Designed UPS Availability
True online VFI efficiency.
Designed, built and tested in Switzerland.
The load concentrates. The thermal profile sustains. The decisions that matter are architectural.
Module count doesn’t always eliminate failure paths. In AI environments, topology decides whether a single component can drop the entire critical load.
GPU clusters don’t ramp gradually. Load moves from idle to sustained maximum simultaneously across every rack, every time.
Continuous AI training means no maintenance windows. True safe hot-swap under live load isn’t optional, it’s the operational baseline.
GPU allocations have a commissioning date. When UPS delivery is measured in months, the facility isn’t ready when the hardware arrives.
One in five outages now costs more than €1M. That consequence lands before the root cause is identified while the cluster is still down and the GPU capacity it was built to run sits idle.
The person who owns the architecture decision owns the outcome. That is why this decision deserves evidence before the PO: the one-line diagram, the availability model, the field record.
A UPS for an AI facility is bought once and runs through several GPU generations. Two things decide whether it was the right call: what it costs over its life, and how much of the building had to change to fit it.
DARA removes that decision point entirely. When loads spike together, faults are contained at module level. Under correct installation and configuration, the load continues without interruption.
Redundancy only works if the architecture beneath it has no shared failure paths, no centralised control logic that can make a system-wide decision. DARA removes these at the topology level by design, not by configuration.
DDM™ replaces master-slave control. No single component decides for the whole system. Each module assesses, communicates, and responds collectively so one fault never becomes a system-wide event.
No bypass transfer. No production exposure. Service happens on the system’s terms, not the workload’s. Faulty modules are replaced under live load and the workload continues without interruption.
Each new AI pod adds compute capacity. The power architecture was designed to match that increment. No redesign, no load interruption, no stranded investment.
Prefabricated power pods have become the dominant deployment model for high-density GPU infrastructure compressing timelines, enabling repeatable design, and scaling in discrete increments alongside AI compute expansion.
At 40–130+ kW per rack, every square metre carries a thermal consequence. Space consumed by the UPS is space unavailable for CDUs, liquid cooling distribution, and thermal buffer capacity. A direct constraint on how many GPU racks the facility can operate.
GPU clusters run at sustained high load. The UPS thermal profile, how the cabinet dissipates heat and in which direction, must be compatible with the facility’s cooling architecture, not imposed on it.
Large cross-section cables, 300mm² and above, create mechanical stress and thermal consequences depending on how they enter the cabinet.
A forced double bend in a confined, enclosed space limits heat dissipation and accelerates cable degradation.
Full flange connection across all busbars, with multiple flange types to match the project’s busbar specification.
Or fully cabled. For POD deployments, AC IN/OUT follows whichever method the POD manufacturer requires.
Architecture translates into operations. These are the measurable results.
A module fails mid-training run. The fault stays inside that module while the remaining modules carry the load. No bypass transfer. No load interruption. The training run continues.
Planned service happens under live load. No bypass exposure, no unplanned interruption, no notification to the AI operations team. The MOP becomes routine. Emergency response becomes the exception.
When the next GPU generation arrives, capacity is added under operation without stranding existing investment. The frame that protects today’s cluster protects the next one.
One capacitor change. No chemistry lock-in at battery replacement. No mid-life UPS swap in a live AI facility. Acquisition is one line in a three-decade model, and there, the architecture pays.
How DARA removes SPOFs at the architecture level. DDM explanation, SPOF analysis, and availability mathematics.
How the architecture performs against correlated GPU ramp events and sustained high-density load profiles conventional load banks can’t simulate.
How 4–6 week typical lead time works in practice: production slots, delivery window, and what protects your critical path.
30-year economics: acquisition, 97.6% efficiency, one capacitor change, no mid-lifecycle replacement in a format finance accepts.
Architecture and operation. Both documented.
The engineering foundation that eliminates failure paths before the first training job runs.
Designed for up to 99.9999999% availability. Distributed control verified across triple-redundant communication paths.
IEC 62040 compliance. Power-path architecture aligned with IEEE, NFPA, and Uptime Institute Tier III/IV expectations, scoped to the UPS infrastructure.
Safe hot-swap under live load. No bypass transfer. No maintenance window. No production impact. Modules are replaced while the critical load stays protected.
Capacity added under operation, without redesigning the frame or interrupting the live load. The next cluster lands without a rip-and-replace.
Centiel was created to solve a structural problem in mission-critical power: systems that only reveal their weaknesses once they are live. This is the engineering intent behind every architecture decision that follows.
Field performance, deployment record, and third-party recognition.
Measurable outcomes from named facilities.
22 MW of legacy UPS replaced with zero downtime. Efficiency from below 90% to above 97%. Two halls now configured for AI workloads: 12 MW of additional power protection planned for 2026.
Government procurement selection. Tight schedule delivered. Modular architecture supporting national AI, big data, and cloud infrastructure.
Not a startup, not a rebrand. Four decades of UPS architecture development and the engineering continuity that comes with it.
Financial transparency and institutional accountability. The governance structure of a company built for 30-year infrastructure partnerships.
Different roles, different next steps. Facilities teams need the maintenance and SPOF evidence. Consultants need documentation that defends the specification. Procurement and finance need the lead time and the 30-year cost case.
4–6 week typical delivery. How production slots work, what’s required at order, what protects your commissioning window.*
Acquisition, 97.6% efficiency, 15+ year components, downtime risk, in a format finance accepts.
SPOF elimination evidence and concurrent maintainability documentation, scoped to the UPS power path. The material that defends your specification.
Redcentric, Heathrow: 22 MW replaced with zero downtime. Taipower, Taiwan: delivered on a tight schedule. Both on modular architecture ready for AI load.
Performance figures are based on architectural modelling and configuration-specific assumptions. Actual performance depends on system configuration, installation environment, maintenance practices, and operating conditions.
Tier certification is granted by independent third-party assessment bodies and depends on full compliance with facility design.
Typical factory lead time is four to six weeks from confirmed order and final technical release, subject to configuration, production capacity, and prevailing supply conditions.