Modular UPS For AI Data CentersRedundancy Architecture

N+1 with shared control isn't N+1

A single control board. A single static bypass. A shared parallel bus. Each one a point where everything depends on one component.
Module count doesn't change that. Topology does.

Modular UPS For AI Data CentersLoad Behavior

The UPS spec passed.
The AI load didn't care.

When a training job starts, every GPU ramps simultaneously. The load concentrates — it doesn't diversify.
The real test isn't commissioning. It's the first production run.

Modular UPS For AI Data CentersLifecycle Economics

The cheapest UPS decision
is rarely the cheapest UPS.

30+ year design life. Components that don't need replacing for 15+ years.
The acquisition price is one line in a model that runs three decades.

Modular UPS For AI Data CentersIntegration & Deployment

The cabinet was designed for how AI facilities are built.

Bottom or top entry. Any busbar. Any chemistry. Airflow in either direction.
When the cabinet adapts to the installation, one engineering problem disappears from the list.

9 nines

Designed UPS Availability

97.6%

True online VFI efficiency.

30+ years

UPS design life

15+ years

Component replacement life

4-6 weeks*

Typical order to
delivery lead time

32 years

Mission-critical engineering

High-density AI infrastructure changes what matters in a UPS decision — and in what order.

The load concentrates. The thermal profile sustains. The decisions that matter are architectural.

Architecture, Not Only Redundancy

Module count doesn’t always eliminate failure paths. In AI environments, topology decides whether a single component can drop the entire critical load.

AI Load Behavior

GPU clusters don’t ramp gradually. Load moves from idle to sustained maximum simultaneously — across every rack, every time.

Maintainability Under Load

Continuous AI training means no maintenance windows. True safe hot-swap under live load isn’t optional — it’s the operational baseline.

Lifecycle Economics

At AI scale, every efficiency point and every replacement cycle compounds. A 30-year frame with 15+ year components changes what TCO actually measures.

Schedule & Supply Chain

GPU allocations have a commissioning date. When UPS delivery is measured in months, the facility isn’t ready when the hardware arrives.

Downtime is no longer a pause. It is an erasure.

In AI infrastructure, a power failure doesn’t pause the workload. It erases it. The cluster restarts from zero. The financial consequence arrives before the root cause is clear. And the infrastructure leader who approved the architecture is the first person in the room.

At $1M+ per significant outage, the financial consequence arrives before the root cause is identified.

The person who owns the architecture decision owns the outcome. That accountability is the reason this decision deserves the right evidence behind it

The question isn’t whether a power event is possible. It’s whether the architecture was designed so that when the grid fails, the load doesn’t even know.

In AI infrastructure, loads spike together. Failures can't. Enter DARA™.

When every GPU ramps simultaneously, the UPS sees a correlated load event traditional IT architectures were never sized for. Sustained. Concentrated. Unaveraged.

In a centralised design, one control board processes that event for the entire system. One component. System-wide consequence.

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.

Architecture That Makes Redundancy Real

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.

Distributed Decision Making

No single component decides for the whole system. Each module assesses, communicates, and responds collectively — so one fault never becomes a system-wide event.

Concurrent Maintainability

No bypass transfer. No production exposure. Service happens on the system’s terms, not the workload’s. Faulty modules are replaced under live load — the workload continues without interruption.

Scale Without Redesign

Each new AI pod adds compute capacity. The power architecture was designed to match that increment — no redesign, no load interruption, no stranded investment.
Download: UPS Architecture for AI Data Centers

Primary deployment model for AI data centers

ePOD is not a packaging option. It's how AI power infrastructure is built.

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.

Centiel works directly with POD manufacturers. The UPS adapts to the environment — not the other way around. AC IN/OUT via flange or cable, matched to the POD configuration. No design compromises on the pod’s thermal or structural architecture.

Typical delivery: 4–6 weeks.* The UPS is not the variable that slips.

1 MW per square metre

In AI data centers, the UPS footprint is a thermal budget decision.

At 40–100+ 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.

Centiel’s modular UPS delivers up to 1 MW per square metre. The power chain protects the load. The space around it protects the temperature.

Airflow designed for the installation, not the datasheet

Airflow configuration is an engineering decision, not a default setting.

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.

Centiel cabinets support front-to-top and front-to-back airflow. Configured at project stage to match the hot aisle / cold aisle layout. An airflow path that matches the pod’s thermal architecture disappears from the problem list entirely.

Bottom or top — single smooth bend, not a forced double

How the cable enters the cabinet is a thermal and mechanical engineering decision.

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.

Centiel’s cabinet is natively designed for bottom entry — a single smooth bend, the path the cable geometry naturally wants to take. No confined routing. No thermal derating.

Top entry is equally supported. The choice is driven by the facility design, not the cabinet.

Flanges, busbars, cables — configured to the project

The connection method adapts to the facility. The facility does not adapt to the cabinet.

Centiel supports full flange connection across all busbars — multiple flange types available to match the project’s busbar specification precisely. Or fully cabled. The connection method is a project decision, not a product constraint. For POD deployments, AC IN/OUT via whichever method the POD manufacturer requires.

The result is a clean installation that matches the design intent — not a workaround the site team has to justify for the life of the system.

Chemistry today. Chemistry tomorrow. Frame for 30 years.

The battery decision should not be locked in at UPS procurement.

Battery technology for AI data centers is not settled. Chemistry economics, fire safety regulations, and performance requirements are all evolving — what is right today may not be right at the first replacement cycle.

Centiel’s architecture is battery-agnostic: Lead Acid, Zinc, Lithium, and any future chemistry. The battery decision is not locked in at the beginning of a 30-year frame life.

That cycle should be driven by technology and economics — not by the UPS vendor’s design constraints.

Delivery as a critical path decision

When the GPU allocation has a commissioning date, the UPS cannot be the variable that slips.

In AI data center builds, time-to-revenue is a board-level metric. GPU clusters are ordered, financed, and expected to generate returns on a fixed schedule. Infrastructure that can’t match it converts capital investment into a waiting problem.

When UPS delivery is measured in months, commissioning contingency disappears. Critical path dependencies cascade. The facility isn’t ready when the hardware arrives.

Typical Centiel lead time: 4–6 weeks.* Production slot reserved at order confirmation. In ePOD environments where power integration, commissioning, and GPU installation are tightly sequenced, this is not a procurement preference. It is a schedule decision.

⚠️ Disclaimer: Subject to configuration, order confirmation, and prevailing supply conditions.

Design choices compound. In AI infrastructure, so do the consequences.

Architecture translates into operations. These are the measurable results.

Continuous operation under AI load

When a module fails during a live training run, the remaining modules absorb the load instantly. The fault is contained within the affected module. Remaining modules absorb the load. No bypass transfer. No load interruption.The training run continues. The fault is contained. Nothing stops.

Maintenance without production impact

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.

Density growth absorbed without redesign

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.

TCO that survives 30 years of AI infrastructure evolution

One capacitor change. No forced battery replacement cycles. No mid-lifecycle UPS replacement disrupting a live AI facility. The acquisition cost is one line in a model that runs three decades — and in that model, the architecture pays.

Schedule integrity from order to commissioning

When the power chain arrives on schedule, the commissioning window holds. GPU allocations find a facility that is ready.

In AI infrastructure, every role carries a different version of the same risk. Start from yours.

Different roles, different pressures, different entry points.

Eliminate failure paths from the topology

(Facilities Managers, Critical Infrastructure Leaders, MEP Engineers)

 

How DARA removes SPOFs at the architecture level — DDM explanation, SPOF analysis, and availability mathematics.

Validate the architecture against AI load behavior

(MEP Engineers, MEP Consultants, Critical Facilities Engineers)

 

How the architecture performs against correlated GPU ramp events and sustained high-density load profiles conventional load banks can’t simulate.

Protect the commissioning schedule

(MEP Consultants, Procurement, Operations & Engineering)

 

How 4–6 week typical lead time works in practice — production slots, delivery window, and what protects your critical path.

Build the lifecycle cost case

(Economic Buyers, Procurement, Finance)

 

30-year economics: acquisition, 97.6% efficiency, one capacitor change, no mid-lifecycle replacement — in a format finance accepts.

Confirm Tier III/IV pathway alignment

(MEP Consultants, Design Authorities, Engineers of Record)

 

Architecture and documentation aligned with Uptime Institute Tier III/IV power-path expectations. SPOF elimination evidence that defends your specification.

Two layers of proof: how the system is designed, and how it performs in the field.

The difference between a specification and a proof is a deployment.

Architecture and operation. Both documented.

Designed Not to Fail

The engineering foundation that eliminates failure paths before the first training job runs.

DARA distributed architecture

Designed for up to 99.9999999% availability. Distributed control verified across triple-redundant communication paths.

Standards alignment

IEC 62040 compliance. Architecture aligned with IEEE, NFPA, and Uptime Institute Tier III/IV power-path expectations.

Concurrent maintainability

Safe hot-swap under live load. No bypass transfer. No maintenance window. No production impact.

Modular scalability

Capacity added under operation, without redesigning the frame or interrupting the live load.

Proven in operation

Field performance, deployment record, and third-party recognition.

Not marketing summaries — measurable outcomes from named facilities.

Redcentric, Heathrow, London

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.

Taipower, Taiwan

Government procurement selection. Tight schedule delivered. Modular architecture supporting national AI, big data, and cloud infrastructure.

32 years in mission-critical power

Not a startup, not a rebrand. Three decades of UPS architecture development — and the engineering continuity that comes with it.

Listed on the SIX Swiss Exchange

Financial transparency and institutional accountability. The governance structure of a company built for 30-year infrastructure partnerships.

The evidence exists. Select the layer that moves your decision forward.

Different roles, different next steps.

Validate the lead time

4–6 week typical delivery. How production slots work, what’s required at order, what protects your commissioning window.*

Build the 30-year cost case

Acquisition, 97.6% efficiency at sustained AI load, 15+ year components, downtime risk — in a format finance accepts.

Confirm Tier III/IV power-path alignment

SPOF elimination evidence, concurrent maintainability documentation that defend your specification.

See AI Data Center deployments

Redcentric Heathrow. Taipower Taiwan. Zero downtime, schedule delivered, AI-ready architecture.
Download: UPS Architecture for AI Data Centers

Disclaimer: 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 lead time 4–6 weeks, subject to configuration, order confirmation, and prevailing supply conditions.

Download: AI Data Center Power Resilience Guide