# Redundant Power Feeds Reliability in a data hall is achieved by arranging that nothing is unique. Power reaches each rack by two independent paths, and each path is built so that any component in it can be shut down, serviced, or fail outright while the other path carries the load. The discipline has a name worth remembering: there should be no single element whose loss stops the work. > [!note] The one-sentence version > Build two complete paths, feed every machine from both, and then any one thing can be taken out of service on a Tuesday afternoon without anyone noticing. ## Why it exists Equipment requires maintenance, and maintenance requires de-energization. A facility with one path must choose between deferring maintenance and accepting outages, and both choices degrade over time. Two paths convert maintenance from an interruption into a routine operation, and they convert a component failure from an incident into an event the monitoring system records. The same arrangement answers the harder question of what happens when something fails unexpectedly, because the answer is already proven every time maintenance is performed. ## Prerequisites - **N, N+1, 2N, 2(N+1):** N is exactly enough capacity; N+1 adds a spare unit; 2N is two complete independent systems; 2(N+1) is two systems each carrying its own spare. - **Concurrently maintainable:** any single component can be removed from service for planned work with the load fully supported. - **Fault tolerant:** the facility survives an unplanned failure of any single component, including during maintenance, which requires more than concurrent maintainability. - **Dual-corded equipment:** servers and switches with two power supplies, each fed from a different path. - **Static transfer switch:** a fast electronic switch providing two sources to equipment that has only one power input. - **Common-mode failure:** a fault that defeats both paths at once, such as a shared controller, a shared room, a shared fuel supply, or a shared operator error. Finding and eliminating these is the real work of redundancy design. ## How it is built Two paths are established at the earliest point the design allows, ideally at two separate utility sources into the campus substation, and they are kept physically and electrically separate the whole way down: separate transformers, separate switchgear sections, separate uninterruptible power systems, separate busways, separate rack power distribution units entering the rack from opposite sides. Each path is sized to carry the full load alone, which means each runs at roughly half capacity in normal operation; that apparent inefficiency is the price of the guarantee, and in modern designs part of it is recovered through operating modes that improve conversion efficiency at partial load. The discipline extends past electricity. Cooling is arranged on the same logic, with redundant pumps, redundant coolant distribution units, and valved piping so that a loop can be isolated; see [[wiki/Data Center Cooling|Data Center Cooling]]. Networking follows it too, with two fabrics and two paths out of the building; see [[wiki/Fiber Route Diversity|Fiber Route Diversity]]. A facility is only as redundant as its least redundant dependency, and the dependency people forget is usually controls: a single building management system commanding both paths is a common-mode failure wearing a redundant costume. ## Stack position - **Upstream:** utility sources, generation, and equipment supply, since redundancy roughly doubles the equipment count. - **Downstream:** every machine in the hall, and the service commitments made to whoever runs on them. - **Function and vendor:** the redundancy class is a commercial decision about what continuity is worth; several industry classification schemes exist to describe it in contracts. - **Degraded case:** a single-path facility is cheaper to build and spends the saving on outages and deferred maintenance. ## Lineage and trajectory The practice came from telephone exchanges and hospitals, where continuity preceded computing by decades, and was codified for data centers through tier and class systems from the 1990s onward. The current movement is subtle: as workloads become distributed across many facilities, some operators deliberately build less redundancy per site and achieve continuity at the fleet level instead, accepting that a building may fail if the service does not. Training workloads with checkpointing tolerate that trade well; see [[wiki/Checkpointing|Checkpointing]] and [[wiki/Failure Domain|Failure Domain]]. Both strategies are rational, and the choice is now made explicitly rather than by default. ## Relationships - **implemented through:** [[wiki/Electrical Substation|Electrical Substation]] · [[wiki/Medium-Voltage Distribution|Medium-Voltage Distribution]] · [[wiki/Uninterruptible Power Supply|Uninterruptible Power Supply]] · [[wiki/Rack Power Distribution Unit|Rack Power Distribution Unit]] - **coordinated by:** [[wiki/Automatic Transfer Switch|Automatic Transfer Switch]] - **parallel discipline in:** [[wiki/Data Center Cooling|Data Center Cooling]] · [[wiki/Fiber Route Diversity|Fiber Route Diversity]] - **traded against:** [[wiki/Failure Domain|Failure Domain]] · [[wiki/Capital Intensity|Capital Intensity]] - **validated by:** [[wiki/Commissioning|Commissioning]] - **source articles:** [[articles/Vertically Integrating an AI Superpower from AI Factory to Citizen|Vertically Integrating an AI Superpower from AI Factory to Citizen]] · [[articles/The Gigawatt Is the New Measure of War|The Gigawatt Is the New Measure of War]] - **collection route:** [[collections/American Infrastructure|American Infrastructure]] · [[collections/American|American Collection]] ## Evidence posture - **Established:** the N, N+1, 2N, and 2(N+1) classes; the distinction between concurrently maintainable and fault tolerant; dual-corded equipment and static transfer switches; common-mode failure as the governing design concern; industry tier and class classification schemes. - **Established (datestamped):** as of October 2026, independent A and B critical power and cooling paths with redundant uninterruptible power are standard in operator disclosures across the industry. - **Strongly indicated:** fleet-level redundancy substituting for site-level redundancy in some distributed deployments; operator architecture statements support it and comparative availability data is not public. ## Sources - Data center reliability classification practice and electrical standards, as current October 2026.