The core tradeoff

Deploying rack-level UPS eliminates the electrical room — but it also changes how you size for growth, plan maintenance windows, and think about failure domains. Neither architecture is universally superior. The right choice depends on load density, site constraints, redundancy requirements, and how fast your environment needs to scale.

This guide walks through both architectures with enough specificity to inform real design decisions. If you’re looking for a feature checklist, the comparison table is below. If you’re trying to figure out which architecture is right for your deployment, start with the decision framework.

Architecture overview

Rack-based (distributed) UPS

Rack UPS systems are deployed directly within IT racks or adjacent enclosures. Power protection is distributed across the infrastructure, with each unit protecting a discrete rack or row. Capacity scales with the load — you add UPS capacity when you add compute, not before.

Because the UPS sits close to the load, branch circuit runs are short, distribution losses are minimized, and the failure domain is localized. A fault in one rack UPS affects that rack — not the row, not the floor.

This model is well-suited for environments where racks are factory-integrated with servers, networking, and PDUs before shipment. The UPS ships as part of the rack stack, reducing on-site integration work and eliminating a class of field wiring errors.

Centralized UPS

Centralized UPS systems are installed in a dedicated electrical room and distribute power to downstream loads through facility wiring. A single system — or a small number of large systems — protects the entire load. Redundancy is typically achieved at the system level through N+1 or 2N configurations.

This approach is appropriate where a centralized operations team manages power, where facility infrastructure already exists, or where the scale of the deployment justifies the upfront engineering investment. It concentrates both the protection and the single point of failure in one place — which can be a strength or a liability depending on how well that infrastructure is designed and maintained.

Design considerations

Load density and thermal management

High-density AI compute and GPU clusters are driving rack power loads well above traditional data center averages. As rack densities climb past 20–30 kW per rack, the efficiency of distribution from a centralized UPS becomes a more significant factor. Losses across long distribution runs add up — and at high density, they add up faster.

Rack UPS systems, positioned close to the load, reduce distribution path length and associated losses. At high rack density, this can meaningfully improve total efficiency.

Redundancy and failure domains

In a centralized architecture, the UPS system is the single protection point for a broad load. A well-designed centralized system with N+1 or 2N redundancy is robust — but a fault or maintenance event at the UPS level affects a larger scope of load than a localized rack failure would.

Distributed rack UPS architecture creates smaller, independent failure domains. A battery fault, inverter issue, or bypass event is contained to the affected rack. This doesn’t eliminate the need for redundancy planning — parallel-capable or redundant rack UPS configurations still require design attention — but it limits blast radius.

Maintenance windows

Centralized UPS maintenance typically requires either a planned bypass event affecting a wide load scope, or a parallel redundant system to cover the load during service. Rack UPS systems with maintenance bypass switches can be serviced at the rack level without affecting adjacent equipment. For environments with tight uptime requirements and limited maintenance windows, this is a meaningful operational consideration.

Physical space and cabling

Rack UPS architecture uses existing rack space and eliminates the need for a dedicated electrical room. In edge sites, colocation deployments, or space-constrained facilities, this is a practical advantage. Centralized systems require floor-standing cabinets, cable management for long distribution runs, and coordination with facility electrical infrastructure.

The cabling difference isn’t just about cost — long branch circuit runs increase exposure to voltage drop and require careful coordination between the electrical and IT teams. Rack-level power keeps that boundary simpler.

Scalability and capacity planning

Rack UPS systems scale incrementally. You size for what you’re deploying today, and add capacity when you expand. There’s no need to oversize the initial UPS installation to accommodate projected future load.

Centralized systems require upfront capacity planning. Undersizing creates a future constraint; oversizing means you’re paying for capacity you’re not using. For environments with predictable, stable load profiles this is manageable — for environments that scale rapidly or unevenly, it’s a consistent source of friction.

Comparison

How to choose

The following criteria will point you toward the right architecture for most deployments. If your situation spans both columns, the edge usually goes to rack-based — the deployment flexibility advantages compound over time.

The Ai90 as a deployed example

Platform reference

Ai90 Rack-Integrated UPS

The Ai90 integrates UPS conversion, hot-swap battery modules, maintenance bypass, and power distribution within a single rack footprint — eliminating the gap between power protection and compute infrastructure. It’s designed to ship as part of a complete rack stack, reducing on-site integration to a physical install.

Key specifications: modular battery architecture supports incremental runtime extension; integrated maintenance bypass enables service without load interruption; touchscreen controller provides local monitoring and management.

View product details Platform overview →

UPS conversion + touchscreen controller Hot-swap battery module array Redundant cooling / fan tray Integrated rack enclosure

Related comparisons

If you’re evaluating the Ai90 against specific alternatives, these technical comparisons cover performance, efficiency, and deployment differences in detail:

• Ai90 vs. Schneider Electric Galaxy VS →
• Ai90 vs. Eaton 93PM →
• Ai90 vs. Vertiv APM2 →