Artificial intelligence is forcing a quiet but profound rethink of how data centres are powered. As computing density rises and workloads run continuously rather than intermittently, electrical infrastructure is emerging as one of the decisive factors determining whether AI systems operate reliably at scale. New developments in modular uninterruptible power supply architecture, outlined in recent technical guidance from Legrand, illustrate how power continuity is becoming inseparable from AI performance itself.
High-density computing environments increasingly depend on uninterrupted electrical supply, not simply to prevent outages but to maintain stable operating conditions for sensitive processors and accelerated workloads. The Keor FLEX UPS architecture has been introduced as a response to these evolving requirements, designed specifically to support environments where availability, resilience and energy efficiency must coexist.
Rather than relying on traditional monolithic power protection systems, the design adopts a modular structure intended to remove single points of failure and allow infrastructure to evolve alongside growing computational demand.
Redundancy redesigned for continuous computing
The architecture distributes intelligence, power conversion and redundancy across multiple subsystems, ensuring that faults or maintenance operations do not interrupt supply to critical loads. Power modules operate in parallel and can be replaced while the system remains live, allowing remaining modules to compensate instantly if one component fails.
Redundancy extends beyond the power modules themselves. Dual communication buses maintain synchronisation between system components, preventing coordination failures caused by a single communication fault. Auxiliary subsystems, including low-voltage power supplies and monitoring circuits, are also duplicated to prevent supporting systems from becoming vulnerabilities.
An automatic static bypass operates independently from the main controller through its own digital control unit, ensuring continuity even in the event of a major system failure. The overall architecture has been assessed using Failure Modes, Effects and Criticality Analysis methods to verify reliability across operational scenarios.
For AI data centres, where downtime can interrupt training cycles or halt automated decision systems, such layered redundancy reflects a growing emphasis on continuous operation rather than simple backup protection.
Maintenance without interruption
Traditional maintenance procedures often require scheduled downtime, a constraint increasingly incompatible with always-on AI workloads. The system’s hot maintenance capability allows technicians to replace modules without interrupting service, enabled by hot-swappable components and secure connector design intended to prevent electrical arcing.
Technicians retain direct access to connectors for inspection and monitoring, allowing thermographic checks and component replacement while critical loads remain powered. This approach reduces operational risk while supporting continuous availability, an increasingly important requirement as AI systems operate around the clock.
Managing unpredictable AI demand
Artificial intelligence workloads frequently generate rapid fluctuations in power demand, particularly during training or inference spikes. The system is designed to stabilise these variations by drawing temporarily on internal battery reserves when consumption peaks occur, reducing stress on generators or external networks.
After demand subsides, batteries recharge automatically through controlled micro-cycles of charging and discharging. This process maintains stable voltage and frequency while preserving power quality across changing load conditions.
As rack densities increase, scalability becomes another critical factor. The architecture allows internal expansion up to 1,200 kW through the addition of modular power units, while multiple systems can be operated in parallel to reach capacities of 4.8 MW. Such scaling is intended to accommodate hyperscale and AI supercomputing environments where energy requirements continue to grow.
Security as part of power reliability
Cybersecurity considerations are also integrated into the system’s design. Physical access controls restrict entry to authorised personnel, while operational commands must be performed locally following authentication. Remote supervision is supported through isolated networks, but remote control functions are intentionally excluded.
Data exchanges are encrypted and tied to certificate validation, and communication systems operate independently from the core UPS functions. Even if internal communications were compromised, essential power delivery would remain unaffected.
These safeguards reflect the increasing convergence of operational technology and cybersecurity concerns as power infrastructure becomes digitally managed.
As artificial intelligence reshapes computing, the resilience of electrical systems is emerging as a strategic issue rather than an engineering detail. The shift towards modular, redundant and secure power architectures signals a broader reality: the success of AI increasingly depends not only on algorithms and processors, but on whether the infrastructure beneath them can deliver uninterrupted energy in an era defined by continuous computation.
