Rehlko warns AI data centres need dynamic power checks

Rehlko has published an eBook on power infrastructure for AI-driven data centres, outlining its view of how changing AI workloads are reshaping critical power requirements.

The publication argues that operators should assess AI readiness by how power systems perform under variable real-world loads, not by installed capacity alone. It examines the effect of large language models and other AI applications on data centre energy demand, particularly in environments built around graphics processing units.

Data centre operators are under pressure to add AI capacity while utility connections in many markets remain constrained. At the same time, AI computing loads can fluctuate sharply and persist for long periods, creating operating conditions that differ from those of more conventional data centre workloads.

That shift affects the design and operation of generators, alternators, battery systems and uninterruptible power supply architecture. Operators increasingly need to assess system stability, power quality, redundancy planning and battery cycling as AI training and inference place sustained stress on electrical infrastructure.

Brian Melka, President and Chief Executive Officer of Rehlko, said traditional critical power design is being challenged by these new load profiles. "For decades, power infrastructure was designed around capacity and redundancy. AI introduces a fundamentally different operating reality," he said. "Rehlko is optimizing its products and redefining AI readiness as the ability to maintain stability, control and performance under continuous dynamic loads with five nines reliability."

Load profile shift

Rehlko's central argument is that AI infrastructure planning now depends as much on dynamic behaviour as on nameplate capacity. In practice, that means operators must look beyond how much electricity a site can draw and focus more closely on how systems respond when demand changes constantly across computing clusters.

It says it works with customers from initial design through ongoing operation, covering power generation, controls, energy storage integration, distributed energy systems and modular infrastructure. The company frames that work around what it calls energy resilience, supporting both grid-connected and islanded data centre strategies.

Options highlighted include on-site gas generation, combined heat and power systems, battery energy storage, generator backup and modular power infrastructure delivered through its e-POD platform. Modelling tools, including digital twin methods where appropriate, can help operators test performance before deployment.

This focus on pre-deployment modelling reflects a broader industry concern about how AI facilities will perform in live operating conditions. Operators expanding AI estates must weigh the risk that equipment designed for older computing patterns may not respond in the same way under sustained and highly variable demand.

Broader market

Rehlko is an industrial energy resilience provider with operations across 80 locations worldwide. It says it has deployed more than 14GW of capacity, with more than 90% designed for specific applications, and serves hyperscale and non-hyperscale data centre operators as well as customers in healthcare, telecoms and industrial markets.

The company traces its roots back more than a century and says its lifecycle work covers engineering, manufacturing, deployment, servicing and operation of power systems. That long operating history forms part of its pitch to customers navigating AI expansion, grid instability and broader electrification trends.

For data centre operators, the immediate issue is how quickly they can bring power online without compromising reliability. AI roll-outs have intensified scrutiny of backup and bridging systems and their ability to maintain performance when workloads no longer follow predictable patterns.

Rehlko's latest publication is the first in a broader set of materials it plans to release on AI readiness. The initial eBook makes a simple point that is becoming harder for data centre owners to ignore: installed power capacity alone may no longer be enough to show whether a facility is prepared for AI demand.