Collaborative testing between data center operators and infrastructure providers reveals how AI workloads can destabilize diesel generators—and demonstrates the advanced UPS features that restore stability without compromising battery life or equipment longevity.
Vertiv, ST Telemedia Global Data Centers, and Baudouin collaborated to evaluate the performance of the Vertiv™ PowerUPS 9000 system under realistic AI workload conditions. Using the Vertiv™ AI Load Simulator, the research team replicated real-world GPU transient profiles observed in operational data halls running high-performance GPU clusters. Testing was conducted in a configuration that included supporting onsite power generation infrastructure, enabling assessment of system behavior under dynamic AI-driven load conditions.
This strategic alliance reflects a collective commitment to advancing data center resilience for AI workloads, with each partner contributing specialized knowledge: Vertiv's power protection expertise, ST Telemedia’s operational insights from managing hyperscale facilities, and Baudouin's generator engineering capabilities. Together, they worked on a testing methodology designed to validate infrastructure performance under the demanding, dynamic conditions that characterize modern AI computing environments.
This initiative offers a practical roadmap for how we design, operate, and maintain AI-ready facilities. The results redefined what power system resilience can look like and reveal why traditional infrastructure approaches may fall short when supporting next-generation GPU clusters.
Collaboration for testing and mitigation strategies
GPUs performing training or inference operations create highly transient load profiles. These processors synchronize their activity, generating steep power peaks and troughs every 40 milliseconds. The result is a constantly fluctuating demand pattern that behaves nothing like traditional IT equipment.
Artificial intelligence (AI) workloads’ continuous fluctuations expose the critical vulnerability that most data centers face: reliance on diesel generators for backup power. Generators are designed for smooth load ramps and are unable to maintain stable frequencies and voltages under rapid AI-driven swings. While the industry has focused extensively on the computational demands of AI workloads, a critical challenge has remained largely unexplored: how these workloads interact with the fundamental power infrastructure that keeps them running.
Previous industry analyses examined individual components: how uninterruptible power supplies (UPS) handle AI loads, or how these demands might impact the utility grid. What remained unexplored was how the entire critical power chain performs when supporting AI workloads on generator power, the backbone of data center resilience during utility outages.
Researchers implemented the study in a laboratory-controlled environment across four scenarios:
- Scenario 1: Currently observed load profile from H100 GPUs
- Scenario 2: Currently observed load profile H100 GPUs with data hall power increased to 100%
- Scenario 3: Worst-case GPU loading
- Scenario 4: Rack-based load smoothing
Impacts and solutions that work
When operating on generator power without mitigation strategies, the rapid current peaks and valleys from AI workloads caused generators to struggle in maintaining voltage and frequency within acceptable ranges. Output frequency fluctuated significantly, preventing the UPS from synchronizing with its bypass path; a critical capability for maintenance operations and system transfers.
The severity of these impacts correlated directly with the proportion of AI versus traditional loads. When AI workloads exceeded approximately 75% of total IT capacity, frequency variations became severe enough to prevent bypass synchronization entirely, even with certain mitigation measures in place.
This threshold varies based on multiple factors, such as generator rating and specific infrastructure configuration, among others. However, the fundamental challenge remains consistent across scenarios.
The test also validated effective mitigation strategies. Among them are advanced load smoothing capabilities that leverage battery systems to preemptively reduce current peaks before they reach the generator. Even at just 20% smoothing, the system sufficiently restored stability under realistic operating conditions, reducing generator voltage and frequency variations from approximately ±10% to ±0.5%.
Note that this approach doesn’t simply shift the problem to battery systems. Extended endurance testing revealed no significant impact on battery life for either lithium-ion or nickel-zinc technologies. After over 300 hours of testing, battery temperature remained stable, and no measurable degradation occurred. Moreover, battery autonomy increased by 4%, likely due to the brief troughs in the AI load profile.
The generator itself also fared better than anticipated. Post-test inspection showed that after extended exposure to AI load patterns, the wear on primary friction components was equivalent to standard operating conditions. Advanced fuel injection algorithms enabled the engine to respond effectively to rapid load changes without accelerated degradation.
Design implications and moving forward
Facilities should configure UPS systems with higher frequency slew rates—the speed at which they adjust frequency to match input sources. Operating UPS systems in double conversion mode becomes essential to isolate downstream loads from frequency fluctuations.
A more crucial implication of this study on design is how facility operators should consider their operational procedures. Gradually walking loads onto generators during transfers from utility power, rather than sudden switching, helps prevent step-load conditions. Maintenance sequences should avoid placing UPS systems on bypass while operating directly on generators supporting high AI load percentages.
As AI deployments scale and GPU densities increase, understanding these power system interactions becomes increasingly critical. The gap between traditional infrastructure design assumptions and AI workload realities represents a genuine risk to uptime and reliability.
The good news is that with proper configuration, advanced features, and thoughtful design, these challenges are manageable. The infrastructure to support AI workloads exists, but it requires intentional implementation of capabilities specifically designed for these demanding profiles.
To read all the technical details, analyses, specifications, and insights from this collaboration, download the white paper today. To inquire about how Vertiv can support your facility’s power needs from design to deployment, book a meeting here.
