In a hypothetical, illustrative scenario—not an actual incident—utility work at 02:13 drops the A feed for one row. Six GPU servers remain online, yet the B-side rack PDU alarms immediately and its upstream circuit approaches a trip condition. Nothing is wrong with the individual power modules: each 3000w server power supply is operating within its DC rating. The problem is that every server transferred its full demand to the same surviving AC path at nearly the same moment.
This is the failure-domain change introduced by 3kW-class modules. The PSU rating still matters, but rack survival increasingly depends on what the surviving feed, branch circuit, PDU, receptacles, cords, and connectors can carry after redundancy collapses onto one side. Designing only from the normal A/B split can produce a rack that looks comfortable in steady operation and becomes electrically fragile during the exact event redundancy is meant to cover.
The number on the PSU describes DC output
A 3000W module rating normally describes the maximum regulated DC power the PSU can deliver under a manufacturer-defined operating envelope. That envelope may specify acceptable AC voltage and frequency, inlet temperature, airflow, altitude, platform configuration, firmware, and other conditions. The complete output may be available only on particular input ranges. A module could have a lower output limit, different behavior, or no approved operation at another voltage.
The 3000W DC figure is therefore not the same as 3000W drawn from the wall. AC input must cover both useful DC output and conversion losses. If a server receives 2,400W of DC power and its power conversion efficiency is illustratively assumed to be 95% at that operating point, its approximate real AC input power is:
2,400W ÷ 0.95 = 2,526W AC
This arithmetic is useful for modeling, but assumed efficiency should not replace manufacturer data or a measured input value. Efficiency changes with load, voltage, temperature, and implementation. Platform fans, PSU fan behavior, and power measured at different boundaries can also make apparently similar readings disagree.
AC current introduces another layer. A rough single-phase estimate may divide real input power by voltage and power factor. With an illustrative 0.99 power factor, 2,526W at 240V would be about 10.6A:
2,526W ÷ (240V × 0.99) ≈ 10.6A
At an assumed 208V, the same real power and power factor would imply approximately 12.3A. Voltage at the rack can vary, however, and the PSU data sheet may state a maximum input current based on conditions different from this simplified example. Circuit and PDU planning should use the applicable manufacturer input specification, approved platform envelope, site voltage range, and local electrical requirements rather than treating calculated nominal current as a universal design value.
Normal sharing conceals the surviving-feed load
Consider a server with two 3000W hot-swap modules configured for 1+1 redundancy. During normal operation, the pair may share the load. If the server requires 2,400W DC, each module might carry approximately half under balanced conditions. That does not mean the server has a redundant 6,000W budget.
For 1+1 operation to preserve the workload after one module or one feed is lost, either installed module must be able to support the full required server load by itself, within the approved limits of both the PSU and platform. The relevant redundant capacity is one module’s usable capacity under the actual input and environmental conditions.
If a platform consumes more than the single-module operating limit, two modules may support it while both are present, but the state is no longer full-load 1+1 redundancy. Following a module or feed failure, the platform might throttle, shed accelerators, restrict power, or shut down. The exact response depends on server firmware and power policy; it should not be inferred from the sum of the PSU nameplates.
At rack level, the same distinction becomes more consequential. In normal operation, six servers may divide their input between A and B. Each PDU displays only the share carried on its side. When A disappears, the B-connected modules in all six servers can become responsible for nearly the entire rack demand. The B feed has not gained capacity merely because the transfer was expected.
A conditional six-server transfer
The following example is illustrative, not a circuit recommendation. It assumes six identical servers, two 3000W modules per server, a valid 1+1 platform configuration, and one module connected to A while the other connects to B. It also assumes each server requires 2,400W DC at the tested workload, conversion efficiency is 95%, power factor is 0.99, and rack voltage is 240V. Real projects must substitute approved and measured values.
| Modeled quantity | Per server | Six-server rack |
|---|---|---|
| Server DC demand | 2,400W | 14,400W |
| Approximate AC real power | 2,526W | 15,158W |
| Approximate total AC current at 240V and PF 0.99 | 10.6A | 63.8A |
| Illustrative current on each feed before failure, assuming equal sharing | 5.3A | 31.9A per side |
| Illustrative current on B after complete A-feed loss | 10.6A | 63.8A on B |
The operational trap is visible in the last two rows. A planner observing roughly 32A on each side may conclude that both paths have ample room relative to a selected architecture. Yet the failure state asks one path to carry approximately 64A, before considering voltage variation, unequal power sharing, measurement tolerance, additional rack equipment, or workload growth.
The correct engineering comparison is not between normal PDU current and the PDU’s printed headline rating. It is between the credible post-failure demand and every applicable limit in the surviving path. Those limits can include the upstream branch circuit, protective device behavior, conductor and connector ratings, PDU input, PDU outlet groups, individual outlets, cords, PSU inlets, and site rules for continuous loading. Because installation standards and electrical architectures differ, a generic article cannot turn that comparison into one universal breaker or circuit recommendation.
The whole rack does not transfer as a gentle average
Spreadsheet totals usually describe a settled state. An A-feed loss is a dynamic event. Multiple power supplies detect the loss, use available hold-up energy, and increase output contribution from their surviving partners within short, overlapping intervals. The rack PDU and upstream protection see the aggregate behavior, not six independent events politely spaced over several minutes.
This transfer should not automatically be treated as identical to cold-start inrush. In a live 1+1 system, the surviving modules may already be energized and sharing current, while server DC buses remain active. A cold start can involve input capacitor charging and platform boot sequencing that do not occur in the same way during feed loss. Conversely, transfer behavior can still include brief current changes, control-loop response, temporary overload handling, or power capping.
The relevant evidence comes from PSU and platform documentation plus an instrumented test of the intended configuration. Important time-domain details include the server’s ride-through behavior, whether the surviving module accepts the full load without a DC interruption, how quickly sharing changes, and whether the PDU or upstream device reports or reacts to the aggregate event. Slow facility telemetry may show only the stable before-and-after readings and miss a short peak.
Protective devices also do not behave as simple digital gates at a single displayed amperage. Their response depends on device characteristics, duration, ambient conditions, installation, and applicable engineering rules. That is another reason not to derive branch protection from nameplate watts alone.
The outlet and cord can become the narrowest boundary
A rack may have sufficient aggregate PDU capacity while one outlet, outlet bank, plug, cord, or PSU inlet is used outside its approved current or temperature conditions. This risk rises when each surviving PSU can draw several kilowatts and when rack cabling has been selected from normal shared load rather than single-feed operation.
The cord set must match the PSU inlet and the selected PDU outlet electrically and mechanically, but connector fit is not proof of suitability. The cord, plug, receptacle, and PDU outlet each have their own approved ratings and installation conditions. Cable routing matters as well. Dense bundles and restricted rear-rack airflow can raise local conductor and connector temperatures even when a room sensor reports an acceptable ambient value.
Voltage drop deserves attention because it can reinforce current stress. For a regulated server load, lower input voltage generally requires more current to deliver similar real power. Resistance in a long or marginal path contributes additional drop and heating. The practical evaluation should therefore include voltage at the rack under transferred load rather than only the nominal facility voltage.
PDU metering granularity also changes what can be learned. An input-level reading can identify total feed demand but hide a heavily loaded outlet group. Per-outlet telemetry can expose imbalance, although its accuracy, update interval, and alarm behavior still need to be understood. Neither type of meter overrides the PDU manufacturer’s installation limits.
Three temperatures belong in the decision
Rack power discussions often reduce heat to the room’s total cooling capacity. A 3kW-class PSU requires a more local view. At least three thermal conditions influence a credible failure assessment: PSU inlet air temperature, conversion losses inside the module, and connector or cable temperature along the surviving AC path.
Using the earlier illustrative 2,400W DC and 95% efficiency assumptions, conversion loss is about 126W per server:
2,526W AC − 2,400W DC ≈ 126W
That figure is only an assumed operating-point estimate. It is not a claimed efficiency for any Linklieo or third-party product. Nevertheless, it shows why efficiency affects both input current and local thermal load. Nearly all electrical power entering the server ultimately appears as heat in the data hall, while PSU conversion loss is concentrated near the module and its airflow path.
Inlet temperature may be higher than the nominal cold-aisle reading because of recirculation, blanking gaps, rear-to-front leakage, adjacent equipment, or airflow restriction. Manufacturer output ratings can be conditional on inlet temperature and may include derating or protection behavior outside specified regions. A design that provides single-module capacity in a cool lab may not provide the same margin at the hottest rack location.
Feed loss can also alter PSU fan operation. The surviving module moves from partial contribution to full contribution and may increase fan speed. If airflow was already constrained, the event changes both electrical and thermal conditions together. Acceptance testing at an unrealistically light workload or unusually cool inlet can therefore validate the wrong operating point.
Power policy defines whether redundancy survives peak compute
AI, GPU, storage, and telecom servers rarely present one fixed demand. Accelerator utilization, CPU boost states, storage activity, fan response, and firmware limits can change system power over short and long intervals. Average rack consumption may be well below the single-module boundary while a sustained training phase or a simultaneous fan ramp approaches it.
A useful rack model separates at least three states: representative production demand, an approved maximum or capped demand, and post-failure demand on one feed. If platform power capping is part of the design, the cap must be persistent and effective during PSU or feed failure. A management policy that is disabled after firmware reset, lost during controller communication failure, or configured differently across nodes should not be counted as firm electrical capacity.
Power capping can be an intentional architectural control rather than a last-minute workaround. For example, if measured workloads occasionally exceed the single-module envelope, a platform-supported cap may preserve 1+1 operation at reduced performance. The tradeoff should be explicit: the organization is choosing bounded performance during all relevant states in exchange for electrical redundancy. The cap value still has to account for platform behavior and PSU input conditions, not merely match 3000W because that number appears on each module.
Acceptance should recreate the failure domain
A meaningful acceptance run begins with the intended firmware, PSU population, cords, PDU paths, rack position, and server power policy. The workload should reach a documented production-relevant state and remain there long enough for temperatures and fan behavior to stabilize. Record per-server input power where available, both feed currents, voltage at the PDU, PSU inlet temperature, and any platform-reported DC or power-cap data. The boundaries of each measurement should be noted because server telemetry and PDU telemetry do not necessarily measure the same point.
With personnel following the site’s approved operating and safety procedure, the planned feed-loss test should remove one power path in a controlled way. The purpose is to observe the simultaneous rack response: server continuity, single-module load, surviving PDU current, outlet or branch alarms, voltage movement, thermal behavior, and any throttling or event logs. Restoring the feed also deserves observation because rebalancing and module re-entry are separate transitions.
A second run at the hottest permitted or realistically reproduced inlet condition can reveal a conditional output or cooling issue hidden by the first test. Where a full live-feed test is not authorized, staged load-bank, platform, or facility testing may provide evidence, but its differences from the final rack need to be documented. Passing a single server test proves little about six servers transferring together onto one PDU.
The acceptance record should also distinguish a PSU removal from an A-feed loss. Pulling one module tests a local single-PSU event. Opening the rack’s A path tests common-mode transfer across every attached server and exposes the larger failure domain. Both are informative; they answer different questions.
The procurement boundary now extends to the rack path
When sourcing a 3kW CRPS or redundant hot-swap module, wattage and form-factor language alone do not establish interchangeability. Mechanical fit, electrical interfaces, communication behavior, firmware support, airflow direction, approved input range, output availability, and platform qualification all matter. Manufacturer documentation should identify the conditions under which the claimed DC output is available and the host should explicitly support the selected module.
For Linklieo customers, the productive request is therefore broader than “supply a 3000W unit.” It should state the host platform, required redundancy mode, expected DC demand, available AC range, inlet environment, and relevant integration constraints. That information connects module selection to the failure state the rack must actually survive.
If one approved 3000W module can carry each server’s capped worst-case demand at the site’s lowest expected input voltage and highest permitted inlet temperature, and the complete B path can carry every server’s simultaneous transferred AC current within all applicable limits, then 1+1 deployment is technically coherent. If either condition fails, the decision should be to reduce or cap server demand, redistribute nodes across additional failure domains, or redesign the AC path before treating the rack as redundant.
