AI data center glossary: PUE, UPS, BESS and more

30 terms used in planning AI data centers, from PUE and UPS to interconnection queues and radiators in orbit, in plain words.

Average power use: Average power drawn as a share of the maximum, including idle time. Energy and cost scale with it.
Backup battery (BESS): A battery energy storage system that carries the critical load for minutes during a power cut, until generators start.
Backup generator: An engine-driven generator that starts after a power cut and can run the facility for hours or days.
Beta angle: The angle between an orbit’s plane and the direction of the Sun. At 0° a satellite spends the longest time in the Earth’s shadow; above a limit set by altitude it never enters it.
Cooling distribution unit (CDU): Pumps and a heat exchanger that move coolant to the chips and hand the heat to the building’s water loop.
COP (coefficient of performance): How many kilowatts of heat a cooling system removes per kilowatt of electricity it uses. Higher is better.
Critical load: The equipment that must never lose power: the IT itself and the liquid cooling that keeps it from overheating.
Critical path: The equipment whose delivery decides the earliest possible opening date.
Dawn-dusk sun-synchronous orbit: A near-polar low orbit that rides the line between day and night, so its solar arrays face the Sun almost all the time and eclipses are rare and short.
Direct-to-chip cooling: Liquid flows through cold plates on the hottest chips; fans and air handle the remaining heat.
Free cooling (economizer): Using cool outside air or water instead of running chillers. It saves energy in cool weather.
Immersion cooling: Servers sit in tanks of non-conductive liquid that absorbs almost all of their heat.
Interconnection queue: The line of projects waiting for the grid operator to study and approve a new connection. Waits are often several years.
IT load: The power drawn by the computing equipment itself: GPUs, CPUs, memory, networking and storage.
Lead time: How long equipment takes to arrive after you order it.
MVA (megavolt-ampere): The unit for the size of an electrical connection or transformer. Roughly a megawatt, adjusted for the power factor.
Non-blocking fabric (fat tree): A network of leaf and spine switches where every GPU can send at full speed to any other at once, because each switch has as many links up as down.
Power conversion system (PCS): The electronics that connect a battery to the facility’s AC power.
Power distribution unit (PDU): The power strips inside a rack that feed each server.
Power transformer: Steps grid voltage down to what the facility uses. Large ones are custom-built and often take years to deliver.
PUE (power usage effectiveness): Total facility energy divided by IT energy. 1.0 would mean every kilowatt-hour goes to computing; typical values are 1.1 to 1.6.
Rack density: Power drawn by one rack, in kilowatts. It decides which cooling methods can keep up.
Radiation design margin: How many times the expected mission dose a part can take before it fails. Spacecraft are usually built to a margin of at least 2.
Redundancy (N, N+1, 2N): N means just enough equipment; N+1 adds one spare unit; 2N duplicates everything so either half can carry the load.
Single-event effects: Glitches caused when one energetic particle strikes a chip: a flipped bit, a silent wrong answer, a memory error ECC cannot fix, or a processor that hangs until reset.
Switchgear: Large switches and breakers that route and protect high-voltage power inside the site.
Tokens: The pieces of text an AI model reads and writes: a word or part of a word.
Total ionizing dose: Radiation energy absorbed by a chip over its life, in rad(Si). It slowly shifts transistor behaviour until parts fail; shielding and tolerant parts keep it in check.
UPS (uninterruptible power supply): Keeps power steady and bridges the gap between a power cut and backup power taking over.
WUE (water usage effectiveness): Litres of water used per kilowatt-hour of IT energy, mostly for evaporative cooling.