Data center power infrastructure has an engineering problem. Faced with the demands of AI training and inference clusters, a millisecond gap has emerged between AI load behavior and traditional power system response. The core issue is a mismatch in reaction speed between three things that all have to work together: the load, the power electronics, and the mechanical generation.

When thousands of GPUs in a training cluster shift in lockstep between compute-intensive and communication-intensive phases, power draw doesn't ramp gently, it swings sharply and repeatedly in the 10-200 millisecond range. This isn't one server spiking; it's a synchronized, cluster-wide event, so the amplitude is large and the pattern repeats continuously for the duration of the job.

Joint research by Microsoft, Nvidia, and OpenAI has confirmed that AI workloads introduce high-frequency, repetitive electrical oscillations that cannot be addressed by mechanical generation systems alone. 

A diesel generator, even a well-designed one, typically takes 200-450 milliseconds to respond to a step change in load. That's because it depends on mechanical inertia and electromagnetic response in rotating machinery, physical processes with hard speed limits. A UPS is much faster, but once again, traditional units are engineered for occasional ride-through during outages, not for absorbing continuous, high-frequency cycling.

Why the Gap Matters

When the load can change faster than a generator can even begin to react, there are consequences. 

The fluctuation doesn't get absorbed; it propagates. If nothing intercepts the swing at the point where it originates, it travels upstream through switchgear, transformers, and eventually toward the grid interconnection or generator systems, where it may affect overall power quality and system stability.

Equipment gets stressed in ways it wasn't designed for. UPS batteries built for infrequent backup start cycling constantly, accelerating wear. Generators exposed directly to fast and repetitive AI-driven load profiles may experience operating conditions that differ materially from those seen in traditional data center environments, and long-term field data remains limited. Nobody yet has the long-term field data to say exactly what the impact will be over time.

Traditional data center loads were steady or slowly varying, so architectures were sized for capacity and occasional transients, not for constant millisecond-scale oscillation. Oversizing a generator doesn't fix a speed problem, because a bigger machine still can't outrun its own mechanical inertia.

AI workloads create power disturbances on a timescale that only power electronics –fast-responding UPS or battery buffering in the 2-4 millisecond range – can handle. The generator is simply too slow to serve as the first line of defense. This is exactly why e-POD, our prefabricated, power solution, doesn't try to make generators faster. It removes the need for them to be fast by keeping the first response in the domain of power electronics.

Absorbing the Impact of AI Loads

Instead of oversizing generators or bolting on isolated mitigation devices, e-POD separates the power-stabilization problem into three distinct layers, each matched to the timescale it is physically capable of managing.

Layer 1: StoreModule

The only element in the architecture with dedicated energy storage, StoreModule is built around battery-based buffering. Not designed as a backup energy source; its job is transient absorption. It releases power at the millisecond scale, reshaping the sharp, repetitive ramp rates produced by synchronized GPU activity into a smoother demand profile, absorbing and reshaping the volatility before it propagates through the rest of the system. 

Layer 2: ShieldModule

Sitting between the volatile load and the long-duration power sources, ShieldModule carries no storage of its own. Built around Rehlko's MF Series UPS, its function is confined to power-electronic conversion, voltage and frequency regulation, and electrical isolation. This division of labor is deliberate. Conventional UPS batteries are designed for occasional ride-through events during outages, not for the constant, high-frequency charge-discharge cycling that AI workloads would otherwise impose on them. 

By keeping fast buffering in the StoreModule and conditioning in the ShieldModule, the architecture helps reduce excessive cycling that could contribute to premature UPS battery aging.

Layer 3: GenModule

Supplying sustained power during extended grid interruptions, GenModule is an ordinary backup generator that does what generators are good at, but with a difference. Sitting behind Layer 1 and 2, it is never directly exposed to AI load volatility. The two layers absorb and smooth out the millisecond-scale chaos before it reaches the generator.

Interception Key to Generator Protection

Essentially, our approach with e-POD is to intercept fast transients at the StoreModule and ShieldModule layers, preventing generators from being tested by long-term AI-driven cycles. Meanwhile, GenModule continues to operate inside the well-understood performance envelope that generators have always been validated against, rather than serving as an unproven first line of defense against millisecond-scale swings.

The practical takeaway is twofold. First, third-party research from the hyperscalers has established that millisecond-scale AI load volatility is real and well-documented, not a theoretical edge case. Second, while long-term data on how generators would perform under direct exposure to that volatility is still limited, e-POD sidesteps the question rather than waiting to answer it. 

By shielding the GenModule, the generator is only ever asked to operate within the load profile that it’s already validated for. In effect, unproven risk is channeled away from the generator, arguably the power train's most capital-intensive and hardest-to-retrofit component, and toward power electronics layers purpose-built to absorb it.

To learn more about this topic, check out our ebook, Stabilizing AI Power at Scale with e-POD: A Layered Critical Power Architecture.