A data center generator has one job: carry the critical load from the moment the UPS batteries run out until the mains comes back, for as many hours as the site has promised its clients. It is sized on the whole site load rather than the IT load, it has to accept that load in a few seconds, and it has to be tested and maintained without ever dropping the servers. Everything else in a data center power specification follows from those four sentences.
Why the generator is only the second half of the chain
A data center does not run on one power source, it runs on a chain, and each link covers a different length of time. The UPS covers the seconds between the mains failing and the generator being ready; the generator covers the hours until the mains returns; the fuel covers the generator. Reading the chain this way is what stops the two classic mistakes: buying a generator that is too slow to catch the load, and buying one that runs dry before the outage ends.
| Layer | What it covers | The limiting figure |
|---|---|---|
| Utility supply | Normal operation | Availability of the local network, not the site |
| Switchgear and ATS | Detects the failure and transfers | Detection and transfer time, typically seconds |
| UPS and batteries | Ride-through while the generator starts | Battery autonomy at full load, typically minutes |
| Diesel generator | The whole outage after the batteries | Fuel autonomy in hours, and load acceptance in seconds |
| Fuel system | Everything the generator does | Litres on site, and fuel quality over time |
The battery is a bridge, not a source. A site with ten minutes of battery autonomy and a generator that takes twenty seconds to close and accept load is comfortable. The same site with a generator that fails to start on the first attempt is depending on a maintenance regime, and that is a different risk. This is why the sequence is tested as a sequence, not as separate items: mains off, UPS carries, generator starts, generator closes, generator accepts, batteries recharge.
Start with the load, not the kVA
Data center generator sizing starts from the IT load in kW and then adds everything the site uses to keep that IT load running. Modern sites are usually designed to a power usage effectiveness (PUE) of roughly 1.2 to 1.6, which means the cooling and the losses add between 20 % and 60 % on top of the IT load; older or hotter-climate sites can be worse. A 500 kW IT load in a site designed to 1.4 is a 700 kW electrical load before any margin.
| Step | Figure | Where it comes from |
|---|---|---|
| IT load | 500 kW | The racks, measured or specified |
| UPS and distribution losses | 25 to 40 kW | Typically 5 % to 8 % of the IT load |
| Cooling and auxiliaries | 175 to 260 kW | The balance of the PUE figure |
| Total electrical load | 700 to 800 kW | Sum of the three lines above |
| Load in kVA at 0.8 power factor | 875 to 1000 kVA | kW divided by 0.8 |
| Generator to quote | 1000 kVA and up | Chosen so the load sits below the rating, not on it |
Two consequences follow from that arithmetic. First, the generator is not a rounding error next to the UPS; it is usually the largest single electrical machine on the site. Second, the number you quote from is a rating, not a kVA figure on its own: a 1000 kVA set declared for standby duty has no overload allowance, while a prime-rated set of the same kVA may be allowed 10 % for one hour in twelve. Which one the site needs depends on whether the generator is also expected to carry the load for long, planned periods and on how many hours a year it is expected to run.
Six numbers to fix before the generator is quoted
A data center enquiry that says only “1000 kVA” cannot be answered properly, because five other numbers move the machine. They are worth collecting once, on one page, before any supplier is asked for a price.
| Figure | Typical range | Why it changes the machine |
|---|---|---|
| Fuel autonomy | 24 to 48 hours, sometimes 72 | Sets the day tank and bulk tank size, the refuelling contract and the bunding |
| Largest load step | Chiller and UPS block load | Decides whether one set alone can accept the step, or whether the load has to be staged |
| Ambient temperature | 45 to 50 °C in Gulf summer | Derates the engine and the alternator and sizes the radiator |
| Altitude | Sea level on the coast | Derates naturally aspirated and turbocharged engines above roughly 1000 m |
| Annual running hours | Testing plus outages, low hundreds | Decides whether the standby rating is valid or a prime rating is needed |
| Redundancy | N+1 or 2N | Decides how many sets, how each one is sized, and how the bus is arranged |
The reference conditions matter as much as the figures themselves. A rating is declared against a reference ambient and altitude, so a 1000 kVA set in a 50 °C plant room is not the same machine as the same kVA figure at 25 °C. Ask for the derating calculation on the datasheet rather than a sentence in an email, and read it together with the duty rating the price is quoted against.
N+1 or 2N: what the topology changes
Redundancy in a data center is described as a topology, and the topology decides how many generators there are and how each one is sized. The letters are not marketing: they change what happens during maintenance as well as during a failure.
| Topology | What it means at the generator | What it protects | Cost |
|---|---|---|---|
| N | Enough sets to carry the load between them | Nothing beyond the load itself | Lowest |
| N+1 | One extra set beyond the load | A single set failing, and one set being serviced | One extra set |
| 2N | Two independent sets of plant, each able to carry the load | A failure in either path, including switchgear | Doubles the plant |
| 2(N+1) | Two independent paths, each with its own spare | Failure plus maintenance on either side | Highest |
Two practical points follow. Each generator in a redundant arrangement is usually sized to carry the full critical load on its own, which is why an N+1 site often has two sets of the same rating rather than one large and one small. And redundancy only covers what is actually duplicated: two generators sharing one fuel tank, one control system or one bus section is not redundancy, it is one failure away from the same outage. Where two sets are arranged to run together, the electrical conditions are the ones in generator paralleling, and the transfer equipment is the subject of ATS versus AMF.
The failures that actually take a data center down
Over years of running standby plant, the failures that reach the servers are rarely dramatic. They are the boring ones, and they are almost all avoidable by specification and routine:
- Fuel. Water, microbial growth and aged diesel are the most common reason a set that started perfectly on test will not carry load on the day. Filters block, injectors suffer, and the engine derates or stops.
- Batteries and chargers. A starting battery that has lost capacity, or a charger that has failed quietly, means the generator never turns over. The set is fine; the start is not.
- Cooling. A radiator that is fouled with dust, or a plant room that recirculates hot air, turns a 50 °C day into a derated or tripping set. Airflow is part of the specification, not an installation detail.
- Running too lightly. A generator that spends its life at 10 % to 20 % of rating runs cold, builds carbon and unburnt fuel, and eventually smokes and loses capacity. Wet stacking is a maintenance symptom with a sizing cause.
- The bus and the switchgear. Single points of failure hiding inside an apparently redundant arrangement: one fuel line, one control supply, one breaker that both paths pass through.
- Control and protection settings. A set that trips on a load step because the protection is set for a smaller machine, or a transfer that re-transfers too early when the mains flickers.
None of these is a reason to distrust diesel generators in a data center. They are the reason the specification includes fuel polishing, battery monitoring, load testing and protection coordination rather than stopping at the kVA figure.
How the generator is proved without dropping the load
A data center generator is only as good as its last test, and the tests are designed so that the IT load never sees them. The right pattern is scheduled, recorded, and repeated at fixed intervals.
| Test | Typical interval | What it proves |
|---|---|---|
| Start and run off load | Weekly or monthly | Starting system, fuel supply, cooling, alarms |
| Transfer test | Monthly | Detection, transfer and re-transfer settings |
| Load bank test | Annually | Performance at 50 % to 100 % load, cooling, fuel system |
| Fuel quality test | Annually, or after any delivery | Water, particulates, microbial growth |
| Protection test | Annually | That the protection still trips at the settings it did at commissioning |
| Black building test | At commissioning, then as agreed | The whole chain from mains loss to IT load on generator |
The load bank test is the one that is most often skipped and most often regretted, because it is the only test that puts heat into the engine and the radiator. A generator can pass years of monthly no-load runs and still fail on the first real transfer. Where the site cannot isolate the IT load for a full black building test, the load bank is what fills the gap. The wider routine behind these tests is set out in diesel generator maintenance schedule.
What to put in the specification
| Item | What to state |
|---|---|
| Load | IT load in kW, plus the designed PUE, and the resulting site load |
| Rating | Standby, prime or continuous, with the expected annual running hours |
| Voltage and frequency | For example 400/230 V, 50 Hz, three phase, four wire |
| Topology | N+1 or 2N, and whether sets run singly or in parallel |
| Load step | The largest single step, and the starting method of the chillers |
| Autonomy | Hours at full load, and the fuel storage arrangement |
| Site conditions | Ambient, altitude, plant room airflow, and noise limit at the boundary |
| Enclosure | Containerized or purpose-built plant room, and the acoustic requirement |
| Controls | BMS or DCIM interface, remote monitoring, event logging |
| Testing | Factory tests and documentation required before shipment |
When a diesel generator is the wrong answer
The generator is the default for data center backup because it is the only mature technology that stores hours of energy on site at a workable density. It is still the wrong answer in a few situations, and saying so is part of the specification:
- No space or no permit for fuel. A site in a dense urban area with no bunded tank space and no refuelling access cannot support the autonomy it wants to advertise.
- Emission or noise limits that a diesel set cannot meet. Some jurisdictions cap NOx or require a permit for the running hours the site needs for testing.
- Outages that a UPS can cover on its own. If the real requirement is five minutes of ride-through, a battery or flywheel system may be cheaper than a generator plus fuel system plus maintenance contract.
- Maintenance access that does not exist. A rooftop set with no crane route for a replacement engine is a set that will eventually be run past its overhaul hours.
Where the site does need diesel, the form of the set follows the site: a containerized set for a plant yard with limited building work, an open skid inside a purpose-built plant room where airflow can be engineered, and parallel sets where the redundancy topology requires them. The data center application page describes how the same requirement is expressed at each power band.
Questions
How many hours of fuel should a data center generator have?
Common requirements are 24, 48 or 72 hours of autonomy at full load, with 24 to 48 hours the usual range for colocation. The figure belongs to the promise the site makes to its clients, so it has to be decided by the operator rather than the generator supplier. As a check: a 1000 kVA set at 0.8 power factor carries 800 kW, burns roughly 200 litres an hour at full load, and therefore needs about 4800 litres for 24 hours – before any allowance for the tank that cannot be used and for the days when refuelling is delayed.
Does the generator carry the whole site load or only the IT load?
The whole electrical load that the site needs in order to keep the IT load running: IT, UPS losses, cooling and auxiliaries. In a typical modern site the cooling and losses add 20 % to 60 % on top of the IT load, so a generator sized on the IT load alone is undersized. The exact figure is the site’s designed PUE, and it should be stated in the enquiry.
Can a generator start fast enough to cover a UPS?
Yes, if the sequence is designed and tested. A diesel set typically starts and reaches speed in about ten seconds, and closes and accepts load within a few seconds of that; a UPS with ten minutes of battery autonomy covers that comfortably. The risk is not the normal start, it is the start that does not happen on the first attempt – which is why the transfer test and the start battery are part of the routine, not the commissioning.
Do data centers need N+1 or 2N generators?
N+1 covers a single generator failing or being serviced while the load continues; 2N duplicates the entire path so that a fault in the switchgear of one path cannot affect the other. The choice is driven by the site’s availability commitment and its clients, and it is usually decided before the generators are specified. What matters at the generator is that each unit is sized to carry the load alone where the topology requires it – which is why an N+1 site often has two identical sets rather than one large one.
What size generator for a 500 kW IT load?
Start from the site load, not the IT load. A 500 kW IT load in a site designed to a PUE of 1.4 is a 700 kW load, which is 875 kVA at 0.8 power factor – so the quotation starts at a 1000 kVA set rather than a 700 or 800 kVA one, and a larger site with a bigger cooling load moves towards 2000 kVA. The number also has to be checked against the largest load step, which is usually a chiller or a UPS block load rather than the steady state.
Is a load bank test really necessary?
For a data center, yes. Monthly no-load running proves that the set starts; it does not prove that the engine and radiator can reject the heat of a full load, that the fuel system delivers at rated power, or that the protection is coordinated. An annual load bank test at 50 % to 100 % of rating is the standard way to prove those things on a site where the IT load cannot be used as the test load.
Send the load, the topology and the autonomy
IT load, designed PUE, redundancy topology and fuel autonomy are enough for us to size the sets and quote the plant the first time.
Written and maintained by the KVA Grid sales engineering desk. PUE ranges, autonomy figures and test intervals are industry practice, not standard limits; the site design, the fuel supplier’s specification and the manufacturer’s manual govern. Reviewed 2 October 2026.
