inside the fence, medium-voltage feeders run in duct banks to pad-mounted or indoor unit substations at each building.

• Duct banks and manholes need corridors that avoid future building pads, which affects the campus master plan.
• Each unit substation and its switchgear line-up takes floor or pad area next to the hall it serves.
• Medium-voltage switchgear and transformers are long-lead items; see transformer and switchgear lead times before committing to a schedule.

Static UPS systems, built from power electronics and batteries, are the most common. Eaton’s widely cited comparison describes three static topologies that data center managers have traditionally chosen among: standby, line-interactive and double-conversion, with widely varying levels of efficiency, performance and protection.⁸ Double-conversion units rectify incoming AC to DC and invert it back to clean AC continuously, isolating the load from the grid. Line-interactive and standby units pass utility power through and switch to the inverter on a disturbance, which loses less energy but protects less.

Modern large UPS modules blur the line. Multi-mode units run in a high-efficiency mode and switch to full double conversion when power quality warrants.⁸ ENERGY STAR estimates that bypass, or eco, mode can cut data center energy costs by as much as 2–8%, while warning that the trade-off in performance or reliability varies by vendor.⁹ EPA’s UPS framework likewise notes that eco-mode trades off power conditioning for efficiency and set out to rate products without it.¹¹

Rotary systems are the main alternative. A diesel rotary UPS (DRUPS) couples a motor-generator and flywheel to a diesel engine, giving only about 8–12 seconds of stored ride-through before the engine carries the load. These units are usually installed outdoors in purpose-built enclosures, while static UPS sit inside the building.¹²

Efficiency matters because every lost kilowatt becomes heat the cooling plant must remove. ENERGY STAR puts new efficient UPS units at roughly 92–95% efficiency and older legacy units below 90%, notes that efficiency drops at the part loads most facilities run, and estimates certified units cut UPS losses by 30–55%.⁹ These losses feed directly into PUE.

Because generators should be on line within seconds, UPS batteries no longer need long runtimes. Older designs often sized lead-acid strings for about 15 minutes, a habit that inflates floor area when designers compare options.¹² Lithium-ion has been displacing valve-regulated lead-acid (VRLA) in centralized UPS plants: Uptime Institute’s 2021 survey found close to half of operators had adopted lithium-ion for centralized UPS, up from about a quarter three years earlier.¹³ Lithium-ion can reduce battery floor space substantially, by as much as 80% in some cases.¹²

Lithium-ion brings code questions that lead-acid did not. The 2018 International Fire Code set energy storage thresholds of 20 kWh for lithium-ion batteries and 70 kWh for lead-acid, above which energy storage system provisions apply.⁶ NFPA 855, the energy storage standard, adds limits often cited as 50 kWh per group, 250 kWh per listed array and 600 kWh per fire area, which shape how battery cabinets are grouped and separated. A UL 9540A fire test report can let the authority having jurisdiction approve larger units or closer spacing.¹⁴ Which edition applies depends on what the jurisdiction has adopted; our guide to building and fire codes for data centers explains how the code path works, and the fire marshal has the final word.

Redundancy describes how many spare components and paths exist beyond what the load needs (N). N+1 adds one spare unit to a shared system. 2N builds two complete, independent systems in an A/B arrangement with the load split evenly, so losing one side, or taking it down for maintenance, leaves the facility at N.³ The cost is that every generator, UPS and switchboard is duplicated, and each runs at no more than half its rating.

Distributed redundant designs spread the reserve across more systems. The common three-to-make-two (3M2) arrangement shares the load across three independent systems, any two of which can carry it, so more of the installed capacity is used than in 2N.³ Caterpillar notes that in a 3M2 generator system each unit is sized for the full block load and carries about 66.7% of its capacity under normal conditions.¹⁵ Block redundant, or catcher, designs run several primary blocks near full load and hold one or two reserve blocks that pick up a failed block through static transfer switches. Socomec describes an example with six normal streams and one or two redundant streams, and claims about 38% less footprint than 2N, a vendor figure.¹⁰

Uptime Institute’s Tier system frames the goals rather than the topology: Tier III requires a second delivery path so any component can be maintained without shutting down IT, and Tier IV adds fault tolerance so that equipment failures stop short of IT operations.¹⁶ Our guide to tier levels and redundancy covers certification and how tenants use it.

Electrical rooms are one of the largest non-IT uses of a data center building. Schneider Electric gives the example of a 1 MW static UPS system with five minutes of runtime needing nearly 2,000 square feet, a vendor figure that varies widely with battery chemistry.⁵ Rotary systems shift that space to outdoor enclosures, and published claims of 40–60% space savings for DRUPS assume a lead-acid, long-runtime static comparison.¹² Multiply by the redundancy factor: a 2N design carries roughly twice the UPS and switchgear of an N design, which is why block designs advertise smaller footprints.¹⁰

On the site plan, that translates into room for medium-voltage duct banks, transformer and generator yards, setbacks around outdoor equipment and access for replacing heavy gear. Our guide to cooling plant and equipment yard space shows how these yards compete with building pads.

AI loads are starting to change the chain itself. NVIDIA has proposed an 800 V DC architecture that removes AC conversion steps between the grid and GPUs, with a future DC power block converting grid power directly to 800 VDC.¹⁷ As of 2026, NVIDIA, Google and Microsoft were developing it through the Open Compute Project, with a row-level design using an overhead 800 VDC busway of up to 2 MW per row expected in 2027.¹⁸ For sites planned around high-density GPU racks, electrical rooms and yard layouts may look different from today’s AC designs.

Electrical design is the engineer’s job, but site decisions made before design begins can constrain it, especially where buildings, yards and duct banks go. Questions worth settling early:

1. 01What service voltage will the utility provide, and will the substation be utility- or customer-owned?
2. 02What redundancy will the target tenant require (N+1, 2N, distributed or block), and how does that scale the electrical rooms and yards?
3. 03Where will medium-voltage duct banks run, and do they keep future building pads clear?
4. 04Will the UPS use lithium-ion, and has the fire marshal’s adopted code edition been confirmed?
5. 05Is there yard space for generators, transformers and any rotary UPS enclosures, with setbacks and noise limits met?
6. 06Have long-lead transformers and switchgear been reflected in the schedule?

If you are weighing whether a parcel can fit the power chain a tenant expects, get a site reviewed before the layout is fixed.