Data center power distribution chain from utility to rack

The power chain, and what it has to carry end to end

The data center power distribution chain is the path electricity takes from the utility connection to the processor on a server board, and at every stage along it the voltage steps down while the protection and redundancy step up. Power that arrives at the property as medium voltage leaves the last stage as a few volts of DC at the chip. In between sit the transformer, the switchgear, the UPS, the floor distribution, and the rack strip, each one doing a specific job and each one a place the load can be lost.

This guide is the map that ties the others together. The UPS topology and redundancy guide goes deep on the protected bus. The PDU and RPP commissioning guide goes deep on the last distribution stage. This one walks the whole path end to end, names every stage, and shows where the redundancy has to hold continuously from the service entrance to the rack.

One question runs the length of the chain: what happens to the load if this single thing fails or has to be opened for service. Ask it at the utility, the generator, the transformer, the UPS, the PDU, and the rack feed. The chain is only as redundant as its weakest stage, and the weak stage is usually the one nobody walked.

How does power get from the utility to a server?

Power reaches the server through a fixed sequence of stages, each stepping the voltage down and adding protection. The utility delivers medium voltage to the site. A transformer steps it to the building utilization voltage, commonly 480 V three-phase in North America. Switchgear distributes that and protects it. The UPS conditions it and bridges a utility loss on stored energy. From the UPS, distribution carries power across the floor through a floor PDU, an RPP, or overhead busway. The rack PDU breaks it into outlets, and the server's own power supplies make the final conversion to the low DC voltages the boards run on.

Two things ride alongside that path the whole way. A second, independent path runs in parallel on redundant designs, so the A and B feeds give every dual-corded server two sources. And a backup source, the standby generator, stands ready to pick up the whole site when the utility drops. Walk the one-line and the same pattern repeats: a source, a transfer or protection point, a step down, then the next stage.

The order is fixed, but which stages exist depends on the design. A small room might run utility to a single transformer to a UPS to rack strips. A large hall has medium-voltage switchgear, paralleled generators, multiple transformers, a 2N UPS plant, and busway over every row. Read the actual gear, not the generic block diagram.

The utility service and medium voltage

The chain starts where the utility hands power to the site, and on anything larger than a small room that handoff is at medium voltage. Medium voltage is roughly the 1 kV to 35 kV band, and the distribution levels a utility commonly brings to a campus fall somewhere in 5 kV to 35 kV depending on the region and the load. Bringing power in at medium voltage and transforming it on site keeps the conductors smaller for the same power, because higher voltage means lower current for the same kilowatts.

The service entrance is where the utility's responsibility ends and the owner's begins, usually at a metering point and a main service disconnect or medium-voltage switchgear. On a large site this is a lineup of MV switchgear, not a single can. Where the design counts on the utility for availability, the site takes more than one feed, ideally from separate utility substations on separate routes, so a single utility fault or a backhoe through one duct bank does not take the whole site dark.

Two utility feeds are worth less than they look if they come off the same substation transformer or share a duct bank for the last mile. Diverse feeds means electrically and physically separate, all the way back. Confirm the actual diversity with the utility, because a one-line can show two feeds that quietly share a common point upstream.

The generator and the backup source

The standby generators are the site's answer to losing the utility, and they enter the chain through transfer equipment, not into the UPS directly. When the utility drops, the UPS carries the load on its stored energy for the seconds to minutes it takes the generators to start, come up to speed, and accept load. Then a transfer switch shifts the site from the failed utility to the generator bus, and the engines carry the building until the utility is back.

On a small site that transfer is a single automatic transfer switch, the ATS, sitting between the utility and generator sources ahead of the load. On a large site it is generator paralleling switchgear: several engines synchronized onto a common bus, sized so the loss of one engine still leaves enough capacity, which is N+1 on the generator plant. The generator side has its own redundancy question, asked the same way as the rest of the chain.

The generator is where two timing windows have to line up. The UPS runtime has to outlast the generator start and transfer with margin, and the fuel system has to keep the engines running for the contracted ride-through, commonly measured in hours to days of on-site fuel. A UPS that runs out before the generator picks up, and a generator that starts but cannot get fuel, both end the same way. This guide stays at the chain level. Size the runtime and the redundancy against the basis of design.

The transformer steps the voltage down

The transformer takes the medium voltage from the utility or generator side and steps it down to the building utilization voltage, commonly 480 V three-phase in North America. This is the conversion that makes the rest of the chain workable. Medium voltage is efficient to bring across the site, but the switchgear, the UPS, and the distribution all run at the lower utilization voltage. The transformer is the bridge between the two.

On a large site the transformer is usually part of a unit substation: the primary MV section, the transformer itself, and the secondary LV switchgear built as one coordinated lineup. It is also a stage with its own loss and its own failure mode. It runs warm, it has a finite life, and a transformer failure takes out everything fed from it, so the redundancy question applies here too. Designs that need it run redundant transformers so either one can carry the load.

Voltages vary by region and design. Outside North America the utilization voltage is commonly 400 V three-phase, and some data center designs distribute at higher voltages to cut current and copper. The level matters less than the principle. Each transformation is a deliberate step, sized and protected, and it sets the voltage every stage below it sees.

Switchgear and the distribution switchboard

Switchgear is where the building voltage gets distributed and protected. After the transformer, power lands on the main low-voltage switchgear or switchboard: a lineup of a main breaker, a bus, and the feeder breakers that send power out to the UPS inputs, the mechanical plant, and the rest of the building. The breakers are the protection. Each one is set to clear a fault on the circuit it feeds without taking down the circuits beside it, which is the selective coordination this guide covers further down.

The switchgear is also a main metering point for the facility, where the monitoring system reads the incoming power, the major feeders, and the power quality. On a redundant design the switchgear itself is doubled or arranged so a section can be de-energized for maintenance with the load still fed from the other side. A tie breaker between two main sections lets the lineup run split for redundancy and tie together when one source is out, which is a common arrangement worth understanding before you operate it.

The thing that bites crews here is working on a lineup whose actual breaker settings and interlock scheme do not match the drawings. The protection only coordinates if the settings in the trip units match the coordination study. Confirm them as set, not as designed.

The UPS in the chain

The UPS, the uninterruptible power supply, is the stage that conditions the power and bridges a utility loss without a break. On a data center it is almost always a double-conversion online UPS. It rectifies incoming AC to DC, holds energy on a battery or flywheel, and inverts back to clean AC, so the load is always fed from the inverter and a utility disturbance never reaches it. The transfer to stored energy is instantaneous because the load was never on the raw utility in the first place.

In the chain, the UPS sits between the switchgear and the floor distribution, and its output feeds the PDUs. That is most of what this guide says about the inside of the UPS. How the modules and paths are arranged for redundancy, N+1 versus 2N, parallel versus distributed redundant, the static and maintenance bypass, and how the topology maps to an Uptime Tier, is the whole subject of the UPS topology and redundancy guide. Read that one for the protected bus. This one keeps walking the path.

What is a static transfer switch?

A static transfer switch (STS) is a solid-state switch that picks between two independent power sources and transfers the load from one to the other in a few milliseconds, fast enough that the equipment downstream never sees the gap. It uses thyristors, the SCRs, rather than mechanical contacts, which is what makes it fast. The STS feeds a load that has only one cord but needs the resilience of two sources, by choosing the live source automatically.

This is the stage that gives a single-corded load A and B resilience. A dual-corded server does not need an STS, because it already has two power supplies plugged into two paths. A single-corded device, a network switch or a piece of legacy gear, gets fed from an STS that watches both the A and B sources, stays on the preferred one until it degrades, then transfers to the other. The transfer has to be fast because a server power supply rides through only a few milliseconds of missing input.

Do not confuse the STS with the ATS. The ATS is the mechanical transfer between utility and generator, taking seconds, sitting up at the building source. The STS is the fast solid-state transfer between two already-conditioned UPS outputs, taking milliseconds, sitting down near the load. Both are transfer switches. They live at different stages and work on different time scales.

The floor PDU

The power distribution unit, the floor PDU, takes the UPS output and distributes it to the white space, usually stepping it down through an internal transformer on the way. UPS output is commonly 480 V and the racks want 208 V or 120 V, so the floor PDU carries a transformer that drops 480 V to a 208Y/120 V or 415Y/240 V secondary and breaks it into branch circuits or subfeeds. It is a floor-standing cabinet with a transformer, a main, panelboards, and metering.

In the chain, the floor PDU is the first stage of the last distribution leg, the handoff from the protected plant to the floor. Everything above it is the power plant. Everything below it heads to the rack. The transformer makes the PDU a separately derived system, which changes how it is grounded, and on a data center it is usually a K-rated transformer built to take the harmonic heat the IT load throws back at it.

The commissioning of all of that, the transformer test, the phase balance, the branch monitoring, and the A and B feeds, is the subject of the PDU and RPP commissioning guide. Read that one for the floor PDU in depth. Here it is one stage on the path: the unit that transforms and distributes the protected power onto the floor.

The RPP, closer to the racks

The remote power panel, the RPP, is a panelboard fed from the floor PDU and set out in the rows, close to the racks it serves. It has no transformer. It takes a subfeed from the PDU at the utilization voltage and breaks it into the branch circuits that run to the racks, so the long, heavy feeder does the distance from the PDU and only the short branch whips run from the RPP to the cabinets.

The RPP exists to keep the branch runs short. A single PDU in a corner feeding the whole hall would run long whips across the floor, and voltage drop and copper both climb with the distance. Pushing RPPs out into the rows keeps the branch breaker within a row or two of its rack. Whether a hall uses RPPs at all is a design choice. Some feed racks straight off the PDU, some off busway, some use the RPP as the only floor-level panel. The RPP and the PDU together, and how they get commissioned, are covered in the PDU and RPP guide.

Overhead busway and bus duct

Busway is a prefabricated overhead power run with plug-in tap-off boxes spaced along its length, and it is the flexible alternative to fixed floor PDU whips. Instead of hard-wiring a whip from a panel to each rack, the busway runs above the rows as a continuous bus, and a tap-off box plugs in over any rack to drop power to it. Move a rack, add a rack, change the density, and you move or add a tap-off box instead of pulling new conductor.

The reasons busway has taken over the dense floor are real. Overhead, it leaves the underfloor clear so the cooling air moves the way the design intended, where hard pipe and whips under a raised floor used to choke it. The tap-off boxes are factory-built with their own protection, so adding a circuit is a plug-in instead of an electrician terminating a whip on site. And the run is rated as a whole, commonly in the few-hundred to over-a-thousand-amp range, which is what the high-density and AI rows now demand.

Busway carries its own discipline. A loaded high-current busway is heavy, especially in a liquid-cooled AI hall where it shares the overhead with coolant pipe and cable tray, so the structure has to carry it. And every tap-off box is a connection point, which means it is a place to torque correctly and to thermal-scan, because a loose bus joint makes heat in the one spot nobody is looking at.

The rack PDU, the strip in the cabinet

The rack PDU is the power strip mounted in the rack that the servers actually plug into, and it is the last piece of fixed distribution before the cord. It takes the branch circuit from the whip, the RPP, or the busway tap-off, and breaks it into the outlets down the rack. Despite sharing the three letters, it is not the floor PDU. The floor PDU is a cabinet that transforms and distributes. The rack PDU is a strip that hands out receptacles.

Rack PDUs run from a basic strip to a fully intelligent unit. A metered rack PDU reports the current it carries, which is what lets the floor stay inside the branch rating and balance the phases. A switched unit can turn individual outlets on and off remotely. The trend on dense racks is the intelligent, per-outlet-metered, three-phase strip, because a rack pulling tens of kilowatts needs three-phase input and outlet-level visibility to manage it.

Single-phase versus three-phase is the first thing to get right. A low-density rack runs fine on a single-phase strip. A dense rack needs three-phase to carry the load without oversized conductors and to balance across the phases, and putting two single-phase strips where three-phase was needed is a rework nobody wants after the rack is populated. Size the rack PDU to the rack's real power and the outlet types the gear actually uses.

What is A/B power?

A and B power is the practice of feeding a load from two independent power paths, an A path and a B path, each carried all the way from a separate source through separate distribution, so either path can fail with the load still up. It is the heart of data center power resilience. A dual-corded server has two power supplies. You plug one into the A feed and one into the B feed, and the server keeps running if either entire path goes down for a failure or for maintenance.

The two paths have to be independent the whole way to mean anything. A is its own UPS, its own PDU, its own RPP or busway, its own breaker, its own rack strip. B is a second set of all of it. They only come together inside the dual-supply server, where the two cords meet at the server's internal power. The moment A and B share a stage, a single UPS, a single PDU, a single breaker, the independence is gone and one failure can drop the load even though it has two cords.

This is also the most commonly defeated protection on the floor. Plug both cords of a dual-corded server into the same strip, or feed the A and B strips from the same source, and the server still has two cords but they no longer protect it. The two cords have to land on two paths that go back to two sources. Trace them. Do not trust the labels.

Redundancy through the whole chain

Redundancy in the power chain is written as N, N+1, or 2N, and the point that matters here is that it has to hold continuously from the service entrance to the rack. N is just enough to carry the load. N+1 adds one spare unit so any one can fail or be serviced. 2N is two complete independent systems, the A and B paths, either one able to carry the whole load alone. The notations and the topology are the subject of the UPS redundancy guide. The chain-level lesson is different.

A chain is only as redundant as its least redundant stage. You can build a 2N UPS plant and still lose the load to a single transformer, a single switchgear section, a single breaker, or a single rack feed if the redundancy necks down anywhere along the path. The expensive failures are almost always a stage where the redundancy quietly dropped from two paths to one and nobody walked it. A 2N plant feeding a single-corded server through a single strip is not 2N to that server.

This is where the chain ties to the Uptime Institute Tier classification. The Tier is a statement about how much of the chain can be maintained or can fail concurrently with the load still up, and it only holds if the redundancy is continuous end to end. Map the Tier against the whole path, not just the UPS, and verify it stage by stage.

What is selective coordination?

Selective coordination is the arrangement of the protective devices so that a fault trips only the breaker nearest the fault, leaving every breaker upstream closed and the rest of the load up. Without it, a short in one branch circuit can trip a breaker several stages upstream and drop a whole switchboard, turning a one-rack problem into a one-hall outage. It is the difference between losing a circuit and losing a floor.

It works through the breakers' time-current behavior. A fault draws huge current, and each breaker has a trip curve that says how fast it opens at that current. Coordination sets the curves so the downstream device always opens first and clears the fault before the upstream device decides to act. On a critical system this is proven by a coordination study, and the trip units have to be set to match it. A breaker on the wrong curve, or a molded-case breaker where a settable one was assumed, breaks the coordination quietly.

The chain has a built-in tension between coordination and the available fault current. The closer you sit to the transformer and the larger the transformer, the more fault current is available, and the harder the devices are to coordinate. This is engineering work, not field guesswork, but the field has to verify the settings as installed match the study, because a coordinated design with the wrong settings in the trip units is not coordinated.

Grounding and bonding through the chain

Grounding runs the length of the chain alongside the power, and each stage has a job in it. The service has the main bonding and the grounding electrode system. Each transformer secondary, the floor PDU included, is a separately derived system that gets its own grounding and bonding at the point it creates a new voltage. The equipment grounding conductors and bonding tie all the metal together so a fault has a low-impedance path back to the source and the breaker trips.

Data centers add a layer on top of the code-required grounding: a signal reference grid or common bonding network under the floor and through the racks, meant to hold the equipment at a common reference so the IT gear is not fighting ground noise. That is its own subject. The chain-level point is that the grounding has to be continuous and bonded across every stage, because a break shows up as nuisance trips, noise on the gear, or, at the bad end, a fault current that does not have the path it needs to clear.

The thing that gets missed is the separately derived system at each transformer. Every step-down creates a new system that has to be grounded and bonded correctly at that point, and a PDU transformer wired as if it were still on the upstream ground is a classic finding. Confirm the bonding jumper and the grounding at each derived system, not just at the service.

Metering and monitoring across the stages

The chain is instrumented at every stage, and the system that watches it is the EPMS, the electrical power monitoring system. Meters at the service, the switchgear, the UPS, the PDU, and on a good design every branch circuit at the rack feed a common platform that shows where the power is, how the phases are balanced, and where a problem is starting. You cannot manage or troubleshoot what you cannot see, which is the reason the whole chain gets metered.

The metering also feeds the two numbers the facility runs on. Capacity management lives on the meters: how much of each stage's rating is used, so the floor fills without overloading a transformer, a PDU, or a branch. And the efficiency number, the PUE, comes from metering the total facility power against the IT power, which you can only compute if both are actually measured. A floor that meters the chain can answer where its power goes. A floor that does not is guessing.

Branch circuit monitoring at the PDU and the rack strip is where the chain-level metering meets the load. It catches a phase going out of balance, a circuit approaching its rating, or a feed that is not carrying what it should because it secretly shares a source with its partner. The detailed mapping of that monitoring is covered in the PDU and RPP guide.

Losses and efficiency through the chain

Every conversion in the chain loses a little power as heat, and those losses stack from the utility to the chip. The transformer has a loss, the UPS has a loss, the PDU transformer has a loss, and each set of conductors drops a little voltage along the way. None of it is large at a single stage, but a chain of small percentages multiplied together is the difference between a lean facility and a wasteful one.

This is what the PUE measures: power usage effectiveness, the ratio of total facility power to the power that actually reaches the IT load. A PUE of 1.5 means half again as much power goes in as comes out at the racks, spent on cooling, conversion losses, and distribution. The power chain is one of the two big levers on it, cooling being the other. A more efficient UPS operating mode, a higher distribution voltage that cuts current and copper loss, and fewer conversions all pull the number down.

Higher distribution voltage is the lever the chain gives you. Distributing at 415Y/240 V instead of 208Y/120 V to the rack, or moving to higher-voltage DC on the newest designs, cuts the current for the same power, which cuts the copper loss and the conductor size. The losses are real money over the life of the site, so the conversions and the distribution voltage are a design decision, not an afterthought.

The power chain under AI and high-density loads

AI compute has pushed rack densities from the old 5 kW to 15 kW range into 50 kW, 100 kW, and beyond, and that jump reshapes the whole chain. A rack that pulled a couple of branch circuits now pulls what a small row used to, so the transformer sizing, the UPS plant, the distribution voltage, and the rack feed all get rethought together. The chain that fed a traditional hall will not feed an AI hall without resizing from the service down.

The visible changes are at the bottom of the chain. The rack feed moves to higher current and higher voltage. Three-phase at 415Y/240 V is common and higher-voltage DC is coming, because at a 100 kW rack the only way to carry the power without absurd conductor is to raise the voltage and drop the current. Overhead busway in the few-hundred to over-a-thousand-amp range replaces floor whips, because the density and the flexibility both demand it. And the busway shares the overhead with the liquid cooling that high-density compute now requires, so the structure carries more.

The newest direction worth knowing is the move toward high-voltage DC distribution to the rack on the largest AI builds, which cuts the conversion stages and the copper for the same power. Alongside it, the Open Compute Rack approach pushes the conversion into the rack itself, with a shared power shelf feeding a 48 V DC busbar that the servers tap directly, and solid-state transformers are being explored to collapse stages further. None of these are the norm yet. The point for anyone planning a power chain now is that the AI load changes the sizing of every stage above the rack, so do not assume a hall built for traditional density has the chain to feed AI racks without a real resize.

Commissioning the chain end to end

The power chain gets commissioned in two passes: each stage proven on its own, then the whole chain proven together. Stage by stage, the transformer is tested, the switchgear breakers are trip-tested and the settings verified against the coordination study, the UPS is run through its modes including the transfer to stored energy and to bypass, the generator is load-tested, and the PDUs and rack feeds are checked for phase balance with the A and B feeds verified as truly separate. None of that proves the chain works as a chain.

The integrated test is what proves it. You pull the utility, on purpose, with load on the system, and watch the whole sequence run: the UPS holds the load, the generators start and come up, the transfer shifts the site to generator, and nothing downstream drops. Then you fail a UPS, fail a path, open a breaker that should be redundant, and confirm the load rides through each one. This is the test that finds the stage where the redundancy necked down, because it is the only test that exercises the whole path at once.

The failures the integrated test catches are the ones no single-stage test can. An A and B feed that shares a source looks fine until you fail the source. A generator that starts but cannot accept the step load looks fine until the step comes. A coordination error trips the wrong breaker only when a real fault flows. Commission the stages, then commission the chain, and budget the time for the integrated test, because it is the one that tells you whether you actually built what the one-line shows.

What to document for the chain

The record for a power chain is the one-line walked stage by stage, with the rating, the redundancy, and the as-built condition of each stage captured, not assumed. The question the record has to answer later is simple: at every stage, what carries the load if this fails, and is that still true. A chain whose redundancy was real at turnover and quietly compromised by a later change is the common way a Tier slips, and only a maintained record catches it.

Common mistakes

• Running a single path to a load that needed two, so a single failure anywhere drops it.
• Feeding both cords of a dual-corded server from the same path, defeating the A and B protection.
• Sizing one stage for today's load while the rest of the chain carries more, so the undersized stage caps the whole path.
• Building a 2N UPS plant and letting the redundancy neck down to a single transformer, breaker, or rack feed downstream.
• Treating the coordination study as done without verifying the trip-unit settings as installed match it.
• Ignoring the transformer and UPS conversion losses when sizing upstream capacity and figuring the PUE.
• Confusing the floor PDU with the rack PDU, or the ATS with the STS, and commissioning the wrong scope.
• Calling two utility feeds diverse when they share a substation or a duct bank back to a common point.

Field checklist

Standards and references

The NEC, NFPA 70, governs the installation: conductor sizing and overcurrent protection, the grounding and bonding of each separately derived system, the transfer equipment, and the working clearances on the gear. The figures and article numbers shift between code cycles, so confirm them against the edition the jurisdiction has actually adopted and any local amendments before you cite them on a submittal. NFPA 110 covers emergency and standby power systems, the generator and its transfer, where that level of standby applies.

For the availability and redundancy framework, the Uptime Institute Tier classification is the common language for how much of the chain can fail or be maintained concurrently with the load up, and TIA-942 gives a parallel set of data center infrastructure ratings. Selective coordination, fault current, and protective device behavior fall under IEEE work, including the color-book guidance on industrial and commercial power systems and protection. NETA gives the acceptance and maintenance testing standards the commissioning of the gear is measured against.

Cite the standard that controls the point, and let the project basis of design and the owner's requirements override a rule of thumb when they are stricter. None of these replace the engineer of record's calculations, the coordination study, or the manufacturer's listed requirements for the specific gear installed. Verify, do not assume.

Units, terms, and abbreviations

The chain uses a stack of abbreviations that get used loosely, so the same three letters can mean different things on one floor. Pin the term to the stage before you act on it.

MV / LV

Medium voltage, roughly 1 kV to 35 kV, and low voltage, 1000 V and below

ATS

Automatic transfer switch, the mechanical transfer between utility and generator, in seconds

STS

Static transfer switch, the fast solid-state transfer between two sources, in milliseconds

UPS

Uninterruptible power supply, conditions power and bridges a utility loss on stored energy

PDU

Power distribution unit; the floor PDU is a cabinet, the rack PDU is the strip in the rack

RPP

Remote power panel, a panelboard fed from the PDU and set near the rows

A and B power

Two independent power paths to a load so either can fail with the load up

N, N+1, 2N

Redundancy notations: just enough, one spare, and two full independent systems

EPMS

Electrical power monitoring system, the platform that meters the chain

PUE

Power usage effectiveness, total facility power divided by IT load power

Related tools

Calculators and readiness checks for this work

Compare your options

FAQ

People also ask

• What does N+1 redundancy mean?
• What is the difference between a PDU and an RPP in a data center?
• Why are data centers moving to DC power for AI racks?
• What is a rack PDU?
• What is a commissioning hold point on a UPS?