In the past few years, AI interest has exploded, leaving the data center industry gasping for the power needed per rack to fuel digital transformation. As AI workloads become more demanding, power needs have only increased.

For Nicole Dierksheide, global category director of large power at Rehlko, the disruption is not simply about the amount of power AI consumes. It’s about the way AI consumes power. Whereas traditional data centers were built around diverse workloads across thousands of servers, she describes AI training environments as “fundamentally different”.

“They’re complex, they could be using tens of thousands of GPUs operating in synchronization, moving repeatedly between compute-intensive and communication-intensive phases,” she says. “That synchronized behavior can create large, facility-wide power swings, rapid ramp rates, and recurring oscillations in power demand.”

Evaluating resilience

For critical power systems, this new level of demands means many of the assumptions that have guided data center design for decades are changing.

Where historically generators, UPS systems, and electrical infrastructure were designed to manage occasional transient events, AI introduces a world where dynamic loading is a continuous operating condition rather than an exception, Dierksheide explains.

“Reliability is evolving from a simple question of ‘is power available?’ to ‘can the power system continuously maintain stable performance under highly variable operating conditions?’ – that shift fundamentally changes how resilience must be evaluated,” she adds.

As data center environments become more complex, AI inevitably raises expectations for power system performance, testing, and validation. For Rob Danforth, director of advanced development and simulation at Rehlko, these workloads force the industry to validate performance under sustained dynamic operation instead.

“Operators increasingly need confidence that infrastructure can continuously maintain voltage stability, frequency control, and predictable response during ongoing load variability,” he explains. “That requires more advanced modeling, extensive testing, and validation methods that accurately reflect how AI environments actually behave.”

With things changing so quickly, operators can also overlook certain parts of the process. Dierksheide explains that one of the biggest mistakes is focusing exclusively on megawatts.

“Many organizations begin with the assumption that if they have sufficient utility capacity and backup generation, they are AI-ready. In reality, AI readiness is increasingly determined by how a system performs, not simply how much capacity it has installed,” she says.

Another common oversight is evaluating equipment independently instead of evaluating the entire power ecosystem. Each component can perform exactly as specified individually, yet unexpected behaviors might emerge when they interact under AI-driven load conditions.

“Operators can underestimate how quickly AI infrastructure is scaling,” she says. “AI-focused data center electricity demand is projected to grow dramatically over the remainder of the decade, while grid expansion timelines continue to lag behind demand growth.”

This ultimately creates pressure not only on capacity planning, but also on resiliency strategies, energy storage integration, and onsite generation. Rehlko is helping to address these challenges via Clarke Energy, a Rehlko company, to look at prime power, continuous power, and bridging power to consider how entire systems will react.

Given the speed in which AI infrastructure is expanding, Dierksheide acknowledges grid constraints versus power requirements:

“The key lesson is that AI readiness is not a procurement exercise. It’s a systems-engineering exercise.”

Building strategy

As the industry enters a period where power infrastructure needs to evolve as quickly as compute infrastructure, Rehlko’s role is to help customers navigate that transition with an integrated, system-level approach.

The energy resilience leader is focused on looking at the full solution, looking at backup power and continuous power generation. It also has additional expertise in areas like energy storage, power conditioning, and advanced controls.

“We bring these together into a coordinated energy resilience strategy,” Dierksheide notes. “Rehlko’s focus is helping customers understand how infrastructure will perform before deployment, validate performance under real-world operating conditions, and create architectures that remain resilient even as technology, workloads, and energy requirements continue to evolve.”

With obvious AI-driven load growth electricity demands, Rehlko says operators need partners that understand the needs of both the data center and the realities of the evolving energy ecosystem.

“What we’re really doing is stepping back to de-risk and avoid finding issues at the time of installation and commissioning,” Dierksheide adds.

Previously everything was treated independently and there was no real interplay between the various components on account of things happening at such slow speed. Today, however, there is interconnection.

For Danforth, AI workloads expose the limitations of this component-by-component approach.

“This is why infrastructure evaluation needs to consider the complete powertrain,” he says. “The key question is no longer whether each individual component works. It is whether the integrated system can deliver stable, predictable power under sustained AI-driven dynamics.”

As AI data centers become even more demanding, the first step is to help operators move beyond historical assumptions. Dierksheide explains this is because of the critical importance of looking at things from a system level.

“It’s about making sure they’re educated and have awareness. AI workloads introduce operating characteristics that traditional validation methodologies were not necessarily designed to address,” she says. “That’s why Rehlko is investing heavily in advanced modeling, simulation, and real-world testing that evaluates performance under AI-representative load profiles.”

Beyond validation, Rehlko helps customers implement integrated architectures that combine generation, battery energy storage, UPS technologies, and advanced controls. It does this as research across the industry increasingly points toward layered approaches to managing AI load variability, with tech giants like Microsoft emphasizing the importance of combining infrastructure-level and storage-based solutions to stabilize power behavior.

“Solutions like Rehlko’s e-POD architecture reflect that philosophy by aligning different technologies with the timescales they manage best, from millisecond-scale buffering to long-duration resilience,” Dierksheide adds. “That helps to simplify not only installation and commissioning, but also the design perspective – getting a broader scope of the solution from a single supplier that’s looking holistically at the system.”

“Ultimately, our objective is to help customers reduce deployment risk while creating infrastructure that can grow alongside AI demand.”

Quality and quantity

Designing for real-world loads, rather than focusing only on installed capacity and traditional redundancy, is critical to long-term resilience. This is because resilience is determined by performance.

Danforth explains that AI-driven environments can remain within their overall power capacity while simultaneously creating operating conditions that stress generators, UPS systems, controls, and distribution networks in ways traditional validation methods never anticipated.

“Designing for real-world loads means understanding how the infrastructure responds to dynamic behavior over time,” he says. “Capacity is important, but long-term resilience depends on stable voltage, frequency control, coordinated system response, and validated performance under actual operating conditions.”

Historically, loads were considered based on nameplate rating and the load didn’t typically exceed 90 or 100 percent. AI loads today are more pulsive and go up to 150 percent for shorter durations.

“This needs to be factored in when looking at things that historically had steady or slow-changing loads, without that dynamic loading effect on the various components,” Danforth adds. “That needs to be looked at, and it’s critical for long‑term resilience.”

He explains how the data center industry should evolve from availability-based metrics toward performance-based metrics. While availability is important, AI environments need greater visibility into how systems perform while they’re operating, not just whether they remain online, he says.

“That means increased focus on voltage stability, frequency regulation, transient response, dynamic load handling, control system coordination, and overall stability of the system,” he notes. “Reliability measures need to include looking at the overall mechanical and thermal stress presented to the components associated with the cyclical nature of AI nature loading.

“The future of reliability is about delivering consistent power quality across the entire facility under continuously changing conditions.”

He also makes a case for testing and validation needing to evolve for AI environments, as they require testing that reflects sustained, repetitive, and synchronized load behavior. At Rehlko, this involves combining real-world generator testing with advanced modeling and simulation to understand how systems behave across a broader range of operating conditions.

“We’re increasingly evaluating not only steady-state performance but also stability and reliability associated with dynamic AI loading,” Danforth explains. “Rehlko has Simulink models of its components which can be incorporated into the complete powertrain analysis.

“We’ve had to evolve, create new ways to load faster, and validate our various models so that there’s high confidence and trust when they’re put into a larger system.”

Designing for change

The IEA expects global data center electricity demand to more than double by 2030, with many regions facing transformer shortages, generation constraints, and length grid interconnection timelines.

This means there is an even greater need for infrastructure to support flexibility as much as capacity, as AI power densities, rack architectures, cooling requirements, and utility constraints continue to rapidly evolve. For Rehlko, ensuring this means designing for change rather than looking at today’s requirements alone.

“A scalable strategy typically includes modular architectures, integrated energy storage, flexible generation assets, advanced controls, and the ability to support multiple operating modes, whether grid-connected, grid-constrained, or microgrid-based,” Dierksheide says. “Future-ready infrastructure is not simply designed to withstand disruption – it’s designed to evolve with the business.”

She advises that operators make modular changes so they don’t have to redesign everything, in addition to paying attention to control systems they have in place.

“Yes, you want to be resilient, but you have to be scalable,” she adds. “What does it mean to scale beyond designing for the future – it’s doing concrete things today to enable that.”

To make these processes clearer, Relhko launched AI Readiness Starts with Power, an eBook outlining the company’s approach to AI readiness. Its core message is simple: AI readiness starts with understanding real-world power behavior.

The eBook makes a case for operators needing confidence in both component specifications and system behavior. This involves validating voltage stability, frequency response, transient performance, control interaction, and resilience under realistic operating scenarios.

“Rehlko is focused on helping operators bridge that gap through integrated architectures, advanced validation, and energy resilience solutions designed specifically for emerging AI demands,” she explains.

The broader industry is reaching similar conclusions, acknowledging that supporting next-generation AI training environments require coordinated solutions across hardware, software, and infrastructure layers. Historically, resilience often meant ensuring backup power was available during an outage. Now, the goalposts of what it means to be an energy resilience partner are changing.

For Rehlko, resilience going forward means helping customers continuously adapt to an evolving energy landscape.

“AI growth, utility constraints, changing regulatory environments, new power architectures, and shifting sustainability goals are all converging at the same time,” Dierksheide explains. “When we think about being an energy resilience partner, our focus is pulling everything together, knowing we have the breadth of capability, solutions, and expertise to build from to get the right solution for the customer.”

As AI continues to reshape infrastructure requirements, customers will need partners who can help them anticipate change, reduce complexity, validate performance, and scale confidently. That’s the role Rehlko is committed to playing.

Rehlko is defining a new standard for AI Readiness here.