The Engines of the Hall [/ðə ˈɛnʤənz əv ðə ˈhɔl/] n - A chain of equipment whose sole charge is to take messy, occasionally-absent utility power and render it clean, always-present power at the server. Every component below exists to strike one single point of failure from that chain.
The Road the Current Travels
- Grid to building. [/ˈgrid ˈto ˈbuildɪŋ/] n - Power arrives at the Where the Grid Comes In, usually at medium voltage, and is stepped down and protected by Medium-voltage switchgear.
- Bridging outages. [/ˈbridgɪŋ ˈoutages/] n - The Uninterruptible power supply carries the load instantly when utility power dips or fails, drawing on Battery energy storage for the seconds-to-minutes gap.
- Sustained backup. [/ˈsustainɪd ˈbakkup/] n - A The Standby Engine starts and, by way of an Automatic transfer switch, takes up the building load for as long as the outage endures.
- Distribution. [/ˈdistribuʃən/] n - Conditioned power flows through the Power distribution unit, optional Remote power panels, and Busway distribution out to each row.
- At the rack. [/ˈat ˈthe ˈrakk/] n - A Rack power distribution unit delivers power to individual servers, oft fed from two independent paths through a Static transfer switch.
Of the Notions That Bind It
- The whole is built around The Current That Comes in Threes and careful Grounding and bonding.
- How many of each component one needs is a question of Redundancy topologies (N+1, 2N, …).
- The resulting design is graded by the Uptime Institute tier classification.
- The efficiency of the whole plant is summed up in Power usage effectiveness.
Whither the Chain Tends
The AI build-out strains every link above. The shifts worth the watching are gathered in The New Buildout: Direct-to-chip liquid cooling, 800-volt DC distribution, On-site power generation, and the carrying of the network over to light by Silicon photonics. The most distant shape of that same pressure is the question of whether computing must one day quit the planet; see Data Centers on the Moon for what it would ask of us.
For how the largest operators have pressed efficiency, see Efficient datacenter case studies.
Of the Work as It Is Lived
For the messier truth, the small troubles that yielded to a clever fix and the great ones that did not, see Field Notes and Problems We Could Only Work Around.
For why these pages are ordered as they are, see About these pages.