The way we used to think about data center design was built around the assumption that most of the power chain would spend most of its life in a stable, predictable state, with occasional deviations when something went wrong. Generators, UPS systems, transfer switches, even the way reliability itself was measured, were all engineered around the premise that normal operation is steady, disturbances are rare, and the job of backup infrastructure is to survive the occasional exception.

All that changed with the arrival of AI training and inference clusters that no longer behave like traditional IT loads. GPUs swing hard and fast between near-maximum draw and near-idle, often in sync across thousands of chips, repeatedly, for hours at a time. When that kind of load lands on a generator, ‘standby’ stops describing what the system actually does. It's no longer waiting for a rare event; it's being asked to perform continuously, under conditions closer to sustained stress testing than an emergency response. 

UPS and battery systems can help absorb fast transients and smooth portions of the load profile, but these layers only reduce exposure to variability rather than eliminate it entirely. System performance will ultimately be decided on how effectively the underlying architecture responds to sustained and rapidly changing demand.

Rather than reducing the importance of standby power, these new load profiles increase the need for power systems engineered to maintain performance under dynamic operating conditions.

We fully explore this topic in our recent ebook, AI Readiness Starts with Power: Designing for Real-World Load Conditions. Check out the full ebook here and keep reading for highlights.

Less Time to Recover and More Stress

A big challenge is recovery time, because a generator doesn't snap back to steady state instantly after a big load swing. There's a settling period of a few seconds where voltage and frequency stabilize.

In a traditional facility, that recovery window isn’t an issue, because the next disturbance may be minutes, hours, or even days away. With AI workloads and GPU clusters running at scale, recovery periods shrink or vanish entirely, and control systems find themselves continuously active rather than occasionally called upon. The generator is still catching up from the last swing when the next one lands.

Mechanical and thermal stresses that used to dissipate between events now start to accumulate. What used to be considered an unusual, high-stress moment for a generator becomes the everyday operating condition, becomes an everyday operating condition.

The result of this sustained activity is additional wear and tear on equipment. Every rotating generator has natural frequencies it wants to vibrate at. When a repeating load swing happens to hit that same rhythm, even a small disturbance can build up and put stress on the shaft. When it happens repeatedly, possibly thousands of times a day, the risk of equipment fatigue increases, making system design, controls and validation increasingly important.

Add in unstable voltage and frequency, and protection systems that shut things down unnecessarily, and a big question arises around backup generators. Can they be relied on to deliver stable, predictable power when they are needed? The answer depends on how systems are designed, validated and tested for these operating conditions. 

Modelling Scenarios to Maintain Five-Nines Reliability

Traditionally, sizing a generator was mostly a matter of asking how big it needs to be to cover peak load, a static, steady-state calculation. With AI clusters, it becomes a physics problem, not a sizing calculation, because it’s no longer about the unit being big enough, it’s about how it behaves, moment to moment, while the load is swinging. Performance under load becomes a core reliability metric rather than a secondary engineering consideration. 

Because steady-state ratings are less relevant, Rehlko built models to show how generators behave during a disturbance, modeling the engine, the speed and voltage controls as well as the generator's own electrical behavior. To get the models right, we used an industry-standard test called Standstill Frequency Response, which measures exactly how a generator responds to fast, repeated disturbances.

Running a physics simulation for every possible AI load scenario is too slow to be practical, so we also developed fast-running, reduced-order models trained on both simulation data and real test-cell measurements. These compact models let engineers predict how a generator's voltage, frequency, and recovery behavior will hold up under thousands of AI-style load combinations in a fraction of the time a full simulation would take. 

All that work doesn't stay in the lab. The same models and test data feed directly into how we design, tune, and validate our generators, so the control systems, exciters, and governors aren't just rated for AI-style loads on paper, they've been validated through testing.  

We ran 900 different load steps, ramping demand from 25% up to 100% of capacity in 5% increments, holding each step for one to five seconds, and varying the idle base load between tests. Some of these transients were spaced closely together, which meant the generator never fully recovered before the next one hit, just like in the real world. The systems stayed stable, and the models predicting their behavior matched measured results within about 2%.

To survive and thrive in the world of AI factories, standby generators have to be engineered to perform reliably under sustained dynamic loading, recover quickly, and continue doing so consistently over time. But one thing hasn’t changed. We might be redefining AI readiness around maintaining stability, control, and performance under continuous dynamic loads, but we are still holding to five-nines reliability, the same benchmark critical power has always been measured against.