Power: Utility, UPS and Generators

The power path

Every server in a data center sits at the end of a chain: the utility feed, an automatic transfer switch (ATS) that can also select the generator, the uninterruptible power supply (UPS), and power distribution units (PDUs) in the room and in the rack. Each link can fail or need maintenance, so serious facilities build the chain twice, as the redundancy page explains. The diagram shows a 2N design with two complete paths, A and B.

Kestrel Energy's colocation racks receive A and B feeds from the provider; Kestrel only sees the last link, the rack PDUs. Its small head-office data center has one utility feed, one UPS and one generator: a single path, with every maintenance job planned in a window (adequate for DR, not for production).

Utility feeds

A utility feed is the connection that brings power from the public grid into the facility. Large data centers connect at medium voltage (typically 10 or 20 kV in the Netherlands) and have their own transformers to step it down to the 400/230 V that the equipment uses. Very large sites connect at higher voltages still and may build their own substation.

Resilient sites get two feeds, ideally from different substations, so that a fault in one part of the grid does not cut both. Even then, the grid has power quality events that matter to IT: short voltage dips when a fault elsewhere is cleared, spikes when large loads switch, and occasional interruptions. The UPS absorbs those; the generator covers long outages.

Getting a feed at all is the hard part today. In much of the Netherlands, large consumers wait for new or bigger connections because of grid congestion, and an existing connection with spare capacity has become one of the most valuable assets a data center has. How the grid itself works is covered in the power grid.

Uninterruptible power supplies

A data center UPS does the same job as the small unit under a desk, at a scale of hundreds of kilowatts per system: it keeps the load running, without a gap, from the moment the utility fails until the generators take over, and it filters the power quality problems of the grid. The international standard IEC 62040-3 classifies UPS designs by what the output depends on.

Double conversion

A double-conversion UPS converts all incoming alternating current (AC) to direct current (DC) in a rectifier, keeps the batteries charged on that DC bus, and converts it back to clean AC in an inverter that feeds the load all the time. When the utility fails, the rectifier stops and the batteries feed the inverter: the load sees no transfer at all. The output voltage and frequency are generated by the UPS, independent of the grid, which also filters dips, spikes and frequency deviations. The price is a few percent of conversion loss, around the clock, which turns into heat that must be cooled; many units offer a high-efficiency mode that bypasses the double conversion while the grid is clean, at the cost of a short transfer when it is not.

The static bypass is a safety valve: if the inverter fails or the load suddenly exceeds what it can deliver (for example a short circuit downstream), the UPS switches the load straight to the input power within milliseconds. Larger installations also have a separate, manually operated maintenance bypass that lets technicians isolate the whole UPS for work. While the load runs on bypass it has no battery protection, which is why bypass operations are planned changes.

Line-interactive

A line-interactive UPS passes utility power through to the load in normal operation and corrects sustained low or high voltage with an autotransformer, without using the battery. When the power fails or drifts too far, it switches to its inverter and battery, with a short transfer time of a few milliseconds that most power supplies ride through. It is more efficient and cheaper than double conversion, which makes it common in network closets and small server rooms, but it does not isolate the load from frequency deviations and adds a transfer at the worst moment. Data halls use double conversion.

Batteries and runtime

UPS batteries are sized to bridge minutes, not hours: long enough for the generators to start, take the load and, if the first start fails, try again. Valve-regulated lead-acid (VRLA) batteries are the traditional choice; lithium-ion batteries, which are smaller and last longer but need their own fire safety measures, are increasingly common. All batteries lose capacity with age and temperature, so they are monitored per block or per string and replaced on a schedule. A battery string that tested fine five years ago may deliver far less today.

# Rough UPS runtime: usable battery energy x inverter efficiency x battery
# health, divided by the load. Real batteries deliver less energy at high
# discharge rates, so the vendor's runtime table always wins.
BATTERY_KWH = 40.0  # usable energy of the battery string when new
EFFICIENCY = 0.95   # DC to AC conversion in the inverter
TARGET_MIN = 5      # cover generator start, retries and transfer


def runtime_min(load_kw, health):
    return BATTERY_KWH * EFFICIENCY * health / load_kw * 60


print("Load      new  aged (80%)")
for load_kw in (100, 150, 200, 300, 400):
    new = runtime_min(load_kw, 1.0)
    aged = runtime_min(load_kw, 0.8)
    flag = "" if aged >= TARGET_MIN else "  below target"
    print(f"{load_kw:3} kW {new:5.1f} min {aged:5.1f} min{flag}")

Load      new  aged (80%)
100 kW  22.8 min  18.2 min
150 kW  15.2 min  12.2 min
200 kW  11.4 min   9.1 min
300 kW   7.6 min   6.1 min
400 kW   5.7 min   4.6 min  below target

Two lessons from even this rough model. Runtime falls as load grows, so a UPS that was comfortable when the hall was half full can be marginal when it is full. And battery aging eats into the margin silently; the target is met with new batteries and missed with old ones. The 5-minute target is an example; each site sets its own.

Generators

A standby generator is an engine-driven generator, usually running on diesel, that starts automatically when the utility fails and supplies the facility for as long as it has fuel. The US standard NFPA 110, a common reference, calls a system that restores power within 10 seconds "Type 10"; data center generators typically start and take load in about that time. The UPS only has to bridge that gap, plus a margin for a failed first start.

Fuel is stored on site in tanks sized for a design runtime, often a day or more, backed by contracts with fuel suppliers for priority deliveries during a long outage. Fuel ages, so it is tested and treated. Generators are also tested regularly: started, run under load (sometimes on a separate load bank), and their starting batteries checked, because the most common generator failure is simply that it does not start when needed.

Transfer switches

Something has to decide which source feeds the load, and switch.

• An automatic transfer switch (ATS) monitors its normal source and, when it fails, switches the load to an alternate source, typically from the utility to the generator, and back once the utility is stable again. Facility ATSs are usually electromechanical, so the transfer is not instantaneous; that is fine because the UPS sits downstream and carries the load during the switch.
• A static transfer switch (STS) does the same job with solid-state switches (thyristors) instead of moving contacts, switching between two UPS-backed sources within a few milliseconds, fast enough that IT equipment does not notice. STSs are used downstream of the UPS, for example to give single-corded equipment a second source.

At rack level, a small rack ATS takes the A and B feeds and gives a single-corded device (an older appliance, a console server) one output that switches between them. It turns a device with one power supply into one that survives the loss of a feed, though the switch itself becomes a single point of failure for that device.

A/B feeds

An A/B power feed is a pair of independent power distribution paths to a rack, each through its own UPS and distribution, so that dual-corded equipment keeps running when either path fails or is switched off for maintenance. A server with two redundant power supplies connects one to the A-side rack PDU and one to the B-side PDU.

In normal operation, a dual-corded device draws roughly half its power from each feed. That creates the most common trap in data center power: both PDUs show a comfortable load, but when one feed fails, the other suddenly carries everything. If that exceeds its breaker, the surviving feed trips too, and a redundant design turns a single failure into a full outage. The rule: the sum of both feeds must stay within what one feed can carry.

# A/B feed check: in normal operation a dual-corded device draws about
# half its power from each feed. If one feed fails, the other must carry
# every dual-corded device on its own; single-corded devices on the lost
# feed simply go dark.
FEED_LIMIT_W = 230 * 16 * 0.8  # one 16 A single-phase rack PDU, 80% margin

racks = {
    "C07": [("app-01", 540, "AB"), ("app-02", 540, "AB"),
            ("db-01", 780, "AB"), ("leaf-a", 310, "AB"), ("leaf-b", 310, "AB")],
    "C08": [("gpu-01", 1450, "AB"), ("gpu-02", 1450, "AB"),
            ("db-02", 780, "AB"), ("old-nas", 420, "A")],
}

print(f"Limit per feed: {FEED_LIMIT_W:.0f} W")
for rack, devices in racks.items():
    normal = {"A": 0.0, "B": 0.0}
    for name, watts, cords in devices:
        for feed in cords:
            normal[feed] += watts / len(cords)
    print(f"Rack {rack}: normal A {normal['A']:.0f} W, B {normal['B']:.0f} W")
    for lost, survivor in (("A", "B"), ("B", "A")):
        load = sum(w for _, w, cords in devices if survivor in cords)
        dark = [n for n, _, cords in devices if survivor not in cords]
        status = "ok" if load <= FEED_LIMIT_W else "OVERLOAD, breaker trips"
        extra = f"; dark: {', '.join(dark)}" if dark else ""
        print(f"  feed {lost} lost: {survivor} carries {load} W  {status}{extra}")

Limit per feed: 2944 W
Rack C07: normal A 1240 W, B 1240 W
  feed A lost: B carries 2480 W  ok
  feed B lost: A carries 2480 W  ok
Rack C08: normal A 2260 W, B 1840 W
  feed A lost: B carries 3680 W  OVERLOAD, breaker trips; dark: old-nas
  feed B lost: A carries 4100 W  OVERLOAD, breaker trips

Rack C08 looks healthy in normal operation: feed A is at 77% of its limit, B at about 63%. Yet losing either feed overloads the other, and the single-corded network-attached storage (NAS) box on feed A has no protection at all. Colocation providers usually state this rule in the contract: on a redundant A/B pair, the combined draw must stay within the rating of a single feed.

kW, kVA and power budgets

As on a desktop UPS, data center power has two ratings. kVA (kilovolt-amperes) is apparent power, volts times amps; kW (kilowatts) is the real power that does work and becomes heat. The ratio is the power factor. Modern server power supplies have a power factor close to 1, so kW and kVA are nearly equal for IT loads, but UPSs, generators and transformers are often rated in kVA with a kW limit next to it. Always compare kW with kW.

The figures assume a power factor of 1. The 80% column reflects a common design margin for continuous loads; the exact rule depends on the country's electrical code and the provider's contract.

A rack's power budget is the power it is allowed to draw, set by the circuits feeding it and by the cooling available for that position; the idea is the same as the power budget for a single computer, at a larger scale. Budgets are planned on measured draw, with a margin, not on the nameplate ratings of power supplies, which can be two or three times the real load.

Densities are rising fast. In the Uptime Institute's 2026 survey, the average typical rack density passed 11 kW, and 24% of respondents had at least some racks of 30 kW or more. AI training racks drawing well over 100 kW are already shipping, and vendors plan racks of several hundred kilowatts. Such racks need dedicated high-current feeds and usually liquid cooling; see cooling and airflow.

What happens in an outage

A utility failure at a well-designed site, step by step:

Things that go wrong, and that designs and tests try to catch: a generator that fails to start (a flat starter battery, a generator controller left in manual mode), batteries that have aged and run out before the generator is on line, an overloaded A/B pair that trips when the load shifts, and cooling that restarts too slowly while dense racks heat up. The disaster recovery page covers what to do when the whole site is lost.

Working safely around power

Data center electrical systems carry currents that can kill, and a short circuit in switchgear can cause an arc flash: an explosive release of energy that burns and blinds. That is why:

• Only qualified, authorized electrical staff operate switchgear, UPS systems, transfer switches and distribution panels, following written switching procedures.
• Before work, equipment is isolated and secured with lockout/tagout: a personal lock and tag on the isolator, so nobody can switch it back on while someone is working.
• IT staff connect equipment only to the outlets assigned to them, never daisy-chain power strips, and report damaged cords or tripped breakers instead of resetting them.
• Power work is a change: planned, approved and communicated, as in change enablement, because even a correct switching operation briefly removes redundancy.

The general rules for electrical work on computers are on the safety and handling page.