A and B Power Feeds: How to Prove Data Center Redundancy Is Physically Independent

• Physically independent redundancy means the two paths share no strip, feed, panel, cassette, tray or conduit that can fail at once.
• In Uptime Institute's latest analysis, power caused 45% of respondents' most recent impactful outages in 2025, and outages linked to PDUs, transfer switching and single corded equipment increased significantly.
• Both supplies of a server can end up on one strip or one feed through ordinary installs, maintenance and moves, with no alarm.
• Each side of a dual path system must be able to carry the full load, so each PDU should stay below half its rating.
• Redundant network links that share a panel, cassette, tray or conduit form a shared risk link group and can fail together.
• Half seated cords, badly handled copper cables and mixed fiber types weaken one side without any warning.
• Proof means a physical audit against the power plan and cutsheet, recorded with photos, and repeated after every change.

Short answer: Two paths are physically independent when they share nothing whose failure can take out both: no common strip, feed, panel, cassette, tray or conduit. Monitoring tells you both paths are healthy; only a physical check tells you they are separate.

Redundancy is built in pairs: two power supplies, two feeds, two network links. The pair only protects you if one failure cannot reach both halves at once, and that depends on where the cords and fibers actually run, not on what the drawing says.

Large operators design for exactly this. Google's Evolve or Die paper describes surviving physical plant failures by using disjoint physical fibers and disjoint power feeds, so no single cut or outage reaches both sides. That is the standard to hold every rack to.

It does not always work out on the floor. In a study of Microsoft's data center networks, Gill, Jain and Nagappan found that network redundancy reduced the median impact of failures by only up to 40%. Redundancy on paper and redundancy on the floor are not the same thing.

Software cannot close that gap. A PDU reports its load, a switch reports its links up, and a server reports two healthy supplies, but none of them knows whether the other half of the pair shares the same strip, panel or tray.

Two healthy paths are a claim. Two separate paths are a fact you have to see.

Short answer: Power is still the leading cause. In Uptime Institute's 2026 analysis, power caused 45% of respondents' most recent impactful outages in 2025, and outages linked to PDUs, transfer switching and single corded IT equipment increased significantly.

Uptime Institute's Annual Outage Analysis 2026 is the closest thing the industry has to a scorecard. Power issues accounted for 45% of respondents' most recent impactful outages in 2025, down from 54% in 2024, and power remains the leading cause.

The detail matters more than the headline. Uptime reports that outages linked to power distribution units, transfer switching and single corded IT equipment increased significantly from its previous survey, and those are exactly the places where redundancy is either physical or imaginary.

People are part of the story. 92% of respondents said human error contributed at least in a minor way to their most recent impactful outage, and failing to follow established procedures remains the leading driver of those outages.

Put those together and the pattern is clear. A cord moved during maintenance, a server plugged into the nearest outlet or a strip swapped without updating the plan can turn a redundant rack into a single point of failure, and nothing alarms until the day the other side drops.

Networks follow the same pattern. Uptime also found that fiber and cable incidents were the leading cause of publicly reported IT service outages in 2025, which is why this guide covers both power and network paths.

Most redundancy fails quietly, long before the outage that proves it.

Short answer: A and B power feeds are two separate paths from the power source to the rack, each sized to carry the full load on its own. Dual corded servers take one cord from each, so either feed can fail without the server going down.

In a typical design, each path has its own UPS, distribution and rack PDU, often color coded so technicians can tell them apart at a glance. A server with two power supplies plugs one cord into the A strip and one into the B strip, and shares its load between them.

If the A path fails, the B path picks up the whole load, and the server never notices. That only works if every link in the chain is truly separate, from the upstream source to the outlet the cord is plugged into.

Two cords do not always mean redundancy

Not every device with two cords is redundant. Schneider Electric points out that some equipment uses a second cord for extra capacity rather than as a backup, so losing either cord can drop the device (vendor). Check the manufacturer's documentation before you count a device as protected.

Single corded devices are a separate case. They need a transfer switch or a rack level source to survive a feed failure, and Uptime's latest analysis names them among the outage causes that grew.

Upstream, the generators and transfer switches that back both paths have their own rules and tests, covered in our guide to emergency and standby power under NFPA 110. This guide stays at the rack, where the paths are easiest to break.

The design is an intention. The cords on the floor are the fact.

Short answer: Through ordinary work: a rushed install, a strip swapped during maintenance, a move that reused the nearest outlet, or two strips fed from the same source. None of these raises an alarm, because both supplies stay healthy until the shared element fails.

Five ways two paths quietly become one

1. 1. Both cords on one strip: the installer runs out of outlets on one side and plugs the second supply into the same rack PDU as the first. The server reports two healthy supplies, and one strip failure takes it down.
2. 2. Both strips on one feed: the rack has an A strip and a B strip, but after a change or a labeling mistake both are fed from the same upstream panel or circuit.
3. 3. The overloaded pair: each side runs past half its rating, so when one fails the survivor trips too, and the pair fails together.
4. 4. The shared element: two redundant network links leave the rack through the same patch panel, cassette, tray or conduit.
5. 5. The change that undid it: a move, add or change after commissioning puts a cord or link back on the wrong path, and nobody checks the rack again.

Each of these passes every software check. The PDU shows load, the server shows both supplies healthy and the switch shows both links up, so the defect stays invisible until the day it matters.

They are also easy to see. Cord color, strip labels and routing show which feed each supply lands on, and a cutsheet compared with the panel shows where each link runs.

Every one of these is visible at the rack. None of them is visible from a dashboard.

Short answer: Because each side has to carry the whole load if the other fails. If both PDUs run above half their rating, losing one pushes the survivor over its limit, and the redundant pair fails together.

In a dual path system, normal operation splits the load between the two sides. Schneider Electric's guidance is that each path must be able to carry the full downstream load, so no circuit should run above 50% of its design rating, and Vertiv makes the same point about rack PDUs (both vendor).

The failover math

• A normal day: two rack PDUs each carry 60% of their rating, and everything looks fine.
• The A side fails: the B side now has to carry 120% of its rating.
• The result: the B side trips, and the servers that were supposed to be protected go down with it.

Failover can add load as well. Schneider notes that fans speed up when a supply is lost, which can raise a device's draw by 10% or more, and Cabling Installation and Maintenance describes how load shifting to the surviving path can cascade until both fail.

Some engineers keep each rack PDU lower still, and TechTarget describes keeping each one below 40% of its rating. Your design rating and your electrical engineer's rule govern, and the number has to be checked on both sides together.

Load comes from power monitoring, not from a walk. But the walk is what tells you which servers sit on which side, so the monitoring numbers mean what you think they mean.

Half plus half only works if both halves are really separate.

Short answer: Yes. Two links that run through the same patch panel, cassette, tray or conduit share a single point of failure, and one pull, cut or disturbance can take out both. Network engineers call this a shared risk link group.

The term comes from network engineering. As IETF RFC 8001 defines it, links form a shared risk link group when they share a resource whose failure can affect all of them.

On a data center floor, that resource is usually physical. It can be a patch panel both links pass through, an MPO cassette carrying both, a tray they share or a conduit leaving the hall.

Where redundant links share a path

• Same patch panel: both links land on one panel, so one bad day at that panel affects both.
• Same cassette: both links ride in one MPO cassette, so one pulled or damaged cassette drops both.
• Same breakout cable: both links split from one trunk, so one cut takes both.
• Same tray or conduit: both run side by side, so one crush, cut or careless pull reaches both.

Maintenance is when shared paths get hurt. In Google's study, nearly 70 of 103 failure events happened while a management operation was in progress, and the Microsoft study found that its periods of most frequent device failures were driven by large scale maintenance.

Network software cannot see any of this. LLDP and cable validation confirm both links are up and connected to the right neighbors, not which panel or tray they pass through, as our guide to what network software cannot detect explains.

When a degraded link does fail over, the backup carries the traffic with less headroom, the pattern described in our guide to intermittent fiber link failures. If the backup shares a cassette with the link that failed, there is no backup at all.

Two links in one cassette are one link with two names.

Short answer: Yes. A half seated power cord, a copper cable bent tight at its connector or the wrong fiber type on a short link all work on day one and leave no margin, so the next knock, move or heat swing takes one side down.

Power cords and retention

A power cord that is not fully seated still powers the server. It works until someone brushes it during nearby work, and then one supply drops and the rack is quietly running on one side.

Retention clips and locking cords exist for this reason. Eaton sells inserts that give about ten times the retention force of a standard C13 connection, and locking C13 connectors cannot be pulled or vibrated out of the inlet (both vendor).

Copper cables

Direct attach copper and active electrical cables connect switches and servers over short runs, and they have handling limits of their own. NVIDIA's specifications give a minimum bend radius close to the connector, and its cable guidelines warn never to insert a cable by its pull tab (vendor).

A copper cable bent hard at the connector or left unlatched can pass traffic for months. Check it like fiber: no tight bends near the plug, fully seated, latched.

Mixed fiber types

Fiber types are color coded, but not uniquely. The Fiber Optic Association lists OM3 and OM4 multimode as aqua, OM5 as lime green and single mode as yellow, so the only way to tell OM3 from OM4 is to read the print on the jacket.

One wrong fiber type on a short link usually passes on day one. It leaves the link with no margin, so a little dirt or a tight bend later is enough to push it into errors.

Small physical defects turn two strong paths into one strong path and one weak one.

Short answer: Compare every cord and link with the power plan and the cutsheet, record each check with a photo and a location, and check again after every change. Leave controlled failover tests to qualified staff in planned windows.

Proof has three layers. The first is a physical audit against the plan, the second is a recheck after every change, and the third is a planned failover test that proves the design under load.

The physical independence audit

• Power: each supply on a different feed per the power plan; cord color and strip labels match the feed; no device counted as redundant that has one cord or uses two for capacity; cords fully seated, with retention clips or locking cords where specified.
• Network: redundant links in different panels, cassettes and trays; not split from the same breakout cable; recorded that way in the cutsheet.
• Copper and fiber: DAC and AEC cables free of tight bends near the connector, seated and latched; fiber type matches the plan, read from the jacket print.
• After every change: recheck the racks that were touched, starting with the cords and links nearest the work.

Record each check, not just the exceptions. A photo of the two cords on their two strips, tied to the rack's location, is what lets the next engineer trust the rack without walking it again, the same principle behind what a camera catches on any round.

Failover tests prove the design, but they are not a substitute for the audit. They are disruptive, they need qualified staff and a planned window, and they test the paths only on the day they run, while the audit catches the cord that moved last week.

The best time to start is at handover. Our data center commissioning checklist puts these checks into Levels 2 and 3, so the record exists before the first server goes live.

If it is not recorded with a photo and a location, it is an assumption.

Short answer: A FacilityOps robot walks every row and checks each server's two cords against your power plan and each redundant link against your cutsheet, then flags anything that shares a path. Every round is compared with the last, so a change that undid your redundancy shows up with before and after images.

FacilityOps AI (RoboSkill Inc.) builds an autonomous inspection platform for data centers. Software shows two healthy paths; we prove they are physically independent.

Our inspection robots run fully autonomously, and a technician can take manual control at any time. The platform is hardware agnostic: inspection robots and sensor hubs are the most common way we capture data, but it works with any capture source.

What the robot checks

• A and B feeds: a FacilityOps robot checks each server's two power cords against the power plan, from cord color and routing to the A and B strips, and flags any server whose supplies land on the same strip or feed.
• Network paths: redundant links are checked against the cutsheet and your design rules, and flagged where both run through the same panel, cassette or tray.
• Cords and retention: power cords visibly not fully seated, and missing retention clips, are flagged for a person to check.
• Copper cables: DAC and AEC cables are checked like fiber for tight bends, seating and latches.
• Fiber types: mixed fiber types on short links are flagged from jacket color and print.
• After every change: each round is compared with the last in our 3D digital twin, so a move that put a cord back on the wrong strip shows up with before and after images.

Built to be trusted

• We check the work. We never do the work. We never touch a cord, never test failover and never work in the power path. PDU load comes from your power monitoring, not from the robot.
• People stay in the loop: anything the robot cannot read clearly goes to a person, a closed door is recorded as not inspected, and a person reviews every finding before it reaches you.
• No false all clear: a rack is reported independent only when both paths were seen, with every cord and link in scope recorded with a photo and its location.
• Everything runs on site: detection and the 3D model run on the robot's onboard computer, and you own every image and record. Patent pending.

Where it stands today

Whole facility rounds, the 3D digital twin and change detection run in a production data center today. Since July 28, 2026, a FacilityOps robot has run inspection rounds at the CRSC Data Center, powered by Solutions IS, in Houston, completing 19 autonomous inspection runs and more than 1,600 observations in its first three weeks.

Rack level power and network path checks run on our own test rack, and a one row trial proves them on your racks, judged on your own data. Give us the power plan and cutsheet for one row, and we will show you what shares a path.

• A and B feeds: two separate power paths from the source to the rack, each able to carry the full load.
• Dual corded: equipment with two power supplies and two cords, one for each feed.
• Single corded: equipment with one power cord, which needs a transfer switch to ride through a feed failure.
• Rack PDU: the power strip in the rack that feeds the equipment, one per feed in a dual path design.
• Failover: the moment one path takes over the full load after the other fails.
• Shared risk link group: links that share a resource whose failure can affect all of them.
• Path diversity: routing redundant links through different physical elements so they cannot fail together.
• DAC: direct attach copper, a passive copper cable for short links.
• AEC: active electrical cable, a copper cable with electronics that extend its reach.
• Retention clip: a clip or insert that holds a power cord firmly in its inlet.
• OM3, OM4 and OM5: grades of multimode fiber; OM3 and OM4 share the aqua jacket color.
• OS2: a grade of single mode fiber, usually with a yellow jacket.

What are A and B power feeds in a data center?

Two separate power paths from the source to the rack, each with its own distribution and rack PDU, and each sized to carry the full load alone. Dual corded servers take one cord from each feed, so the server keeps running if either path fails. The protection only holds if the two paths really are separate all the way to the outlet.

Should both power supplies be plugged into the same PDU?

No. Plugging both supplies of a server into the same PDU or strip means one strip failure, tripped breaker or pulled plug takes the server down. Each supply should land on a different feed, usually the A and B strips, matched against the power plan and checked again after any change in the rack.

Do dual corded servers draw power from both supplies at once?

Usually yes. Most dual corded servers share their load between the two supplies in normal operation, so each side carries roughly half. If one side fails, the other carries everything, which is why each side has to be able to take the full load. Some devices instead use two cords for extra capacity, so check the documentation.

Why should each PDU be loaded to less than half its rating?

Because when one side fails, the other has to carry the whole load. If both PDUs run at 60% of their rating, losing one pushes the survivor to 120%, it trips, and the redundant pair fails together. Vendors such as Schneider Electric and Vertiv recommend keeping each side at or below half of its design rating.

What happens to the B feed when the A feed fails?

The B feed picks up the full load of every dual corded device. If the B side has the capacity and every device really has a cord on it, nothing goes down. If a device had both cords on A, or the B side was already loaded past half its rating, the failure spreads to the side that was meant to protect it.

How do I verify which feed each server power cord is on?

Trace each cord from the supply to the strip it plugs into, using cord color, strip labels and routing, and compare the result with the power plan. Record each rack with a photo of both cords on their strips. Recheck after every move, add or change, because one rushed swap can put both cords back on one side.

What is a shared risk link group?

A set of links that share a resource whose failure can affect all of them, such as one fiber route, conduit, patch panel or cassette. Two links in the same group look redundant in software but can fail together. Keeping redundant links in different groups is how network engineers make backup paths physically independent.

Can redundant network links fail together?

Yes, when they share a physical element. Two links through the same patch panel, MPO cassette, breakout cable, tray or conduit can be taken out by one pull, cut or crush. Network software shows both links up and healthy, so the shared path is only visible by checking where each link physically runs against the cutsheet.

Do I need locking power cords in a data center?

They are worth it wherever a loose cord could drop a server, especially in dense racks where people work close to the cords. Retention inserts and locking connectors hold standard C13 and C19 plugs in their inlets far more firmly than a plain push fit. Either way, check that every cord is fully seated.

What is the minimum bend radius for a DAC cable?

It depends on the cable, so check the manufacturer's specification. NVIDIA, for example, specifies a larger minimum radius close to the connector than for a single bend farther along the cable, both as multiples of the cable's diameter. Never bend a copper cable hard at the plug, and never insert it by its pull tab.

What color is OM4, OM5 and OS2 fiber?

Under the common color code, OM3 and OM4 multimode are aqua, OM5 is lime green and single mode fiber such as OS2 is yellow. Because OM3 and OM4 share a color, the only reliable way to tell them apart is the print on the jacket. Other colors are allowed when the print states the fiber type.

Can a robot check that A and B feeds are really separate?

Yes. A FacilityOps robot walks each row, checks every server's two power cords against the power plan from cord color and routing, and flags any server with both supplies on one strip or feed. It checks redundant network links against the cutsheet the same way, and it never touches a cord or tests failover.

↑ Back to contents

Redundancy is only as good as its weakest shared element. Two power supplies on one strip, two strips on one feed or two links in one cassette all look healthy right up to the failure that proves they were one path all along.

The fix is not more software. It is a physical check of every cord and link against the plan, a record with a photo and a location, and a recheck after every change.

Two green lights are a claim. Two separate paths are a fact you can photograph.

• Donnellan, Lawrence and Weinschenk, Uptime Institute: Annual Outage Analysis 2026, May 2026
• Gill, Jain and Nagappan, Microsoft Research: Understanding Network Failures in Data Centers, ACM SIGCOMM 2011
• Govindan and colleagues, Google: Evolve or Die: High Availability Design Principles Drawn from Google’s Network Infrastructure, ACM SIGCOMM 2016
• IETF: RFC 8001, January 2017
• Schneider Electric: Do two cords guarantee power redundancy for a device?, August 6, 2014 (vendor)
• Schneider Electric: How to ensure true redundancy in a dual path power distribution system, August 30, 2014 (vendor)
• Vertiv: How to avoid downtime with redundant rack PDUs (vendor)
• Julius Neudorfer, TechTarget: Do dual power supply servers increase redundancy?, 2008
• Matt Vincent, Cabling Installation and Maintenance: Overload protection in dual corded data center environments, August 1, 2014
• Eaton: Plug lock insert for C13 power cords (vendor)
• Argosy: IEC lock C13 connector (vendor)
• NVIDIA: QSFP DD direct attach copper cable specifications (vendor)
• NVIDIA: Cable installation and management guidelines (vendor)
• Fiber Optic Association: Color codes for fiber optic cables

• Data Center Commissioning Checklist: What to Inspect Before Handover, Trade by Trade
• What Network Software Cannot Detect About Physical Cabling, and Why LLDP Is Not Enough
• Emergency and Standby Power: Why NFPA 110, Not the Fire Code