Data center power: UPS, PDU and IEC connectors explained
Power is the leading cause of hosting outages. Understanding the chain that runs from the grid to your server, and knowing where redundancy actually happens, is often worth more than comparing rack prices. Here are the links and the right reflexes, without needless jargon.
The power chain, link by link #
Power comes from the grid, backed by a generator for longer outages. It passes through a UPS, which absorbs micro-cuts and surges, then through a PDU that distributes it inside the rack, before reaching your servers. Every link can fail: resilience comes from redundancy, not from any single device.
The UPS: the safety net during an outage #
The UPS supplies power for the seconds or minutes the generators need to start, and smooths electrical quality day to day. You size it on two criteria: power (in watts), with a 20 to 25% margin, and runtime, meaning how long it lasts on battery under the real load.
Not every UPS is the same machine #
People say « the UPS » as if there were only one. There are five families, and they are nothing alike. What separates them is a simple question: does the current reaching your servers pass through the UPS electronics all the time, or only when the mains fails? In the first case it is called double conversion, and the output is clean in all circumstances. In the second the UPS waits in standby and switches over when it detects a fault, with the transfer time that implies. The table below gives the power range of each family, and the ranges speak for themselves: the one under your desk and the one feeding a data hall are not in the same category.
| Family | Power range | Inverter always on | Voltage correction |
|---|---|---|---|
| Standby | 0 to 0.5 kVA | no | no |
| Line-interactive | 0.5 to 5 kVA | depends on the model | depends on the model |
| Standby-ferro | 3 to 15 kVA | no | yes |
| Double conversion | 5 to 5,000 kVA | yes | yes |
| Delta conversion | 5 to 5,000 kVA | yes | yes |
A quarter of a second costs a quarter of an hour #
This is the figure that justifies a UPS on its own, and it always surprises. A 0.25 second dip on the mains is enough to interrupt the work of a computing centre for more than fifteen minutes, the time it takes for operating systems to reboot and applications to come back. The ratio between the fault duration and the incident duration is roughly one to three thousand. It is also why the statistic that matters is not the number of outage hours per year but the number of micro-outages: they do the most damage for the least time lost on the grid. Source: Dorin O. Neacșu, *Telecom Power Systems*, CRC Press, 2017, chapter 14, page 369.
The UPS that spins, and the current nobody sees coming #
Two operational realities that guides tend to skip. First, not every high-power UPS is electronic: there are rotary UPS systems, where a flywheel coupled to a diesel engine holds the voltage while the engine starts. The flywheel covers the few seconds needed, then a clutch engages the diesel. Several European data centres run on this principle, and it changes the conversation about runtime: there is no battery to replace every five years. Second, rising density has a consequence rarely discussed: fault current. A 1 megawatt data centre introduces more than 50,000 amperes of fault current on the low-voltage side of its transformer, and more efficient designs, with less resistance in the cabling, push it higher still. That is what underpins the authorisation procedures and the protective equipment you will see worn in the hall. What it decides for you: if you ask for high density, also ask what intervention conditions come with it, because remote hands on a very dense rack are not priced like remote hands on a 3 kVA rack.
Redundancy: N, N+1, 2N #
N is just the capacity needed to carry the load, with no safety net. N+1 adds a spare module: losing one module cuts nothing. 2N duplicates the whole chain into two independent paths, A and B: an entire path can fail with no impact. 2(N+1) is the most resilient, and the most expensive. A serious data center states its level clearly.
The PDU: distributing inside the rack #
The PDU (power distribution unit) is the rack's power strip, mounted close to the equipment. Vertical (0U format, up to around forty outlets, without using up usable space) or horizontal (1U or 2U). Single-phase or three-phase depending on density. It is what spreads power to each server.
Basic or intelligent PDU #
A basic PDU just distributes. An intelligent (managed) PDU measures consumption outlet by outlet, reports it remotely (SNMP, LCD screen) and helps balance load across phases. It becomes essential as soon as you want to manage density per rack or bill energy back. Two useful variants: the maintenance bypass, which lets you replace the UPS without shutting down, and the ATS, which switches automatically between two sources for single-corded equipment.
Connectors: C13, C14, C19, C20 #
The international IEC 60320 standard (up to 250 V and 16 A) defines the common connectors. The C13/C14 pair powers most servers and network gear. The higher-current C19/C20 pair serves power-hungry equipment such as dense racks or GPU servers. C13 and C19 come in locking versions to prevent accidental disconnection. Knowing your outlet count and your power per rack avoids nasty surprises.
Dual (A/B) power: real resilience in colocation #
A server with two power supplies, plugged into two PDUs fed by two independent power chains A and B, keeps running even if a whole chain fails. This is the standard for demanding colocation. Before signing, check that the data center provides two genuinely independent feeds, and what density (in kW per rack) it allows.
What it means for your hosting #
Beyond the rack price, four questions decide real reliability: the redundancy level (N+1, 2N), whether a dual A/B feed exists, the allowable density in kW per rack, and the type of PDU provided (basic or intelligent). These answers are worth more than a price alone, and a comparator lets you line them up.
FAQ #
The C13/C14 pair covers most servers and network gear. The higher-current C19/C20 pair (up to 16 A) serves power-hungry equipment such as dense racks or GPU servers. On the equipment side, the male plug goes into the PDU.
N+1 adds a spare module to the needed capacity: it survives the loss of one module. 2N duplicates the whole chain into two independent paths A and B: it survives the loss of an entire chain. 2N is more resilient, and more expensive.
Usually no. The data center protects the whole infrastructure, and your servers benefit through the two A and B feeds. In colocation, dual power is the real protection, not a UPS added inside the rack.
A basic PDU distributes power. An intelligent one measures consumption per outlet, is monitored remotely and helps balance load. It becomes necessary as soon as you want to manage density per rack or bill energy precisely.
Written on 1 September 2026.
From reading to comparing: relevant data centers
Estimate my cost → Compare data centers