The UPS only has to bridge seconds — the gap until the generator starts and takes the load. Size the generator wrong, or the UPS battery too short, and that handoff is exactly where an outage becomes a real one.
UPS runtime and generator sizing solve two different halves of the same outage. The UPS battery's job is to carry the critical load for a short, bridging period — typically just a few minutes — covering the instant of utility failure through the time it takes the standby generator to start, reach rated speed and voltage, and have its automatic transfer switch (ATS) close onto it. The generator's job is everything after that: carrying the full critical load, indefinitely, for as long as fuel is supplied, until utility power returns. Confusing the two leads to two very different, both dangerous, design mistakes: sizing a UPS battery as if it needs to ride out a long outage (wasteful and usually impossible at scale), or sizing a generator without properly accounting for the load characteristics of the equipment it's starting, especially for AI/GPU facilities where load can swing dramatically and quickly.
A UPS battery is energy-limited — it stores a fixed amount of chemical energy that discharges in minutes at data center power levels, which is exactly enough to bridge the short, predictable window between utility loss and the generator's automatic transfer switch closing. A generator is fuel-limited, not energy-storage-limited — as long as diesel (or natural gas) keeps flowing to it, it can carry full load indefinitely, which is why data centers plan generator runtime in terms of on-site fuel tank capacity plus contracted resupply logistics, not battery chemistry. Sizing either system for the other's job is a mistake: an oversized UPS battery meant to ride out hours of outage is enormously expensive, heavy, and still eventually depletes, while an undersized UPS runtime margin that assumes a perfect, first-try generator start leaves no cushion for a failed start attempt or a slow transfer — a real risk that most engineering standards explicitly design margin around.
Steady-state average load is only the starting point. Generators must also be sized for transient conditions that can exceed steady-state demand significantly: large motor loads (chillers, CRAH fans) draw a sharp inrush current on startup that can be several times their running current, and UPS systems recharging depleted batteries after an outage add extra load on top of the IT and cooling load the generator is already carrying. AI/GPU facilities add a further wrinkle largely absent from traditional enterprise loads: GPU clusters can swing their power draw dramatically and rapidly — from near-idle to full rated power within seconds — as a training job starts, pauses, or synchronizes across nodes. These rapid load steps can stress a generator's governor and voltage regulation in ways a steady, slowly-varying enterprise load never does, and some facilities specifically evaluate generator response to simulated GPU load-step profiles as part of commissioning, beyond the traditional steady-state and motor-inrush sizing calculations.
Explains why UPS batteries are energy-limited and sized only to bridge the seconds-to-minutes gap until a standby generator takes over, while generators are fuel-limited and sized for peak transient load — including, for AI/GPU facilities, rapid GPU cluster load-step swings — to carry the full load indefinitely.
A UPS battery's stated runtime (commonly 5-15 minutes at full data center load) is intentionally short — its only job is to keep critical loads powered from the instant utility fails until the standby generator starts, reaches rated speed and voltage, and its automatic transfer switch (ATS) closes onto it, typically a 10-30 second process. Design runtime includes margin beyond the ideal-case generator start time to cover a failed first start attempt (most designs allow at least one retry) and any operator intervention time, which is why 5-15 minutes, not 15-30 seconds, is the typical design target.
A generator's electrical rating must cover more than steady-state average demand — it has to handle transient loads including motor inrush current from cooling equipment (chillers, CRAH/CRAC compressors and fans) starting up, and the recharge current UPS systems draw to replenish batteries depleted during the transfer bridge. For AI/GPU facilities, an added consideration is the rapid, large power swings a GPU cluster can exhibit as training jobs start, pause, or synchronize — these fast load steps can stress a generator's governor and voltage regulation differently than the slower-varying loads typical of conventional IT equipment. Once running, a generator's duration is limited by fuel supply, not any internal energy store, which is why data centers plan generator runtime in terms of on-site diesel or natural gas tank capacity (commonly sized for 24-96+ hours) plus contracted fuel resupply logistics for extended outages.
Extending UPS battery runtime to cover hours of potential outage, rather than relying on the generator, would require battery capacity that is prohibitively large, heavy, and expensive at data center power scales — the practical role of a large battery bank is bridging a short, predictable gap, not sustaining long-duration operation. Conversely, a generator cannot substitute for the UPS during the transfer bridge itself, since it takes real time to start and synchronize — the UPS exists specifically to cover that unavoidable startup interval.
Because a "typical" outage duration varies enormously and can't be reliably bounded, while battery energy storage at that scale becomes extremely costly, heavy, and requires large battery rooms with their own fire and ventilation code requirements. It's far more practical and cost-effective to have the UPS bridge a short, predictable gap and let a fuel-supplied generator handle everything beyond that.
Most designs include N+1 or 2N generator redundancy specifically for this reason, along with a UPS runtime margin that allows for at least one failed start attempt and automatic retry before the UPS battery would be fully depleted — this is why UPS runtime specs are set higher than the theoretical minimum needed for a single successful generator start.
Facilities size on-site fuel storage (commonly diesel) for a target minimum runtime, often 24-96 hours depending on the facility's risk tolerance and local grid reliability, and pair that on-site storage with contracted fuel resupply agreements for any outage expected to exceed on-site storage — this combination, not the generator itself, defines the facility's real-world sustained-outage capability.
It is an increasingly discussed consideration in AI/GPU facility design, since rapid, large load steps (a cluster snapping from idle to full power, or vice versa, within seconds) place different transient demands on a generator's governor and automatic voltage regulator than the more gradually varying loads of traditional enterprise IT equipment. Some facility commissioning processes now specifically test generator response under simulated GPU-style load steps rather than relying solely on traditional steady-state and motor-inrush sizing methods.