Backup Power in Data Centers: Generators, UPS, and Runtime

A data center’s backup power is a two-stage relay, not a single device. UPS batteries carry the load for 5-15 minutes to bridge the gap between a utility failure and generator start, while diesel generators — sized by Uptime Institute Tier to 12-96 hours of onsite fuel — take over for anything longer. In 2026, the constraint on that design is no longer engineering; it is a generator order book stretched to 2028 by AI data center demand.

Key takeaways

• UPS batteries are sized for 5-15 minutes, not hours — they exist to bridge to generator start, not to replace it.
• Lithium-ion batteries pack 150-200 Wh/kg versus 30-50 Wh/kg for VRLA lead-acid, with a 10-15 year design life against VRLA’s 3-5 years.
• Uptime Institute Tier III requires N+1 generators with a minimum 12-hour onsite fuel target; Tier IV requires 2N or 2N+1 with a 96-hour bulk tank.
• EPA’s RICE NESHAP (40 CFR Part 63, Subpart ZZZZ) caps emergency generators at 100 hours/year of non-emergency running, with true outages exempt from the limit.
• Diesel gensets in the 1,250-3,250 kW range hyperscale halls specify now quote 52-78 weeks; 3,000-4,000+ kW units run 90-110 weeks, per 2026 industry lead-time tracking.
• Caterpillar’s backlog is reported near $63 billion and Cummins’ data center allocation is reportedly sold out through 2028, pushing operators to reserve engines before a building is even designed.
• Bloom Energy booked roughly $7.65 billion in data center-related fuel cell contracts in a 90-day window in early 2026 — mostly as primary/bridge power to skip grid queues, not as emergency backup.
• Diesel keeps the emergency-standby spec sheet because it starts and accepts full load in seconds and stores its own fuel onsite; natural gas needs longer ramp time and depends on pipeline pressure.

Why backup power is two systems, not one

No single technology covers both failure modes a data center has to survive. A generator cannot start and stabilize output instantly — most designs need anywhere from several seconds to under a minute to reach rated voltage and frequency, and IT load cannot tolerate even a few seconds of interruption. So facilities put a UPS (battery, flywheel, or occasionally an integrated diesel-rotary system) directly between utility power and the load, sized only to ride through that gap.

That division of labor shapes the economics: UPS capacity is bought in minutes of runtime and generator capacity is bought in hours to days. Conflating the two — for example assuming a UPS can substitute for a longer outage — is one of the more common design mistakes flagged in facility due-diligence reviews; see our data center due diligence checklist for the fuller list of what buyers should verify before signing.

UPS batteries: runtime, chemistry, and what changed

Standard practice sizes UPS battery strings for 5-15 minutes of full-load runtime, with many modern designs converging on 5-10 minutes since that is enough to start generators and complete an automatic transfer switch (ATS) sequence, or to execute a controlled shutdown if generators fail to start. Extending battery runtime past that point is rarely cost-effective — it is cheaper to add generator redundancy than to buy hours of battery capacity.

The chemistry choice has shifted meaningfully over the past decade:

Attribute	VRLA (lead-acid)	Lithium-ion
Energy density	30-50 Wh/kg	150-200 Wh/kg
Typical design life	3-5 years	10-15 years
Footprint for equivalent runtime	Baseline	Roughly 40-60% smaller
Thermal tolerance	Degrades faster in heat	More sensitive to thermal runaway risk, needs BMS
Upfront cost	Lower	Higher (but lower TCO over 10+ years)

Lithium-ion’s longer service life and smaller footprint have made it the default for new-build hyperscale and colocation UPS plants, even though upfront capex per kWh is higher than VRLA — the same tradeoff shaping choices in stationary storage more broadly, covered in our BESS price index. VRLA remains common at edge sites and smaller facilities where footprint and thermal management are less constrained and lower upfront cost dominates the decision.

Generator redundancy: N+1, 2N, and what Tier actually requires

Redundancy notation describes how many generators exist beyond the minimum needed to carry the full design load:

Configuration	Meaning	Typical Tier
N	Exactly enough capacity, no spare	Tier I/II (generator optional)
N+1	One extra unit beyond the load requirement	Tier III
2N	The entire plant duplicated; either half covers 100% of load	Tier IV
2N+1	2N plus one additional spare unit	Tier IV (AI/hyperscale halls)

A 4 MW hall built N+1 with 2 MW gensets needs three units — two to carry the load, one spare. The same hall built 2N needs four units split into two fully independent 2 MW plants. AI-dense halls increasingly specify 2N+1 or even 3N because a generator failure during a multi-day grid event is treated as a probable event, not a tail risk, given how long replacement parts and backup units now take to source.

Fuel storage scales with the same logic. Uptime Institute guidance targets a minimum 12-hour onsite supply for Tier III and a 96-hour bulk tank for Tier IV, with the caveat that the number that matters is what remains after the redundant component is pulled out of service for maintenance — not the raw tank capacity. Many operational SLAs commit to 24-72 hours of guaranteed runtime plus contracted resupply, which is why fuel-delivery contracts are as much a part of the design as the tanks themselves. Rack-level power density, covered in our power density guide, pushes the same math per square foot: a denser hall needs proportionally more backup capacity in the same footprint.

Diesel vs. natural gas vs. fuel cells

Diesel remains the default emergency-standby fuel for three practical reasons: it starts and accepts full rated load within seconds, which is what standby-generator codes and Tier III/IV commissioning tests require; it stores energy onsite with no dependence on pipeline pressure during a regional event; and its energy density supports 12-96 hours of runtime in a tank footprint that natural gas cannot match without a dedicated onsite gas storage system.

Natural gas generators and turbines are increasingly used for continuous or near-continuous prime power — often to get a site energized faster than the utility interconnection queue allows — but are less common for pure emergency-standby duty because of slower ramp characteristics and gas-supply curtailment risk during exactly the kind of regional grid stress that triggers a data center outage in the first place. Grid queue delays of five to ten years in constrained US markets, detailed in our grid connection queues guide, are the direct driver behind operators exploring onsite generation as more than a backup asset.

Solid-oxide fuel cells are the newest entrant. Bloom Energy reported roughly $7.65 billion in data-center-related contracts inside a 90-day window in early 2026, positioning fuel cells as bridge or primary power that can be installed in under two months versus years for a new grid connection, at availability ratings from 99.9% up to 99.999% depending on configuration. That is a pitch for continuous power, not emergency backup — fuel cells do not currently displace the generator plant required for standby-code compliance and Tier III/IV certification.

The 2026 generator supply crunch

The bottleneck in backup power design this year is procurement, not engineering. Lead times by unit size, per 2026 industry tracking:

Generator size	Typical 2026 lead time
25-400 kW	12-26 weeks
500-600 kW	12-30 weeks
750-1,000 kW	12-39 weeks
1,250-3,250 kW	52-78 weeks
3,000-4,000+ kW	90-110 weeks

The largest units — the size range hyperscale and AI-dense halls specify — are the most backlogged. Caterpillar’s order backlog is reported near $63 billion, and Cummins’ data center generator allocation is reported sold out through 2028. Switchgear, the equipment that ties generators, UPS and utility feeds together, is running up to 72 weeks on its own, so a generator arriving on schedule does not guarantee the plant is commissionable on schedule. Buyers now reserve production slots before a building’s design is finalized, which reshapes site-selection timing in the same way grid interconnection queues already have — see the financing and development-model tradeoffs in our powered shell vs. turnkey guide.

Emissions rules: what “emergency” actually means

US federal air rules treat backup generators as emergency equipment with a narrow legal definition, and that definition has real teeth. Under the RICE NESHAP (40 CFR Part 63, Subpart ZZZZ), an emergency engine may run for true emergencies — a genuine utility outage, fire, flood — with no hour limit. But non-emergency operation, meaning routine testing, maintenance runs, and any use tied to demand-response or grid-support financial arrangements, is capped at 100 hours per year, with only 50 of those hours available for demand-response-style use. Voltage sag or a utility warning that hasn’t yet caused an outage does not qualify as an emergency under the rule.

Emergency-classified engines are also exempt from the strictest EPA Tier 4 Final emissions standards and may legally run at older Tier 2 emission levels — a carve-out justified by infrequent use that comes under scrutiny whenever a facility’s generators run more often than the emergency label implies. That scrutiny has increased as AI data center campuses cluster dozens of 1-3 MW units in single locations, raising local air-quality objections independent of the federal hour cap. Any site evaluation should confirm which permit class its generators hold and whether logged run-hours are consistent with genuine emergency use, alongside the certification and disclosure checks covered in our data center due diligence guide.

What this means for buyers and site selectors

1. Match battery and generator sizing to actual failure modes. Don’t buy UPS runtime to cover scenarios generators are meant to handle — 5-15 minutes of battery plus a properly redundant generator plant is the standard split, and stretching battery runtime is the more expensive way to buy the same protection.
2. Ask for the post-maintenance fuel number, not the tank size. A Tier IV claim means little if the operator cannot state guaranteed runtime with the redundant fuel system removed from service.
3. Price in the 2026 generator queue before committing to a delivery date. If a build specifies 1,250 kW+ units, a 52-78 week (or longer) lead time is now a baseline planning assumption, not a risk case — confirm the operator’s engines are already on order, not “to be sourced.”
4. Verify emergency-permit compliance, not just Tier certification. RICE NESHAP hour limits and local air permits increasingly shape what a facility can legally do with its generator plant beyond backup duty, which matters for anyone evaluating demand-response revenue claims.
5. Treat fuel cells and gas turbines as speed-to-power tools, not backup replacements. They solve the interconnection-queue problem; the generator plant required for Tier III/IV emergency-standby compliance is a separate line item.

Compare backup-power specifications alongside pricing and power availability across markets in the colocation price index and the full data center catalog, and see current utility tariffs by market in the energy hub.