The term “modular UPS” describes a physical format — power capacity delivered through stackable modules inside a shared frame — but it says nothing about whether the system has eliminated its single points of failure. For a Facilities Manager responsible for a live colocation environment, or an infrastructure lead deploying power for high-density AI workloads, that distinction is not a nuance. It is the architecture question that determines whether a fault in one module stays in one module, or cascades to the load.
What “Modular UPS” Actually Means — and Where the Definition Breaks Down
The word “modular” is applied to UPS systems in three distinct ways, and each describes a different level of actual architectural protection.
The first meaning is component-modular: individual components — typically batteries or certain electronics — can be replaced independently, but the core UPS remains a single, integrated system. The frame is modular in maintenance terms. The architecture is not.
The second meaning is capacity-modular: power modules are added to a shared frame to increase total output. The system grows in steps. This is closer to what most buyers expect when they specify a modular UPS — but it still does not tell you whether the control logic, static bypass, or communication bus is shared across those modules.
The third meaning — the one that changes the availability calculation — is architecturally modular: each module is a fully self-contained UPS. It has its own rectifier, inverter, static bypass, battery charger, control logic, and control panel. No component is shared between modules at the system level. A fault in one module cannot reach another.
The critical question is which of these definitions applies to the system being specified.
A common pattern in the market is a UPS that is capacity-modular — modules can be added to a frame — but uses a single centralised static bypass shared across all of them. That centralised bypass is a single point of failure. If it develops a fault during a maintenance event, the load is unprotected. The modules are modular. The system is not.
An N+1 frame that shares a central static switch between all modules is not N+1 in any meaningful architectural sense. The redundant capacity is real. The redundant protection path is not.
This is the architecture question that determines whether a “modular” UPS delivers the availability its availability claim implies.
“Modular” describes three different levels of protection — and only one keeps redundant capacity and redundant protection together. The DARA white paper is the framework for telling them apart.
True Modular UPS: Architecture That Eliminates Single Points of Failure
The Distributed Active Redundant Architecture — DARA — is the design principle underlying Centiel’s CumulusPower and StratusPower fourth-generation modular UPS systems. Each word in the name describes a specific architectural property, and each property directly affects the availability calculation.
Distributed means no single active component can act as a potential single point of failure. No single control board. No single static bypass. No single parallel bus. Each module within the frame is a fully independent UPS with Centiel’s self-isolating Intelligent Module Technology (IMT), containing all the building blocks of a complete UPS unit: rectifier, inverter, static bypass, battery charger, control logic, and control panel. A fault in one module stays in one module. It does not cascade to the system.
Active refers to how the system responds to a module-level fault. DARA‘s Distributed Decision Making (DDM) technology means there is no single component deciding for the entire system. In a conventional modular architecture where modules share centralised control logic, a fault at that control point can signal all modules to transfer load to bypass — regardless of whether the fault itself threatened the load. In a DARA system, each module communicates instantaneously with all remaining modules via the Triple Mode parallel BUS: three independent, triple-redundant communication paths. The affected module and the system respond together. The critical load remains on inverter throughout.
Redundant describes how additional modules in a DARA frame behave. Each module above the minimum required for the load is an active, load-sharing participant — not a passive standby. This has a direct operational consequence: any single module can be isolated, removed, and replaced while the system continues to support the full load on the remaining modules. No maintenance window. No transfer to bypass. No SLA exposure.
Centiel’s safe hot-swap protocol extends this further. Any replacement module introduced to a live frame is fully isolated and tested within the running system before it accepts any load. Faults in the incoming module are identified before integration — eliminating the risk of a replacement module causing a cascading fault in a live system.
The practical result is a UPS where the failure domain is contained by design at the module level, and where maintenance activities do not introduce system-level risk.
Modular UPS for AI Data Centres: Built for Loads That Don’t Behave
AI and GPU training workloads impose power demands that conventional UPS topologies were not designed to handle. The issue is not peak load — it is the speed and unpredictability of load change. GPU clusters do not ramp the way traditional IT loads do. They spike together, shifting data centre demand by tens of megawatts in milliseconds. A UPS architecture designed around load diversity cannot absorb that without transferring to bypass.
CumulusPower and StratusPower carry a 124% continuous overload rating. This means the system absorbs transient overloads without transferring the load — the UPS stays on inverter through the event. In an AI data centre environment where a bypass transfer exposes the load and stresses upstream generation, this is a functional requirement, not a performance headline.
Rack density is the second pressure point. Current AI deployments operate at 40 to 132 kW per rack. Next-generation GPU infrastructure is targeting 250 to 900 kW per rack. The cabinet architecture of Centiel’s modular systems is designed to scale to these densities without a full infrastructure redesign. Capacity is added by module, not by forklift.
Lead time is the operational constraint that is least visible until it is critical. Centiel’s production lead time is 4 to 6 weeks. Incumbent vendors in the critical power market regularly quote 12 to 24 months for large-format UPS systems. In an AI data centre deployment where GPU delivery schedules are compressed and revenue from compute depends on commissioning speed, that gap in lead time is a material operational variable.
What you are really evaluating is how the architecture behaves under a module fault and sudden AI load. The white paper shows how to compare that behaviour across systems.
Modular UPS for Colocation: Expand Without Downtime, Maintain Without Risk
In a colocation environment, the UPS is not maintained in isolation — it is maintained in a live, multi-tenant facility where every maintenance activity is a potential SLA event. The architecture of the UPS determines whether that is true.
Concurrent maintainability without bypass is the requirement. DARA achieves this by design. Module isolation happens at module level, not system level. A module is taken offline; the remaining modules absorb its share of the load; the maintenance activity proceeds; the module or its replacement is reintroduced and tested before it accepts load. The critical path through that sequence does not touch bypass at any point. Tenant feeds on adjacent modules are unaffected. No customer notification required. No service window negotiation.
Capacity expansion without downtime follows from the same architecture. For a colo operator moving from 30 kW to 50 kW per rack — or responding to a tenant requesting denser compute space — modules are added to a live frame during normal operations. There is no forklift upgrade. There is no forced service window with tenants. The frame grows while it runs.
49% TCO reduction over 15 years — this figure is grounded in the components, not in a financial model assumption. Centiel’s capacitors and fans carry a 15-year service life rating, compared to an industry standard of approximately 4 years. The UPS is designed to a 30-year operational life. Reduced mid-life component replacements, lower reactive maintenance burden, and reduced MTTR under fault conditions compound across the lifecycle to produce the TCO advantage. The architecture is the reason the economics are different.
Tier III/IV power path alignment: the DARA architecture is designed to support Tier III concurrent maintainability requirements. Each power path is independently maintainable without impact to the critical load. Specifying engineers should evaluate architecture fit against Tier III concurrent maintainability criteria; Centiel can provide technical documentation to support the evaluation.
Modular UPS Performance: Efficiency, Overload Rating, and Lead Time
97.1–97.6% efficiency — measured under double-conversion operation, the mode the system runs in continuously. At the Redcentric Heathrow deployment, UPS operating efficiency rose from below 90% to more than 97% following live replacement of legacy equipment.
124% continuous overload rating (CumulusPower and StratusPower) — the system absorbs transient overloads without transferring to bypass. For AI workloads where GPU synchronisation events create sudden load spikes, this means the UPS responds on inverter rather than exposing the load.
4–6 week production lead time — Centiel manufactures in Europe. Comparable-capacity UPS systems from incumbent vendors regularly carry lead times of 12 to 24 months. For any deployment operating against a compressed commissioning schedule, this is a material operational variable.
49% TCO savings over 15 years — the architecture advantage translates directly into lifecycle cost: lower component replacement frequency, reduced reactive maintenance, lower MTTR.
30-year design life — the system is designed to remain in service for the full facility lifecycle, not to be replaced at mid-life.
15-year capacitors and fans — the components that typically drive mid-life service events in conventional UPS systems. Industry standard for capacitors is approximately 4 years. Centiel’s components are rated to 15 years, eliminating the major mid-life service cycle that most operators budget for.
MEM partial load optimisation — Maximum Efficiency Management adjusts module engagement dynamically to maintain efficiency across variable load conditions. At the load levels typical of AI workloads — where utilisation cycles between training, inference, and idle — MEM prevents the efficiency loss that occurs when a fixed-topology system runs below its optimised operating point.
This is where architecture meets operations — concurrent maintainability, Tier III alignment and lifecycle cost. The white paper is the validation tool for that decision.
If you’re evaluating a live project, request an engineering review (ADR). Engineering-led · No sales pitch. Request an engineering review
Centiel Modular UPS Systems
CumulusPower
Three-phase true modular UPS with DARA architecture; engineered for environments requiring maximum availability, full concurrent maintainability, and zero single points of failure at any operating condition.
StratusPower
Three-phase modular UPS with DARA architecture; scalable from module level for demanding colocation and AI data centre applications, with documented deployments in critical facilities across more than 60 countries.
LiFePower
Lithium-ion battery system integrated with Centiel’s modular UPS architecture; longer service life, lower physical footprint, and a reduced maintenance burden compared to VRLA alternatives at equivalent capacity.
CumulusPower X1
Single-phase modular UPS with DARA architecture; scalable hot-swappable modules for critical single-phase environments where a single point of failure in the power protection layer is not an acceptable design position.
Frequently Asked Questions
Is a modular UPS automatically redundant?
Not by format alone — redundancy comes from architecture. In a Distributed Active Redundant Architecture (DARA) each module is a fully independent UPS with its own rectifier, inverter, static bypass, battery charger and control logic, so a fault stays contained in one module while the remaining modules keep supporting the load.
Can a modular UPS module be replaced without downtime?
Yes. With module-level concurrent maintainability you isolate, remove and replace a module while the remaining modules carry the full load — no transfer to bypass, no service window, and no tenant notification in a live colocation facility.
How does a modular UPS handle sudden AI/GPU load spikes?
A 124% continuous overload rating absorbs transient AI/GPU load swings on inverter, keeping the critical load protected instead of transferring to bypass when GPU clusters spike together.
What efficiency and lifetime cost can a modular UPS deliver?
97.1–97.6% double-conversion efficiency, with Maximum Efficiency Management (MEM) optimising across partial loads, and a 49% lower total cost of ownership over 15 years — driven by 15-year capacitors and fans and a 30-year design life.
What is the lead time for a modular UPS?
Centiel manufactures in Europe with a 4–6 week production lead time, compared with the 12–24 months common for comparable large-format UPS systems — a material advantage on compressed AI data centre build schedules.
Discuss Your Power Architecture
If you are specifying a modular UPS for a colocation facility, an AI data centre build, or a capacity expansion in a live environment, the architecture question is the right starting point. Centiel’s engineering team can review your load requirements, redundancy targets, and lead time constraints — and provide a technical assessment of which architecture fits the operational reality of your site.