Independent advisory for colocation, GPU and data center projects

A reply within one business day

Metro Colo Advisory Research

Data Center Power Infrastructure
The Power Chain from Utility Feed to Rack, and How Much of It You Can Actually Lease

Every data center megawatt passes through switchgear, transformers, generators, UPS systems and PDUs before it reaches a rack. The redundancy built into that chain decides how much of it can be leased. Built three plus one, N+1 gives you three of four UPS modules to use.

Send Your Power Requirements
Utility to rack, link by link N+1 vs 2N, worked out Free to cite, CC BY 4.0

Updated November 2026. Every number links to its source.

ResearchUpdated November 2026Every figure links to its source

Inside a data center, utility power runs through medium-voltage switchgear, transformers, transfer switches and generators, UPS systems, PDUs and rack power strips before it reaches a server. Each link is built with spares, and spares are not capacity you can lease. In N+1 built three plus one, three of four UPS modules carry the load, so 75% of the installed UPS is usable. In 2N it is 50%. When you lease wholesale colocation, the number that matters is the critical IT load left after that redundancy.

This guide walks the power path link by link, works out how N+1, block redundant, distributed redundant and 2N designs shrink usable capacity, shows what AI racks at 44, 80 and 130 kW ask of the distribution, and explains why transformer and switchgear lead times set delivery dates. It is the inside-the-building companion to our guides to data center power consumption and how data centers are powered. When you are underwriting a building or a lease, our data center consulting team checks the power chain against the drawings.

Usable in N+1
75%
three plus one: three of four UPS modules carry the load, one is the spare
Usable in 2N
50%
two full systems, each sized to carry the whole load
Outages from power
45%
of operators' most recent impactful outages in 2025
Transformer wait
128 wk
average power transformer lead time, Q2 2025; switchgear 44 weeks

Sources: Schneider Electric White Paper 75, Comparing UPS System Design Configurations (Table 2: four 80 kW modules for a 240 kW load); CoreSite, What is Data Center Redundancy? N, N+1, 2N, 2N+1; Uptime Institute, Annual Outage Analysis 2026; Wood Mackenzie, Mind the gap (Oct 2025). Usable shares are our arithmetic.

Comparing the power behind two sites or two leases?
Send us the load, the redundancy you need and the date you need it. You'll hear back within 24 hours.
Send Your Requirements

The Data Center Power Chain, Utility to Rack

Power arrives at medium voltage from a utility substation, is stepped down to 480 V, passes transfer switches that can swap in generators, runs through UPS systems that bridge any gap, and is split by PDUs and panels into the circuits each rack plugs into. ASHRAE describes the chain the same way: a utility supply at medium voltage "feeds medium-voltage switchgear," a transformer steps it down to "low-voltage (typically 480V and lower)," and "Low-voltage switchgear generally feeds UPS units," which feed room PDUs, remote power panels and finally rack PDUs. Tier III and Tier IV designs run two of these paths, A and B, to every rack.

The power path inside a data centerSimplified two-path design. Voltages are typical North American values.
Utility feed and substationMedium voltageService sized in MVA: the site's ceiling
Medium-voltage switchgeare.g. 13.8 kVTakes in the feeds and protects the site
Generatorse.g. N+1, 12+ hr fuelA Tier III example. Feed both transfer switches
Path ATransformerMV to 480 VSteps power down
Path BTransformerMV to 480 VSteps power down
Path ALV switchgear and ATS480 VSwitches to generators in an outageGenerators feed in here
Path BLV switchgear and ATS480 VSwitches to generators in an outageGenerators feed in here
Path AUPS and batteries480 VBridges 10 to 20 seconds to generator power
Path BUPS and batteries480 VBridges 10 to 20 seconds to generator power
Path APDU, RPP or busway208 or 415 VSplits power into rack circuits
Path BPDU, RPP or busway208 or 415 VSplits power into rack circuits
Rack PDUs, A and BDual-cordedEither feed can carry the whole rack
SharedPath APath BBackup sourceThe rack

Sources: ASHRAE TC 9.9, Data Center Power Equipment Thermal Guidelines and Best Practices; Schneider Electric, UPS design considerations for data centers (UPS bridge "typically 10-20 seconds"); POWER Magazine, Tier III electrical design guide (Dec 2024) (a Tier III design example: N+1, at least 12 hours of fuel, two power inputs per piece of equipment); NVIDIA, 800 VDC architecture for AI factories (13.8 kV example). Diagram is our simplification; designs vary.

LinkWhat it doesWhat it means for usable capacity
Utility service and substationDelivers power at medium voltage under the utility's service agreement.The site's hard ceiling. IT load, cooling and losses all fit inside it. Our power consumption guide converts grid MVA to IT MW.
Medium-voltage switchgearBreakers and switches that take in the utility feeds and send them to the transformers.Adding a feed or a block means new switchgear, which averaged 44 weeks to deliver in Q2 2025.
TransformersStep medium voltage down to 480 V for the building, one or more per path.Wood Mackenzie's 128-week average (Q2 2025) is for power transformers, the large units usually found at the substation. It gives no weeks figure for the smaller pad-mount units that often feed 480 V, only that their lead times rose that quarter. Either way, the transformer count sets how fast capacity can grow.
LV switchgear and transfer switchesDistribute 480 V and switch the load between utility and generators.Transfer switch issues rank just behind UPS failures among the causes of power outages Uptime's 2026 survey cites most often.
GeneratorsStart when the utility fails and carry the site until it returns.They carry cooling as well as IT: cooling is "often backed up by generator power only." POWER Magazine's Tier III design guide calls for N+1 redundancy on every critical system, generators included, and generator fuel storage for at least 12 hours.
UPS and batteriesHold the IT load through sags and outages until the generators take over.The UPS output after redundancy is the critical IT load you can lease. Runtime is "typically" 5 minutes in cloud and colocation sites.
PDUs, RPPs and buswayStep 480 V down to 208/120 V or 240/415 V where needed and split it into branch circuits, or carry it overhead on busway.Block size limits density: Schneider calls 800 A blocks at 240/415 V standard and wants a block to serve at least a full row of an AI cluster.
Rack PDUsPower strips in the cabinet, one on the A feed and one on the B feed.A 60 A, 240/415 V rack PDU is limited to 34.6 kW, and each one runs at half or less so the other can take over.

Sources: ASHRAE TC 9.9, Data Center Power Equipment Thermal Guidelines and Best Practices; Wood Mackenzie, Mind the gap (Oct 2025); Uptime Institute, Annual Outage Analysis 2026; Schneider Electric, UPS design considerations for data centers; POWER Magazine, Tier III electrical design guide (Dec 2024); Schneider Electric White Paper 110, The AI Disruption; Belden, Exploring dual power feeds in data centers.

Redundancy: N, N+1, 2N and What You Can Lease

Redundant modules protect uptime but carry no load you can sell: N+1 built three plus one gives three modules of leasable capacity out of four, 2N gives half, and every other design lands in between. Schneider Electric defines N as the "need" of the critical load, and calls a system N+1 when the "spare" amount of power is at least equal to the capacity of one module. The more modules share the load, the smaller the spare's slice of the total.

How much installed UPS carries the load in N+1Schneider's example: one 240 kW critical load on N+1 modules of different sizes
2 x 240 kW modules50%
3 x 120 kW modules66%
4 x 80 kW modules75%
5 x 60 kW modules80%
Four 80 kW modules, three plus one, install 320 kW to carry 240 kW. The fourth module is redundancy, not leasable capacity. Values are Schneider's "% each UPS module is loaded."

Source: Schneider Electric White Paper 75, Comparing UPS System Design Configurations.

Worked example: a 3.2 MW critical load on 800 kW modules

Hold the load steady and the installed UPS grows with each step up in redundancy. Schneider uses a 3.2 MW data center to compare designs; we filled in the other rows with the same 800 kW module.

Design800 kW modulesInstalled UPSUsable shareModules (navy carry load, outlined are spare)
NCapacity only43.2 MW100%
N+1Parallel redundant, four plus one54.0 MW80%
Distributed redundantSchneider's count64.8 MW66.7%
2NSystem plus system86.4 MW50%
2N+12N plus one module97.2 MW44%
2(N+1)Schneider's count108.0 MW40%

Sources: Schneider Electric White Paper 75, Comparing UPS System Design Configurations ("a 2(N+1) design may require ten 800 kW modules"; distributed redundant "requires only six 800 kW modules"); CoreSite, What is Data Center Redundancy? N, N+1, 2N, 2N+1 (N, N+1, 2N and 2N+1 definitions). Module counts for N, N+1, 2N and 2N+1, installed totals and usable shares are our arithmetic.

The five designs you will see in a spec sheet

N+1, parallel redundant75%usable at three plus one

"A parallel redundant configuration consists of paralleling multiple, same size UPS modules onto a common output bus." Any one module can fail or be serviced while the rest carry the load. Usable share is modules minus one, divided by modules.

Schneider Electric White Paper 75, Comparing UPS System Design Configurations
Block redundant (catcher)75%usable with three blocks plus one reserve

Each block carries its own load, and a reserve block catches any that fails: "three of them supply the primary power and one remains in reserve for failover." The reserve idles, and the design needs load transfer equipment.

STACK Infrastructure, Block vs. Distributed Redundancy
Distributed redundant66.7%usable in three-to-make-two

Three or more modules "with independent input and output feeders," with the load spread so any one can fail. In three-to-make-two, each system carries a third of the load (33.4%), which is up to 66.7% of its rating. Four-to-make-three reaches 75%, at the cost of "complex and precise load planning."

Schneider Electric White Paper 75, Comparing UPS System Design Configurations; Consulting-Specifying Engineer, Data center design considerations; STACK Infrastructure, Block vs. Distributed Redundancy
2N, system plus system50%usable: two full systems

"Simply designing twice as much equipment as needed for the base (i.e., N) load." If four UPS units carry the load, CoreSite's 2N example adds "an additional four UPS units, for a total of eight systems."

Consulting-Specifying Engineer, Data center design considerations; CoreSite, What is Data Center Redundancy? N, N+1, 2N, 2N+1
2N+144%usable with four modules per side

"2N+1 delivers the fully fault-tolerant 2N architecture plus an extra component for an added layer of protection." With N of four, nine modules carry the load of four (our arithmetic).

CoreSite, What is Data Center Redundancy? N, N+1, 2N, 2N+1
At the rack40%of each A/B circuit's breaker rating

Redundancy repeats at the cabinet. "When one power feed fails, the remaining supply supports the entire load," so Belden advises loading each rack PDU to no more than 50%. Add the 80% continuous rating and each circuit carries at most 40% of its breaker (our arithmetic).

Belden, Exploring dual power feeds in data centers; Schneider Electric White Paper 110, The AI Disruption

So ask which design a quoted figure assumes. A building advertised with 10 MW of UPS modules can mean 10 MW of leasable IT with no redundancy, 7.5 MW in N+1 built three plus one, or 5 MW in 2N (our arithmetic). The converter on our power consumption page switches between N, N+1 and 2N for any MW figure.

Leasable Capacity vs the Headline MW

The capacity you can lease is the critical IT load the UPS can carry after redundancy, delivered through circuits that each run well below their rating, and it is always smaller than the utility service or the installed equipment total. One building gets quoted three ways: the utility service in MVA, the installed UPS rating, and the critical IT load. Our data center power consumption guide explains how PUE and power factor connect them and converts any figure. Inside the building, two more haircuts apply: redundancy, covered above, and circuits that run below their breaker rating.

Two 60 A, 415 V breakers
86.4 kW
Nameplate of the A and B pair: 2 x 60 A x 3 phases x 240 V, the formula Schneider uses (our arithmetic)
80% continuous rating
69.1 kW
80% of 86.4 kW, or 2 x 34.6 kW, Schneider's usable figure for one 60 A, 240/415 V rack PDU
Either feed alone
34.6 kW
Each feed runs at half or less so the other can carry the rack
What the rack can use
40%
34.6 kW of 86.4 kW of breaker nameplate (our arithmetic)

Sources: Schneider Electric White Paper 110, The AI Disruption; Belden, Exploring dual power feeds in data centers.

The same logic runs up the chain. A provider that quotes UPS capacity before redundancy, or adds the A and B feeds together toward a rack's kW, is quoting a number you cannot load. Ask for critical IT kW per rack and MW per hall with one module or one feed out of service. That is the figure our readouts use when we compare wholesale colocation options for a client.

Data Center Tiers and Power Paths

Uptime Institute's four Tiers rate a site's whole infrastructure, power and cooling alike. On the power side, Tier I has one path and no spares, Tier II adds spare components, Tier III adds a second distribution path so anything can be maintained without a shutdown, and Tier IV makes the whole system fault tolerant. Each step up usually means more redundant equipment, so a higher Tier tends to mean a smaller share of installed UPS you can lease.

TierPower path (quoted words are Uptime's)UPS design paired with itWhat it means for leasable capacity
Tier IBasic capacityOne path, no redundancy: "The facility will have to shut down completely for preventive maintenance and repairs."Capacity (N)All installed UPS can carry load, and maintenance means downtime.
Tier IIRedundant capacity components"redundant capacity components for power and cooling" on a single path.Isolated redundantA spare module or system is held back from the load.
Tier IIIConcurrently maintainable"redundant distribution paths to serve the critical environment"; "no shutdowns when equipment needs maintenance or replacement."Distributed redundant (Schneider)POWER Magazine's Tier III guide: at least N+1Depends on the design: 66.7% in three-to-make-two; 75% if built as three plus one N+1 on two paths.
Tier IVFault tolerant"several independent and physically isolated systems" as redundant components and distribution paths.System plus system (2N, 2N+1)Half of installed UPS or less.

Sources: Uptime Institute, Tier Classification System; Schneider Electric White Paper 75, Comparing UPS System Design Configurations (Table 1, for Tier I, II, III and IV pairings); POWER Magazine, Tier III electrical design guide (Dec 2024). Schneider's Table 1 also lists parallel redundant (N+1) between isolated and distributed redundant, but the published text does not show which Tier that row sits under, so we do not assign it. Uptime's own Tier page does not use the term N+1. Leasable shares are our arithmetic from the designs above.

For uptime percentages, certification and what each Tier costs to build, see our guide to data center tiers.

High Density and AI Racks: What the Power Chain Has to Deliver

AI racks draw about 44 kW on average in a large H100 cluster, 80 kW in an eight-server training rack and about 120 to 130 kW in an NVIDIA GB200 NVL72, which pushes distribution from 208 V to 415 V, from cords to busway, and toward 800 V DC for the 1 MW racks NVIDIA plans from 2027. At those loads the limits are the circuits and blocks downstream of the UPS, not floor space.

Power per AI rack (kW)Vendor examples. NVIDIA's 1 MW racks planned for 2027 sit far off this scale.
Large H100 cluster, average44 kW
Eight HPE Cray XD670 servers80 kW
NVIDIA GB200 NVL72about 120 to 130 kW

Sources: Schneider Electric White Paper 110, The AI Disruption; NVIDIA, DGX GB200 user guide: hardware ("The rack power consumption is approximately 120kW"); Supermicro, GB200 NVL72 datasheet ("Operating Power 125kW to 135kW"); NVIDIA, 800 VDC architecture for AI factories. The two GB200 NVL72 figures differ, so we give about 120 to 130 kW, the range our AI and GPU colocation guide uses.

How many circuits an AI rack needs

RackRack power60 A, 415 V circuits per feedCircuits with A and B
Large H100 cluster, average rack44 kW24
Eight HPE Cray XD670 servers80 kW36
NVIDIA GB200 NVL72eight 33 kW power shelves in N+N redundancy; Supermicro's 132 kW total is four shelves' worth (our arithmetic)about 120 to 130 kW48

Our arithmetic: rack kW divided by 34.6 kW per 60 A, 240/415 V rack PDU (Schneider Electric White Paper 110, The AI Disruption), rounded up, then doubled for A and B. For comparison, Schneider counts five 60 A circuits at 120/208 V for the 80 kW rack at 1N, and "If 2N was required (although uncommon for AI training loads), this number would double to ten." Rack power and power shelves: NVIDIA, DGX GB200 user guide: hardware; Supermicro, GB200 NVL72 datasheet. At 120 or 130 kW the count is the same, four circuits per feed (120 / 34.6 = 3.5 and 130 / 34.6 = 3.8, both rounded up). Real designs vary; many AI racks take higher-amp feeds or busway.

415 V instead of 208 V

2xpower per amp

"a 208 V, 60 A system provides the same power as a 415 V, 30 A system," and dropping the PDU transformer is worth "a 2 percent efficiency gain," with total gains "typically 4 to 5 percent." Schneider says sites on 120/208 V "should retrofit their distribution to 240/415 V."

Server Technology (Legrand), 415 VAC power distribution white paper; Schneider Electric White Paper 110, The AI Disruption

Busway and bigger blocks

800 Astandard block at 240/415 V

"Distribution block sizes must increase to meet the demands of high-density clusters," at minimum to serve a full row. Schneider's future design feeds "a 3,000 A busway at 415/240 V" from a 2 MW transformer. Arc flash risk "increases as rPDU amp rating increases."

Schneider Electric White Paper 110, The AI Disruption

800 V DC for 1 MW racks

Up to 200 kgof copper busbar for a 1 MW rack at 54 VDC

Today's AI racks "rely on 54 VDC power distribution." NVIDIA's plan converts 13.8 kV AC "directly to 800 VDC at the data center perimeter," improving end-to-end efficiency "by up to 5%," with full-scale production alongside its Kyber racks in 2027.

NVIDIA, 800 VDC architecture for AI factories

Liquid cooling and AI load swings put more on the UPS

Liquid cooling moves power off server fans and onto pumps that cannot stop. Schneider's example server uses 10 kW air-cooled and 7.5 kW liquid-cooled, and because chips heat faster than with air when flow stops, it notes "Placing pumps on a UPS helps address this issue." The GB200 NVL72 rack's in-rack CDU is rated for 250 kW with redundant power supplies and dual pumps. Air-cooled halls often kept cooling on generator power only, so a liquid-cooled hall can need more UPS for the same IT load.

AI loads also move fast. Vertiv describes swings of "80-100% to max load (usually 120-130%)" that can "swing from idle to full load in milliseconds," warns that loads that "repeatedly drive the UPS into overload" stress its electronics, and notes generators tolerate only "a limited rate of power variation." Our AI and GPU colocation guide covers the buildings built for these racks.

Sources: Schneider Electric White Paper 110, The AI Disruption; Supermicro, GB200 NVL72 datasheet; Schneider Electric, UPS design considerations for data centers; Vertiv, Advanced UPS controls for AI workloads.

Lead Times: Why Transformers and Switchgear Set the Delivery Date

New power capacity arrives no faster than its longest-lead equipment: power transformers averaged 128 weeks and switchgear 44 weeks in Wood Mackenzie's Q2 2025 survey, so a site whose power transformers are not yet on order can be roughly two and a half years from power. That is why the delivery date a provider quotes is only as good as the status of its gear: ordered, slot reserved, on site or energized.

Average lead time, Q2 2025 (weeks)Wood Mackenzie supply chain survey
Generator step-up transformers143 wk
Power transformers128 wk
Switchgear44 wk
Wood Mackenzie also reports power transformers and step-up units "seeing sustained lead times of 2 to 3 years." 128 weeks is about 2.5 years (our conversion). Waits kept growing: Wood Mackenzie's Ben Boucher told Reuters in July 2026 that generator step-up transformer lead times passed 160 weeks by the first quarter of 2026, and high-voltage circuit breakers reached 125 weeks in the second half of 2025, against 77 weeks in 2023.

Sources: Wood Mackenzie, Mind the gap (Oct 2025); Wood Mackenzie and American Clean Power, Making the connection (Sept 2025); Reuters via AJOT, US power companies scramble to secure equipment (July 9, 2026).

Equipment prices

+85%circuit breakers; switchgear +47%, transformers +35%

Wood Mackenzie's equipment price index, on a chart that runs from 2019 to 2025. Electrical steel prices had risen "by more than 20% in recent months" as of October 2025, it added.

Wood Mackenzie and American Clean Power, Making the connection (Sept 2025); Wood Mackenzie, Mind the gap (Oct 2025)

Transformers

Up to 3 yearsfor a new transformer ordered today (NEMA, 2025)

"Five years ago, that wait time was four to six weeks," NEMA's Peter Ferrell told Utility Dive in February 2025.

Utility Dive, electric transformer shortage (Feb 2025)

Generators

$4.5BGenerac data center backlog, October 2026

Up from about $1.6 billion in July 2026 and about $400 million at its fourth quarter 2025 results. Generac's 10-K says "the market for large-megawatt backup generators has become significantly supply-constrained." No maker we read publishes a lead time, so we do not give one.

Generac third quarter 2026 results (Oct 28, 2026); Data Center Frontier on Amazon and Generac (Sept 2026); Generac 2025 Annual Report and 10-K

These figures fit the ranges in our powered land guide. It gives 60 to 120 weeks for utility-grade transformers and 90 to 130 weeks or more for units of 100 MVA and up; Wood Mackenzie's 128-week average for power transformers sits in that larger-unit band, and its 44 weeks for switchgear sits inside our 40 to 65. Heavy-duty and combined-cycle gas plants carry longer waits still, 36 to 84 months by Rabobank's count, while aeroderivative turbines take one to two years and engines 12 to 24 months, covered in how data centers are powered. Ordering long-lead gear early is the main way developers protect a delivery date, which is the core of our powered shell and build to suit work.

Source: RaboResearch, The sprint: data centers are building a parallel energy system in the US.

Questions to Ask a Provider About Its Power Infrastructure

Ask for capacity after redundancy, in kW per rack and MW per hall, and for the status of the equipment behind it; a provider that answers all nine clearly has given you what you need to compare its power chain with another. Send these with your requirements, and compare the answers side by side.

Is the MW you quote critical IT after redundancy?

Why: installed UPS in N+1 built three plus one overstates leasable load by a third, and 2N doubles it.

Which UPS design, and how many modules per block?

Why: N+1, block, distributed and 2N leave 50% to 80% or more of installed UPS usable.

What is each rack circuit: voltage, amps, and A and B?

Why: a 60 A, 415 V feed gives 34.6 kW; two of them, A and B, still give 34.6 kW.

Can either feed carry my rack alone?

Why: if A and B are both loaded past half, losing one feed drops the rack.

How many generators, in what redundancy, with how many hours of fuel?

Why: POWER Magazine's Tier III guide calls for N+1 and at least 12 hours of fuel; generators carry cooling as well as IT.

Are cooling pumps and CDUs on UPS?

Why: liquid-cooled chips heat fast when flow stops, and pumps on generator only leave a gap.

Is the Tier an Uptime certification or a design claim?

Why: the Tier tells you whether maintenance or a failure will touch your load.

Is the capacity energized, or waiting on gear still on order?

Why: power transformers averaged 128 weeks and switchgear 44 in Q2 2025; ask for order and delivery dates.

How does the UPS handle AI load swings?

Why: swings past 100% of rating can push a UPS into overload and stress generators.

Need someone to read the power chain before you sign?
Send us the site or the lease, the load and your date. Our data center consulting work checks redundancy, usable capacity and equipment status for leases, acquisitions, powered shell and build to suit projects. You'll hear back within 24 hours.
Send Your Requirements

Frequently Asked Questions

What is N+1 in a data center?

N is the capacity the critical IT load needs, and N+1 adds one more unit of the same size, so any one UPS module, generator or cooling unit can fail or be taken out for maintenance while the rest carry the load. Schneider Electric calls a system N+1 when the spare is at least equal to one module. Built three plus one, three of the four UPS modules carry the load and one is spare, so 75% of the installed UPS capacity can be leased.

What is 2N redundancy in a data center?

2N means two complete, independent power systems, each able to carry the whole load. CoreSite's example: if four UPS units are needed, 2N adds four more for a total of eight. Because either side can carry everything, only half of the installed capacity is usable. 2N is the design Schneider Electric pairs with Uptime Tier IV, and 2N+1 adds one more component on top.

What is the difference between N+1 and 2N?

N+1 shares one spare among the modules on a path; 2N duplicates the entire path. N+1 protects against one module failing, while 2N protects against a whole system or path failing. For a load that needs four modules, N+1 built four plus one installs five and 2N installs eight, so N+1 leaves 80% of installed UPS usable and 2N leaves 50%. POWER Magazine's Tier III design guide describes Tier III as N+1 with two power inputs to each device, and Schneider pairs Tier III with distributed redundant designs; Tier IV is fault tolerant.

What does a UPS do in a data center?

The uninterruptible power supply carries the IT load through voltage sags and outages until the generators start and the transfer switch moves the load to them, which Schneider Electric puts at typically 10 to 20 seconds. Cloud and colocation sites are typically designed with 5 minutes of battery runtime, hyperscale sites with 1 to 2 minutes, and financial firms with 10 to 15. UPS failures are the most often cited cause of power-related outages in Uptime Institute's 2026 survey.

What is a PDU in a data center?

PDU means power distribution unit, and it names two things. A room or floor PDU takes 480 V from the UPS, steps it down where needed to 208/120 V or 240/415 V, and feeds remote power panels and branch circuits. A rack PDU is the power strip inside the cabinet that servers plug into, usually one on the A feed and one on the B feed. Schneider puts the limit of a 60 A, 240/415 V rack PDU at 34.6 kW.

How is power distributed in a data center?

Power arrives from the utility at medium voltage and passes medium-voltage switchgear, then transformers that step it down to 480 V. Low-voltage switchgear and automatic transfer switches connect the generators, UPS systems carry the load through any gap, and room PDUs, remote power panels or overhead busway split it into rack circuits. Tier III and Tier IV designs run two such paths, A and B, so each rack has two independent feeds.

What is distributed redundancy in a data center?

Distributed redundancy spreads the load across three or more UPS systems with independent feeders, so if one fails, its load moves to the others, often through static transfer switches. In a three-to-make-two design each system normally carries a third of the load, which is 66.7% of its rating, so two-thirds of installed UPS is usable. If one system fails, the design drops to N, with the other two fully loaded. Four-to-make-three reaches 75%. Block redundant designs instead keep one reserve block idle to catch any block that fails.

How long does it take to get transformers and switchgear for a data center?

About two and a half years for power transformers. Wood Mackenzie's Q2 2025 survey put the average lead time at 128 weeks for power transformers, 143 weeks for generator step-up transformers and 44 weeks for switchgear. By the first quarter of 2026, a Wood Mackenzie analyst told Reuters, step-up transformer lead times had passed 160 weeks. NEMA said in early 2025 that a new transformer could take up to three years, against four to six weeks five years before. That is why a delivery date depends on whether the gear is already ordered.

What voltage do data centers use?

In North America, power usually arrives from the utility at medium voltage, for example 13.8 kV, and transformers step it down to 480 V for the switchgear and UPS systems. PDUs then deliver it to racks at 208/120 V or 240/415 V. Schneider Electric's AI design guide says sites on 120/208 V should retrofit to 240/415 V, which carries about twice the power per amp. Inside today's AI racks, power runs at 54 VDC, and NVIDIA plans 800 VDC distribution from 2027.

How long can a data center UPS run on battery?

Only minutes, because the battery only has to bridge until the generators start and take the load, which Schneider Electric puts at typically 10 to 20 seconds. Schneider says cloud and colocation sites are typically designed with 5 minutes of battery runtime, hyperscale sites with 1 to 2 minutes, and financial firms with 10 to 15 minutes.

Cite this page

Metro Colo Advisory, Data Center Power Infrastructure, updated November 2026, https://metrocoloadvisory.com/data-center-power-infrastructure/

Free to use in news coverage and research with credit to Metro Colo Advisory and a link to this page, under a Creative Commons Attribution 4.0 license.

How much power: data center power consumption, the four kinds of MW, PUE, and a converter for any MW figure.

Where it comes from: how data centers are powered, from the grid and gas to nuclear, solar and backup.

Leasing it: wholesale colocation, leases from 500 kW to 20 MW, with pricing, providers and the RFP.

Method and corrections

Our research team read every outside figure on this page from the source linked under it in November 2026 and quotes the wording where a definition matters. Usable shares, module counts for the worked example, circuit counts and breaker nameplate figures are our arithmetic from the cited figures, and the power path diagram is a simplified two-path design. Vendor white papers describe example designs, not every building. To report an error, write to us through the contact page.