Two healthy server power supplies and two energized rack power distribution units (PDUs) do not guarantee that a rack will survive losing one electrical feed. Both sides can look comfortably below their limits until maintenance or a fault forces the surviving side to supply nearly the entire protected load.
The relevant capacity is what one remaining power path can safely deliver, not the sum of the A and B feeds. For older rooms, HW Server’s analysis of data-center retrofit limits explains why inherited panels, distribution equipment and incomplete monitoring can constrain upgrades.
What A/B Power Redundancy Is Meant to Protect
In a conventional dual-fed rack, one power cord from each suitably configured server connects to the A-side rack PDU and another connects to the B-side PDU. Those PDUs should lead to independent upstream distribution paths across the failure or maintenance events the design is intended to withstand. Two differently labeled outlets connected to the same upstream breaker do not provide that independence.
The Uptime Institute’s tier classifications distinguish concurrent maintainability from fault tolerance. Tier III is designed to let infrastructure components and distribution paths be taken out for planned maintenance without interrupting IT operations; Tier IV additionally addresses defined unplanned failures. A rack with two cords is not, by itself, evidence that its upstream topology meets either objective.
This architecture matters because planned maintenance often creates the very condition that exposes a bad capacity assumption: one side is intentionally unavailable, and the other has nowhere to send excess load.
The Rack That Appears to Have 12 kW but Can Protect Only 6 kW
Consider a hypothetical rack whose A and B feeds can each safely carry 6 kW of continuous load, after the site’s applicable equipment ratings and electrical limits have been accounted for. The connected equipment is expected to draw 8 kW under its approved demanding operating condition.
Hypothetical illustration assuming equal load sharing before failure and full transfer of protected load afterward. Actual configurations require site-specific assessment.
If the servers share power evenly, metering might show 4 kW on A and 4 kW on B. Both meters look healthy. But after A is lost, B may be asked to carry approximately 8 kW, which is 2 kW more than its 6 kW usable capacity.
The installation is not an 8 kW fault-tolerant rack merely because its two normal-feed capacities add up to 12 kW.
Rack-power manufacturer Raritan describes precisely this circuit-failover overload problem: individually acceptable circuit readings can conceal a combined load that exceeds the surviving circuit’s limit.
The figures above are illustrative, not measurements from a real deployment. Actual assessments must check phase-level current, upstream protective devices, cord and connector ratings, and any required margin. Nameplate PSU wattages and a quiet-hour average are not substitutes for an approved load assessment.
A Second Trap: Redundant PSUs Do Not Always Split Demand 50/50
Even the 4 kW/4 kW example assumes balanced input power. Some servers deliberately favor one supply for efficiency.
In its PowerEdge power-settings guidance, Dell distinguishes nonredundant, PSU-redundant and A/B-grid-redundant modes, and explains that a hot-spare setting can place most demand on the active units rather than dividing it evenly.
That changes what an operator should expect from normal PDU readings. An A-side feed at much higher utilization than B might reflect the selected PSU mode, not a failed meter. It also means a feed-loss calculation should use the equipment’s documented transfer behavior instead of assuming every server shares power equally.
A further complication is PSU output derating at different input voltages. Dell documents examples in which a supply’s available output is lower on low-line AC input than on high-line input.
A server might have two installed supplies but lack enough single-supply capability at the actual feed voltage after a hardware upgrade. Check the specific server and firmware documentation before treating PSU redundancy status as proof of rack resilience.
Where the A/B Protection Stops
Servers with a single power cord cannot benefit directly from two rack PDUs.
Mitsubishi Electric’s explanation of 2N UPS architectures describes independent A and B systems and the role of a static transfer switch in serving single-corded loads. A properly selected transfer device can switch eligible equipment between sources, but its transfer tolerance, capacity, upstream connections and own failure modes still need assessment.
Independence must be traced end to end. Separate rack PDUs do not solve a common upstream switchboard, breaker, UPS output or incorrectly grouped distribution path.
A partially failed PSU, an unavailable transfer switch or an out-of-service branch can also remove the redundancy that the rack diagram appears to show.
Check the Surviving Path Before Approving Maintenance
A useful operational review answers five questions without relying on an improvised live-cord pull:
- Where does each cord actually lead? Reconcile PDU and branch labels with the current electrical one-line diagram, including shared components and maintenance boundaries.
- What is the weakest approved capacity limit? Evaluate the surviving UPS/distribution path, branch circuit, rack PDU, phase, outlets, cords and each device’s usable PSU capability. Use the electrical design team’s applicable continuous-load requirements, not an assumed universal derating percentage.
- How high can the protected demand become? Combine calibrated rack-level measurements, appropriate workload peaks, known PSU operating modes, planned additions and equipment-specific transition behavior. Monitor watts and per-phase current where applicable.
- Which devices will not transfer? Identify single-corded loads, failed or disabled redundant PSUs, transfer-switch dependencies and any equipment that cannot maintain service through the intended event.
- What evidence proves the result? Require an approved commissioning or maintenance procedure, competent electrical and IT oversight, defined acceptance limits and a documented rollback plan. Never simulate a feed failure by casually disconnecting live rack power.
For substantially different high-density power designs, its 800 VDC readiness checks address the wider engineering review; neither article replaces an A/B failover-capacity assessment.
The operational rule is simple: a redundant rack’s usable protected load is bounded by the surviving complete power path.
Document that limit before filling the last empty rack unit or taking one feed out of service. A green status light proves power is present now, not that the alternate path can sustain the next fault.
