Glossary

The vocabulary of the build.

800 V DCalso HVDC, ±400 V: A proposed rack power architecture that rectifies once at the facility edge and distributes high-voltage DC to racks, cutting copper and conversion losses.
Approach temperature: How close a heat exchanger or cooling tower gets its output to the limiting temperature (for example tower outlet minus wet-bulb).
ATSalso Automatic transfer switch: Switches the load from utility to generator when utility power fails, and back.
BESS: Battery energy storage system, typically containerized lithium iron phosphate, used for ride-through, smoothing and grid services.
Build-to-suit: A developer builds a facility to a tenant's specification under a long-term lease.
Buswayalso Bus duct: Overhead enclosed conductors with plug-in tap-off boxes, used instead of individual cables to feed racks.
Capacity factor: Average load divided by peak capacity over a period.
CDUalso Coolant distribution unit: Pumps, heat exchanger and controls that separate the rack coolant loop from facility water.
Co-packaged opticsalso CPO: Optical engines placed on the same package as the switch ASIC to cut electrical trace loss and power.
Colocation: Renting space, power, cooling and connectivity in someone else's data center.
Commissioning: Systematic testing, up to integrated failover tests, that proves the facility works as designed before go-live.
Concurrently maintainable: Any component can be taken out for maintenance without shutting down the IT load (Tier III).
COPalso Coefficient of performance: Heat moved per unit of electricity used. A COP of 6 moves 6 kW of heat with 1 kW of power.
CoWoSalso Chip on wafer on substrate: An advanced packaging technology placing logic dies and HBM on a silicon interposer; a key capacity bottleneck for AI chips.
CRAC / CRAH: Computer room air conditioner (own compressor) / computer room air handler (uses chilled water).
Cross-connect: A physical cable patch between two parties' equipment in a data center.
Cycles of concentration: How many times minerals concentrate in tower water before blowdown; higher cycles mean less water withdrawn.
DAC / AOC: Direct-attach copper cable / active optical cable. DAC reach shrinks as speeds rise.
Dark fiber: Installed but unlit fiber leased and lit with the customer's own equipment.
DCIM / BMS / EPMS: Software for assets and capacity, building mechanical control, and electrical power monitoring.
Delta-T (ΔT): Temperature rise of a fluid across equipment. Heat = mass flow × specific heat × ΔT.
Direct-to-chipalso D2C, DLC, cold plate: Liquid cooling through cold plates mounted on processors.
Distributed redundancyalso Block redundant, Catcher: Several power blocks share reserve capacity (for example 4 blocks sized to carry the load of 3), cheaper than 2N with similar resilience.
Economizeralso Free cooling: Using outside air or water when conditions allow, avoiding compressor energy.
Embodied carbon: Emissions from manufacturing materials and equipment, as opposed to operational emissions from electricity and fuel.
FP4 / FP8 / BF16: Number formats. Fewer bits mean more operations per watt; training and inference choose formats by accuracy needs.
Front-of-meter / behind-the-meteralso FOM, BTM: Front-of-meter power comes from the grid through a utility connection; behind-the-meter power is generated on site.
GPU-hour: One GPU rented for one hour; the unit neoclouds sell.
HBM: High-bandwidth memory: stacked DRAM placed beside the GPU on an interposer.
Hot aisle / cold aisle: Rows arranged so racks draw cold air from one aisle and exhaust into the next; containment seals one of them.
Hyperscaler: A company operating cloud infrastructure at very large global scale with its own campuses, hardware designs and networks.
Immersion cooling: Submerging servers in dielectric fluid, single-phase or two-phase.
InfiniBand: A low-latency fabric with native RDMA, long dominant in HPC and AI clusters.
Interconnection queue: The line of generation and large-load projects waiting for grid studies and approval; a major source of schedule delay.
kW/rackalso Rack density: Power drawn by one rack. Enterprise rooms ran ≈1–5 kW, cloud ≈5–15 kW, liquid-cooled AI racks well above 100 kW.
Lead time: Time from order to delivery. Transformers, switchgear, turbines and accelerators have been the longest in recent build waves.
Leaf-spinealso Clos: Two-tier fabric where every leaf connects to every spine, giving uniform latency.
LPO / XPO: Linear pluggable optics drop the DSP; extended pluggable optics consolidate many lanes in one larger, cold-plated module.
Mantrap: An interlocked entry that admits one authenticated person at a time.
Medium voltage (MV): Roughly 1–35 kV in data-center practice: the campus distribution level between the substation and the buildings.
Meet-me roomalso MMR: The room where carriers and customers interconnect via cross-connects.
MFUalso Model FLOP utilization: Share of theoretical peak FLOPs a training run actually achieves; networking and memory stalls keep it well below 100%.
N, N+1, N+2, 2Nalso Redundancy: N is the capacity needed. N+1 adds one spare unit, N+2 two. 2N duplicates the entire system so either half can carry the full load.
Neocloud: An AI-focused cloud or GPU cloud operator built around accelerated computing rather than general-purpose cloud services, emphasizing speed, utilization, dense clusters and liquid cooling.
North-south / east-west: North-south traffic enters or leaves the data center; east-west traffic moves between servers inside it. AI training is dominated by east-west.
NVLink: NVIDIA's GPU-to-GPU interconnect; NVLink 5 provides 1.8 TB/s per GPU.
Oversubscription: Ratio of downstream to upstream bandwidth at a switch tier. 3:1 is common in cloud; AI backends aim near 1:1.
PDUalso Power distribution unit: Either a floor-standing unit that distributes power to rows, or the power strip / power shelf inside a rack.
Powered shell: A building with utility power connected, ready for tenant fit-out.
Pre-action sprinkler: Sprinkler pipes stay dry until detection confirms a fire, reducing accidental water release.
PUEalso Power usage effectiveness: Total facility energy divided by IT equipment energy. 1.0 would mean zero overhead. Useful, but it ignores water, embodied carbon, grid impacts and how efficiently the IT itself does useful work.
Rack-scale systemalso NVL72: A rack designed and shipped as one system: compute trays, switch trays, power shelves and liquid manifolds integrated.
RDMA: Remote direct memory access: one machine writes into another's memory without involving either CPU.
Rear-door heat exchangeralso RDHx: A liquid coil on a rack's rear door that cools exhaust air.
Reticle limit: The largest area a lithography scanner can pattern in one exposure, roughly 858 mm²; very large GPUs join multiple reticle-sized dies.
RoCEalso RDMA over Converged Ethernet: RDMA carried over Ethernet; the basis for most Ethernet AI fabrics.
Scale-up / scale-out: Scale-up joins GPUs inside a rack or node (NVLink-class); scale-out joins racks over the backend fabric.
Scope 1 / 2 / 3: Greenhouse-gas accounting categories: direct emissions (e.g., generators), purchased electricity, and supply-chain emissions.
Solid-state transformeralso SST: A transformer built from power electronics (SiC or GaN) rather than only magnetic cores; smaller and controllable, still early for data-center use.
Stranded capacity: Power, cooling or space that is built and paid for but cannot be used, for example because racks are denser than the room was designed for.
STSalso Static transfer switch: Solid-state switch that moves load between two sources in a few milliseconds.
Supercapacitor: Energy storage that charges and discharges in milliseconds, used to smooth fast GPU load swings alongside UPS and BESS.
Tensor core: A GPU unit specialized for matrix multiply-accumulate at low precision.
Time-sharing: 1960s technique for sharing one computer among many users by slicing processor time; the origin of usernames, accounts and the computer-utility idea.
Ton of refrigeration: ≈3.517 kW of heat removal; a legacy unit still common for chillers.
Ultra Ethernetalso UEC: An industry consortium (formed 2023) defining Ethernet transport for AI and HPC; 1.0 specification released in 2025.
UPS: Uninterruptible power supply: rectifier, batteries and inverter that carry IT load through outages until generators start.
Uptime Tier I–IValso Tier 3, Tier 4: Uptime Institute's classification: Tier I basic, II redundant components, III concurrently maintainable, IV fault tolerant. Published availability targets range from about 99.671% (Tier I) to 99.995% (Tier IV).
VESDA: Very early smoke detection apparatus: aspirating detection that samples air continuously.
Virtualization: Making physical resources appear as logical ones (virtual machines, virtual drives, virtual networks). The cloud itself is virtualization at planetary scale.
Water consumption: Water not returned to its source, mostly evaporated in cooling towers.
Water withdrawal: Water taken from a source. Some is returned (for example as blowdown); withdrawal is always at least as large as consumption.
Wet-bulb temperature: The lowest temperature evaporation can reach; sets the limit for cooling towers.
WUEalso Water usage effectiveness: Site water use in liters divided by IT energy in kWh. Depends heavily on climate and cooling design; it does not capture water used to generate the electricity.