What is a data-center-rated generator set?

The most common procurement error in generation is comparing two sets on kW without checking what rating that kW is quoted at. The same iron produces a different published number depending on the duty it is warranted for.

The pattern is always the same: the more hours and the higher the average load a rating allows, the lower its kW number for a given machine. Standby is the highest figure and the most restrictive duty. Continuous is the lowest figure and the least restrictive duty.

The spread is material. Cummins publishes the QSK95 at 3000 kW standby at 60 Hz and 2500 kW on its Data Center Continuous rating — the same set, about a 17% difference, purely because the duty changes. Sizing a continuously-running installation off a standby number is how a plant ends up chronically overloaded.

ISO 8528-1 is the international standard the rating vocabulary comes from (Cummins, ISO 8528-1 explained).

• ESP — Emergency Standby Power. For utility outages only. Variable load, roughly 200 hours per year, average load factor around 70% of the rating, and no overload capability. This is the number on most standby genset data sheets.
• LTP — Limited-Time running Power. Full rated load, but only about 500 hours per year. Used for scheduled duty like utility-contracted curtailment.
• PRP — Prime Running Power. Unlimited hours at a variable load, average load factor around 70%, with 10% overload available for one hour in every twelve. The rating for a machine that is the site’s working power source.
• COP — Continuous Operating Power. Unlimited hours at a constant load, 100% of rating for 100% of the hours. No overload. The rating for base-load generation.

The variable-versus-constant distinction between PRP and COP matters more than the hours. Prime assumes the load moves around and averages well below the rating; continuous assumes it does not move at all.

This is the one buyers most often assume is a standard. It is not. Data Center Continuous (DCC) — and Caterpillar’s equivalent mission-critical rating — sits outside ISO 8528-1 and is defined by the manufacturer.

What it is meant to express is a duty the ISO ratings do not cleanly cover: a set that may have to run for an unlimited number of hours, at a load that varies, at up to 100% of its rating, in a facility where losing power is not an acceptable outcome. Prime rating assumes a 70% average load factor that a fully-loaded data hall will exceed. Continuous rating assumes a load that does not vary, which a data hall’s does. DCC is the OEM’s answer in between.

Because it is proprietary, the practical guidance is:

• Ask which rating a quoted kW belongs to, and get it in writing.
• Ask what load profile and ambient the rating assumes, and how it derates with altitude and temperature.
• Do not compare one manufacturer’s DCC number against another’s prime number as though they mean the same thing.
• Check what the warranty says about hours and load factor, because that is where the rating has teeth.

Sets carrying data-center ratings are available through Pantheon in the gensets catalog.

Federal emissions rules for stationary engines split on use, not on nameplate. This is the highest-consequence detail in the entire specification.

Emergency standby engines are exempt from EPA Tier 4 Final and may be certified at Tier 2 — much simpler machines, no aftertreatment. That exemption is bought with a limit: unlimited hours during a genuine emergency, but no more than 100 hours per year for maintenance, testing, and readiness, and explicitly no peak shaving or non-emergency demand response. Some air districts, notably in California, cut the testing allowance to 20–50 hours.

Any non-emergency use reclassifies the engine. Peak shaving, demand-response participation, prime power, or simply running the site because a grid connection has not arrived all pull the unit into Tier 4 Final, which in practice means a diesel particulate filter, selective catalytic reduction, and diesel exhaust fluid. Aftertreatment cannot be bypassed if it fails. The package is physically larger, needs DEF storage and resupply, and adds exhaust backpressure the engine has to be sized around.

The failure mode is buying a fleet of Tier 2 emergency-rated sets for a behind-the-meter site that is, in operational reality, going to run them. Decide the duty first, then buy the emissions tier that matches it.

Rating tells you what the set can carry. Block-load performance tells you whether it can pick it up.

NFPA 110 Level 1, Type 10 is the familiar benchmark: power restored within ten seconds of utility failure. A correctly sized diesel set accepts its full rated load in a single step inside that window (Kohler, NFPA 110). ISO 8528-5 defines the performance classes that describe how well it does so — G1 through G3 — bounding transient voltage dip, frequency dip, and recovery time. Data centers specify G3, because UPS rectifiers and cooling VFDs will drop off a bus that sags or slews too far.

The engineering constraints behind those numbers:

• Turbocharger lag. A large diesel cannot make boost instantly. The first load step is limited by how fast the turbo spools, which is why very large sets sometimes need the load applied in two steps rather than one.
• Gas engines accept much smaller steps than diesels. A lean air-fuel mixture has little margin before misfire or knock. This is the main reason a gas genset is usually not a drop-in replacement for a diesel standby machine.
• The real load is not the IT load. Chiller motors, pump starts, and transformer inrush drive step sizing far more than the servers do. The IT load is behind UPS and appears to the generator as a relatively well-behaved rectifier load; the mechanical plant is what slams the bus.
• Load rejection is the mirror problem. Dropping a large block causes overspeed and voltage overshoot; ISO 8528-5 bounds those too, and load-rejection testing belongs in the commissioning script alongside load acceptance.

Very few critical facilities run one generator per load. They run a plant.

Paralleling means synchronizing multiple sets onto a common bus — matching voltage, frequency, and phase angle before closing each breaker — then sharing load between them. Digital controls handle first-start arbitration (which set closes onto the dead bus), isochronous kW sharing so all units carry proportional real power, and reactive sharing so none of them fight over kVAR.

What that buys:

• Redundancy as a pooled reserve. A 30 MW hall served by eleven 3 MW sets in N+1 has one spare shared across the whole load. The same hall with dedicated per-block sets needs far more spare capacity for the same availability.
• Maintainability. A unit comes off the bus for service while the plant keeps running.
• Better loading. Sets can be sequenced on and off so the running units sit near their efficient load point instead of all idling lightly.

Common topologies include a common-bus plant, catcher or block-redundant arrangements where a spare set backs a group, and isolated-parallel schemes where each set feeds its own bus through a choke and shares through a common ring.

One scheduling point that gets missed: the paralleling switchgear is switchgear. It carries the same multi-year OEM lead times as the rest of the medium-voltage line-up, and a generator plant with no gear to parallel it into is not capacity. See medium-voltage switchgear explained and what is a unit substation for that side of the build.

The rating says a set can run indefinitely. The fuel system decides how long it actually will.

For diesel plants, NFPA 110 Level 1 sets a floor of two hours of on-site fuel. Critical-facility practice runs far past that — twelve hours is a widespread design point, and many operators specify 24 to 72 hours backed by contracted resupply with priority terms. The number that matters is not the tank volume but the replenishment plan, because the outage scenarios that empty a tank are usually regional events during which fuel trucks are also stuck.

Diesel is a perishable asset. Ultra-low-sulfur fuel degrades, absorbs water, and grows microbial contamination, so a stored-fuel program means periodic polishing, biocide dosing, tank turnover, and sampling. The architecture is normally a bulk tank feeding day tanks with redundant transfer pumps, so a single pump or line failure does not starve the plant.

Natural gas removes the storage problem and replaces it with a delivery problem. A pipeline feed has no on-site buffer, so the availability of the generation becomes the availability of the pipeline — a single point of failure a behind-the-meter design has to underwrite explicitly. Dual-fuel is the common hedge: run on gas, keep enough liquid on site to ride through a curtailment.