Data Center Backup Generators: How Many, What Size, and How Much Fuel

In a typical data center the utility feeds the load, UPS batteries carry it for the seconds to minutes after an outage, and engine generators start, synchronize and take over until utility power returns. The UPS covers the gap; the generators cover the duration.

How much the generators must carry depends on the reliability target. Uptime Institute defines the diesel generators as the primary source of power in Tier III and IV facilities and the utility as an economic alternative.¹ That puts two requirements on the generator plant: it must be large enough to carry the entire data center load, and there can be no limit on the hours it can run.¹ See our guide to tier levels and redundancy for how those classifications work.

“Entire data center load” means more than the critical IT load. Chillers, pumps, cooling towers or dry coolers, fans, UPS losses and building systems all have to stay on, and on the hottest day of the year the cooling plant is working hardest. That is why the generator plant is usually sized to the facility’s design-day peak, which is the same number that drives the utility power request. A site with a low annual PUE can still have a high peak overhead.

The same engine can carry several nameplate ratings, and the one that matters depends on how many hours it will run and how steady the load is. ISO 8528-1 sets the base definitions.¹ Under them, a standby rating allows no more than about 200 hours of operation per year on a variable load, prime allows unlimited hours at a 24-hour average load of 70% of rating or less, and continuous allows unlimited hours at full rated output on a constant load.⁸ Limited-time prime is capped at about 500 hours a year and is aimed at peak shaving and similar utility arrangements.⁸

None of those fits a data center neatly. A standby rating fails Uptime’s no-hour-limit test, while a full continuous rating means buying much more engine than the load needs. Uptime references the ISO definitions but does not require them, and manufacturers have responded with ratings developed specifically for data center applications, usually called data center continuous (DCC or DCP).¹ Cummins, for example, lists its QSK95 series with a data center continuous range of 2,500–3,350 kW at 50 or 60 Hz.⁶

Because DCC is a manufacturer rating, its terms vary. One Cummins data sheet notes that its DCC rating is subject to a site-specific assessment.⁷ Ask for the rating basis in writing when comparing bids.

Most large data centers use many medium-voltage or low-voltage diesel sets in the 2–3 MW range rather than a few very large units. That keeps each failure small, matches standard electrical blocks and lets the owner build in phases. Hyperscale campuses can hold dozens or even hundreds of generators.⁵

The count follows from three numbers: the design-day facility load each block must carry, the unit rating, and the redundancy scheme. A block designed N+1 needs enough units to carry its load with one out of service; a 2N design duplicates the whole set. A worked example:

1. 01Start with IT load per building, for example 48 MW.
2. 02Multiply by the peak (not annual) overhead to get facility load, for example 48 × 1.35 ≈ 65 MW.
3. 03Divide by the usable unit rating, for example 2.7 MW DCC, which gives 24 units at N.
4. 04Add redundancy by block: with eight blocks of three units plus one spare each, the building needs 32 units.

These numbers are illustrative; the electrical engineer sets the real block design. The site question is space and access. Each unit needs a pad or enclosure, fuel storage, exhaust clearance and room for crane and tanker access, which is why generator yards take up a large share of the equipment yard. Noise from testing also matters to neighbors and to the noise setbacks and buffers a county may require.

Fuel duration is set by codes, the reliability target and how much the owner trusts fuel delivery. NFPA 110 classifies emergency and standby power systems by Class, which is the minimum time in hours the system is designed to run at rated load without refueling; Class 6, for example, means six hours.⁹ The 2005 edition required a Level 1 system in seismic design categories C through F to be Class X with at least 96 hours of fuel, and one fire marshal’s policy notes that for other buildings the duration is generally set by the systems served and often does not exceed two hours.⁹

For data centers the more common reference is Uptime. Its Owners Advisory Committee set 12 hours of on-site fuel as a starting point, and the Tier Standard: Topology applies that 12-hour minimum at N load to all Tiers.² Storage beyond 12 hours is treated as an operational sustainability decision.² Uptime also counts only the fuel that survives the loss of a redundant component: an owner claiming 48 hours from two 24-hour tanks has only 24 hours of concurrently maintainable fuel.²

Industry practice sits well above Uptime’s floor. One 2026 industry analysis says Tier III and IV facilities typically specify 96 hours of on-site diesel at full load, while some hyperscale operators have compressed contractual commitments to 48–72 hours on site backed by resupply contracts.³ Natural gas-fired backup does not escape the rule: Uptime treats pipeline gas as a utility and expects a minimum of 12 hours of fuel stored on site, which can mean large tanks with their own code and explosion-risk issues.²

Tank sizing is arithmetic once the runtime target is set: fuel burn per hour at the expected load, times hours, plus allowances for fuel that cannot be drawn from the bottom of the tank, thermal expansion and the engineer’s margin. The burn rate comes from the engine’s data sheet, not a generic chart.

At full load that set burns about 664 liters (176 gallons) per hour, or roughly 0.25 liters per kWh generated.⁷ A 3.0 MW unit at full load uses roughly 720–750 liters per hour.³ Holding 96 hours for one 2,680 kW unit therefore takes about 64,000 liters (about 16,800 gallons) of usable fuel before margins, and a building with 32 such units needs more than 2 million liters if every unit must run at full load for the full period. In practice units in redundant blocks run partly loaded, so engineers size storage to the expected load profile.

Storage can be belly tanks under each generator, shared bulk aboveground tanks with day tanks, or a mix. Bulk storage simplifies deliveries but adds piping and pumps that need their own redundancy. Large aboveground volumes also trigger spill-prevention and fire-code requirements, covered in our guide to diesel fuel storage and SPCC plans.

On-site storage only buys time; a long outage is survived by deliveries. The resupply contracts that let some operators hold 48–72 hours rather than 96 depend on suppliers actually reaching the site during a regional emergency.³ Practical points for a site:

• All-weather truck access to every fill point, with turning room for tankers and a route that does not cross the floodplain.
• More than one supplier and terminal, ideally on different road routes, with priority-delivery terms written into the contract.
• Fuel maintenance (testing, polishing and turnover), because stored diesel degrades. One trade analysis notes petroleum diesel is typically stored for up to about a year, while hydrotreated vegetable oil (HVO) can be stored for up to 10 years without major changes.¹⁰

HVO, a renewable diesel, is increasingly used in data center generators because it requires no modification to existing engines and works as a direct replacement for diesel.¹⁰ Microsoft’s Swedish cloud region began running generators on a renewable-blend fuel, Kao Data moved its Harlow, England campus to HVO, and Compass Datacenters started using HVO at campuses in Northern Virginia, Arizona and Texas.¹⁰ Supply is the open question: whether production keeps pace with fast-growing campuses.¹⁰

Federal rules treat a backup diesel as an emergency engine only if it runs within limits. Under 40 CFR 60.4211(f), there is no time limit on emergency use, but maintenance checks and readiness testing are capped at 100 hours per calendar year, and up to 50 of those hours may be used for non-emergency purposes.⁴ Those 50 hours cannot be used for peak shaving, non-emergency demand response or selling power to the grid, except in narrow cases where the local balancing authority or distribution operator dispatches the engine.⁴ An engine run outside these limits must meet the requirements for non-emergency engines.⁴

States are tightening the rules where generators cluster. Virginia’s Department of Environmental Quality revised its permitting guidance, saying the traditional “limited operating hours” basis for exempting emergency data center generators from stricter controls “is no longer appropriate” for permit applications submitted on or after July 1, 2026.⁵ Selective catalytic reduction, diesel particulate filters and continuous emissions monitoring could be required.⁵ The 2026 General Assembly also passed HB 507, which bars DEQ from issuing an air permit for a data center application submitted on or after July 1, 2026 unless each generator’s emissions are no higher than a Tier 4 equivalent engine’s.¹¹⁵ Confirm the enacted version and DEQ’s current guidance before relying on either. DEQ has also posted data center air permits online since November 2024, which makes generator counts on nearby campuses public.¹²

Tier 4 aftertreatment adds equipment, space and cost per unit, and total permitted emissions can trigger major-source review on a large campus. Our guide to air permits for data center generators covers the permitting path, including PSD and Title V.

Generators and fuel touch more of the site plan than most buyers expect: yard area, truck routes, air permits and neighbors. A quick screen of each before closing avoids redesign later.

1. 01Estimate design-day facility load, not just IT load, and the generator count it implies with redundancy.
2. 02Confirm the fuel runtime target with the end user, and check NFPA 110 and fire-code requirements with the local authority.⁹
3. 03Lay out the generator yard, tanks, exhaust and tanker routes early; they drive the site plan and noise setbacks.
4. 04Check state air rules and nearby permitted sources before assuming Tier 2 engines will be allowed.⁵¹¹
5. 05If on-site generation will also serve as prime or bridge power, read our guides to on-site generation and grid power vs. on-site generation; emergency-engine limits will not apply.⁴

BlackForge screens sites for yard space, access, air-permit exposure and neighbors along with power and water. You can get a site reviewed.