🖥️ AI Data Center Terminology Glossary: 55 Essential Terms Explained

Why These Terms Matter

AI data centers are a distinct engineering discipline from traditional enterprise data centers. GPU training and inference clusters draw far more power per rack, reject far more heat per square foot, and depend on liquid cooling technologies that were exotic a decade ago and are now mainstream. Engineers moving into this space — from electrical, mechanical, or controls backgrounds — encounter a dense vocabulary of acronyms and terms borrowed from HVAC, power distribution, networking, and computer hardware, often used in ways that are specific to this application.

Getting these terms right matters because they show up in design basis documents, utility interconnection applications, mechanical schedules, and commissioning checklists. A term like "PUE" or "N+1" carries a precise technical meaning that affects capacity planning, cost estimates, and code compliance. This glossary covers 55 of the most important terms, organized alphabetically, with references to the standards bodies — ASHRAE, IEEE, NFPA, and the Uptime Institute — that define or govern them.

A

Air-Cooled Rack — Traditional cooling method

A server rack cooled entirely by forced air moving across heatsinks, typically supplied by a raised-floor CRAC/CRAH system or hot/cold aisle containment. Air cooling is generally practical up to about 20–30 kW per rack; above that, the volume of air required to remove heat becomes impractical to move and dehumidify economically, which is why high-density AI racks shift to liquid cooling.

AI Factory — Industry term

A colloquial term (popularized by GPU vendors) for a data center purpose-built to run large-scale AI training or inference workloads, emphasizing that the facility's output is computation itself rather than general enterprise IT services. AI factories are typically designed around specific GPU/accelerator platforms from the ground up, including power and cooling infrastructure sized to that hardware's exact thermal and electrical profile.

Ambient Temperature — ASHRAE TC 9.9

The dry-bulb temperature of the air entering IT equipment intakes. ASHRAE's Thermal Guidelines for Data Processing Environments define allowable and recommended ambient ranges (Classes A1–A4) that IT equipment manufacturers design their hardware to tolerate, balancing reliability against economizer (free cooling) hours.

Availability — Reliability engineering term

The percentage of time a system is operational and able to perform its function, typically expressed in "nines" (e.g., 99.99% = "four nines," about 52 minutes of downtime per year). Data center availability targets drive redundancy level (N, N+1, 2N) decisions, since each additional redundant system reduces the probability of an outage from a single component failure.

B

Busway (Busbar Trunking) — Electrical distribution term

A prefabricated metal enclosure containing busbars used to distribute power overhead through a data hall instead of conduit and cable. Busway is favored in hyperscale and AI data centers because it allows plug-in tap boxes to be added or relocated quickly as rack layouts and power densities change — a common occurrence as GPU generations turn over every 1–2 years.

Battery Energy Storage System (BESS) — IEEE 1547 / NFPA 855

A stationary battery installation, often lithium-ion, used in data centers for UPS ride-through, peak shaving, or grid services. AI campuses increasingly pair BESS with on-site generation to smooth the highly volatile, millisecond-scale power swings that GPU training clusters exhibit as workloads synchronize and desynchronize across thousands of chips.

C

Coolant Distribution Unit (CDU) — Liquid cooling term

A unit that separates the facility water loop from the technology cooling loop that runs directly to server components, using a heat exchanger, pumps, and controls. The CDU isolates the sensitive, filtered coolant in direct contact with electronics from the less-controlled facility water system, and regulates supply temperature and flow to the racks it serves. CDUs may be rack-mounted (in-row) or larger centralized units serving many racks.

Cold Plate — Direct-to-chip cooling component

A metal plate with internal microchannels mounted directly on a GPU, CPU, or memory module, through which coolant flows to conduct heat away from the die. Cold plates are the core hardware of direct-to-chip liquid cooling and can remove several hundred watts to over a kilowatt per chip, far more than a heatsink cooled by air alone.

CRAC / CRAH — Computer Room Air Conditioner / Air Handler

CRAC units contain their own direct-expansion refrigeration compressor; CRAH units instead use chilled water supplied from a central chiller plant to cool air passing through a coil. Both distribute cooled air into a data hall, typically through a raised floor or overhead ducting, and are the traditional workhorse of air-cooled data centers before liquid cooling became necessary for AI racks.

Chilled Water Plant — HVAC term

The central mechanical system — chillers, cooling towers or dry coolers, pumps, and piping — that produces chilled water distributed throughout a data center to CRAHs, CDUs, or rear-door heat exchangers. In AI data centers the chilled water plant is often sized and piped for two loops: a warmer "facility water" loop for economizer-friendly liquid cooling and a colder loop reserved for any remaining air-cooled equipment.

Colocation (Colo) — Business model term

A data center model where a facility operator leases rack space, power, and cooling to multiple independent tenant customers within a shared building. AI colocation increasingly requires operators to retrofit facilities designed for 5–10 kW/rack up to 40–100+ kW/rack to accommodate GPU tenants, which is a major driver of liquid cooling retrofits industry-wide.

D

Direct-to-Chip Cooling (D2C) — Liquid cooling category

A liquid cooling method where coolant is piped directly to cold plates mounted on the hottest components (GPUs, CPUs), while lower-heat components like memory and power supplies may still be air-cooled by fans. D2C is the most common liquid cooling approach for current-generation AI accelerators because it targets cooling capacity precisely where the heat is generated without submerging the entire server.

Demand Response — Grid interconnection term

A utility program or contractual arrangement under which a large power consumer, such as an AI data center, agrees to reduce or shift its electrical load during periods of grid stress in exchange for financial incentives or as a condition of interconnection. Demand response participation is increasingly used to justify faster interconnection approval for gigawatt-scale AI campuses in constrained grid regions.

Dry Cooler — Heat rejection equipment

A heat rejection device that transfers heat from a closed water/glycol loop to outdoor air using finned coils and fans, without evaporating water (unlike a cooling tower). Dry coolers are favored in water-constrained regions for AI data centers because they eliminate the ongoing water consumption of evaporative cooling towers, at the cost of somewhat lower efficiency in hot climates.

E

Economizer (Free Cooling) — ASHRAE term

A mode of operation where outside air or water temperature is cool enough to reject IT heat without running mechanical refrigeration (compressors), reducing energy use. Air-side economizers bring in filtered outside air directly; water-side economizers use a heat exchanger between an outdoor loop and the chilled water loop. Facility water loops for liquid-cooled AI racks are typically designed with warmer setpoints specifically to maximize economizer hours.

Effective Power (or IT Load) — Capacity planning term

The actual electrical power delivered to and consumed by IT equipment (servers, GPUs, storage, networking), as distinct from the total facility power that also includes cooling, lighting, and electrical losses. IT load is the denominator against which cooling capacity and PUE are calculated.

F

Facility Water System (FWS) — Liquid cooling term

The building-side water loop that supplies and returns water (often at a warmer, economizer-friendly temperature such as 25–32°C) to CDUs throughout a data center. The FWS is distinct from the technology cooling system (TCS) loop that runs at tighter temperature and purity tolerances directly to cold plates — the CDU is the boundary between the two.

Fault Ride-Through — IEEE 1547 / grid code term

The capability of on-site generation or large loads to remain connected and stable through a brief grid voltage sag or frequency disturbance rather than immediately disconnecting. Utilities increasingly require fault ride-through capability from gigawatt-scale AI campuses because an abrupt disconnection or reconnection of hundreds of megawatts of load can itself destabilize the grid.

G

GPU Pod (or SuperPOD) — Cluster architecture term

A tightly coupled group of GPU servers, typically dozens to hundreds of nodes, interconnected with a high-bandwidth, low-latency fabric (such as NVLink or InfiniBand) to function as a single large compute resource for training a model. GPU pod topology drives rack layout, cable length limits, and cooling loop segmentation in the data hall design.

Grid Interconnection Queue — Utility planning term

The ordered backlog of generation and large-load interconnection requests a utility or grid operator must study and approve before a project can connect. Multi-hundred-megawatt AI data center campuses now frequently face multi-year interconnection queue waits, making queue position and utility relationship management a critical project schedule driver, alongside on-site or behind-the-meter generation as a workaround.

H

Hot Aisle / Cold Aisle Containment — Airflow management term

A data hall layout where racks are arranged in rows facing each other so that cold supply air enters from one aisle (cold aisle) and heated exhaust air is captured in the opposite aisle (hot aisle) and returned to cooling equipment without mixing. Containment (physical barriers sealing the aisle) significantly improves cooling efficiency by preventing hot and cold air from mixing, and remains important even in liquid-cooled halls for the residual air-cooled components.

Hyperscale Data Center — Industry classification

A data center operated at massive scale, typically defined as 5,000+ servers or 10,000+ square meters, usually built and operated by a single large cloud or AI company for its own or a small number of large customers' workloads. Hyperscale AI campuses are increasingly measured in hundreds of megawatts to gigawatts of critical IT load.

I

Immersion Cooling — Liquid cooling category

A cooling method where entire servers (or components) are submerged directly in a dielectric (electrically non-conductive) fluid that absorbs heat, either through natural convection (single-phase immersion) or by boiling and condensing the fluid (two-phase immersion). Immersion cooling can achieve very high rack densities but requires specialized tanks, fluid handling, and server designs, and is less common than direct-to-chip cooling in current AI deployments.

Interconnection Study — Utility grid term

A formal engineering analysis performed by a utility or grid operator to determine what grid upgrades (substations, transmission lines, transformers) are required to safely connect a proposed large load or generator, and who pays for them. AI data center developers now routinely commission or co-fund these studies to accelerate what would otherwise be a multi-year utility-led process.

Inference — AI workload term

The process of running a trained AI model on new input data to produce an output (as opposed to training, which adjusts the model's parameters). Inference workloads generally have lower, steadier power draw per GPU than training but far higher deployment volume, and are increasingly distributed to edge and regional data centers to reduce latency to end users.

K

kW per Rack (Rack Power Density) — Capacity planning metric

The average or peak electrical power, in kilowatts, delivered to a single server rack. Legacy enterprise data centers typically design for 5–10 kW/rack; current-generation GPU training racks commonly require 40–130 kW/rack or more, which is the single biggest driver forcing the transition from air to liquid cooling in new AI data center construction.

L

Liquid-to-Air (L2A) Cooling — Hybrid cooling term

A cooling architecture where heat is removed from chips by a liquid loop (via cold plates) but ultimately rejected to room air through a rear-door or in-row heat exchanger, rather than to a building chilled water plant. L2A systems allow liquid-cooled racks to be deployed inside existing air-cooled data halls without new facility water piping, which is common in colocation retrofits.

Liquid-to-Liquid (L2L) Cooling — Facility cooling term

A cooling architecture where the technology cooling loop transfers heat, through a CDU heat exchanger, directly into the building's facility water loop, which then rejects heat outdoors via cooling towers or dry coolers. L2L systems are more efficient than L2A at scale and are the standard approach in purpose-built, greenfield AI data centers.

M

Megawatt (MW) Critical Load — Capacity metric

The total electrical power, in megawatts, delivered to IT equipment that must be protected by UPS and backup generation (as opposed to non-critical loads like office space or exterior lighting). Utility interconnection applications, generator sizing, and UPS module counts are all specified against critical load, not total site load.

Modular Data Center — Construction method

A data center built from prefabricated, factory-assembled units (power skids, cooling skids, or entire IT modules in shipping-container-like enclosures) that are transported to site and interconnected, rather than built entirely with on-site construction. Modular construction is widely used for AI campuses to compress schedule, since power and cooling skids can be fabricated in parallel with site civil work.

N

N, N+1, 2N, 2N+1 (Redundancy Notation) — Uptime Institute / reliability term

Shorthand describing how many redundant components exist relative to the minimum needed to run the load. "N" means only the minimum required capacity exists with no spare (no redundancy). "N+1" means one extra unit of capacity beyond the minimum is provided, so a single component failure does not cause an outage. "2N" means the entire system is fully duplicated (two complete, independent paths), and "2N+1" adds one further spare unit to each of the two paths. These notations apply to UPS modules, generators, cooling units, and electrical distribution paths, and are central to Uptime Institute Tier certification.

O

Overhead Busway — Electrical distribution term

See Busway. The overhead variant runs along the ceiling or a dedicated support grid above racks, which is the dominant layout in modern AI data halls because it keeps the raised or slab floor clear for structural loading of dense, liquid-cooled racks and their piping.

P

Power Distribution Unit (PDU) — Electrical distribution equipment

Equipment that receives power from an upstream source (busway or feeder) and distributes it, usually stepped down and metered, to individual racks or rack-level power strips (rack PDUs). Data-hall-level PDUs typically include transformers, panelboards, and monitoring; rack PDUs are the strip-style units mounted inside or beside each rack.

Power Usage Effectiveness (PUE) — The Green Grid metric

The ratio of total facility power consumption to IT equipment power consumption (Total Facility Power ÷ IT Equipment Power). A PUE of 1.0 would mean 100% of power goes to IT equipment with zero overhead for cooling, lighting, and losses — not physically achievable in practice. Well-designed hyperscale AI data centers with liquid cooling and economizer-heavy climates commonly report PUE in the 1.1–1.3 range; legacy air-cooled facilities are often 1.5–2.0 or higher.

Purdue Model — See kb-scada-glossary for the full industrial-controls definition; referenced here for the OT/IT network segmentation applied to data center building management systems.

Not typically core AI data center vocabulary but occasionally referenced where a facility's BMS/EPMS integrates with the same OT security segmentation model used in industrial controls, isolating building automation networks from the IT/AI compute network.

R

Rear-Door Heat Exchanger (RDHx) — Hybrid cooling equipment

A liquid-cooled heat exchanger mounted on the back of a server rack in place of a standard perforated rear door, through which chilled water or refrigerant circulates to absorb heat from air exiting the servers before it enters the room. RDHx is a common way to raise a data hall's supported rack density without fully replacing air-cooled IT equipment with direct-to-chip liquid cooling.

Redundancy (Tier Classification) — Uptime Institute Tiers I–IV

The Uptime Institute's four-tier system rates data center infrastructure resilience: Tier I (basic capacity, no redundancy), Tier II (Tier I plus redundant components), Tier III (concurrently maintainable — any component can be taken offline for maintenance without affecting IT operations), and Tier IV (fault tolerant — the facility can withstand an unplanned failure of any component with no impact). AI training campuses, where a training run interruption is costly but not always safety-critical, sometimes accept Tier III rather than Tier IV design to control capital cost, while inference-serving facilities supporting live customer traffic often specify Tier IV-equivalent resilience.

S

Stranded Power — Capacity planning term

Electrical capacity that has been procured or built at a site but cannot be used because cooling capacity, floor space, or network fabric constraints prevent racks from being fully populated. Stranded power is a common problem when legacy facilities designed for low-density air cooling are retrofitted for GPU racks — the electrical service may be adequate while the cooling plant becomes the limiting factor, or vice versa.

Substation (On-Site or Utility) — Electrical infrastructure term

A facility containing transformers and switchgear that steps utility transmission or sub-transmission voltage down to a level usable by the data center's electrical distribution system. Gigawatt-scale AI campuses increasingly build dedicated on-site substations, sometimes at transmission voltage (115 kV and above), rather than relying solely on a utility-owned distribution feed.

T

Technology Cooling System (TCS) — Liquid cooling term

The closed, tightly controlled coolant loop that runs from a CDU directly to the cold plates on IT hardware, carrying treated, filtered fluid at controlled temperature, flow, and water quality (conductivity, pH, particulate levels) to protect sensitive electronics from corrosion or fouling. TCS water quality specifications are typically far stricter than the facility water system that supplies the CDU.

Thermal Design Power (TDP) — Semiconductor specification

The maximum sustained heat output, in watts, that a chip's cooling system must be designed to dissipate under realistic peak workload conditions, as specified by the chip manufacturer. TDP for current AI accelerators commonly ranges from several hundred watts to over a kilowatt per chip, directly driving whether air cooling remains viable or liquid cooling is required.

Training (Model Training) — AI workload term

The compute-intensive process of adjusting a neural network's parameters using large datasets, typically run continuously across thousands of GPUs for days to months. Training workloads produce sustained, highly synchronized power draw across an entire cluster, creating large, fast power swings that stress both the electrical distribution system and the utility grid supplying the facility.

Two-Phase Cooling — Immersion/direct cooling term

A cooling method that relies on a fluid's phase change (liquid boiling to vapor at the heat source, then condensing back to liquid after releasing heat elsewhere) rather than only sensible heat transfer in a liquid that stays liquid throughout (single-phase). Two-phase systems can move more heat per unit of coolant volume than single-phase systems but require engineered fluids and sealed systems to manage vapor pressure.

U

Uninterruptible Power Supply (UPS) — Electrical protection equipment

Equipment that provides continuous, conditioned power to critical IT loads and bridges the gap between a utility power loss and backup generator startup, using batteries, flywheels, or other stored-energy technology. In AI data centers, UPS systems (or dedicated battery buffers) increasingly also serve a power-quality role, smoothing the rapid, large power transients caused by GPU workloads synchronizing across a training cluster.

Uptime Institute — Certification body

The organization that developed and administers the Tier Classification System (Tier I–IV) for data center infrastructure design, construction, and operational sustainability, widely referenced (though not a code or law) across the industry as the de facto standard for describing data center resilience.

W

Water Usage Effectiveness (WUE) — The Green Grid metric

The ratio of total facility water consumption (in liters) to IT equipment power consumption (in kilowatt-hours), analogous to PUE but for water rather than electricity. WUE has become a significant design driver as AI data centers scale in regions facing water scarcity, pushing operators toward dry coolers and closed-loop liquid cooling over evaporative cooling towers despite the efficiency tradeoff.

White Space — Data center planning term

The portion of a data center floor area dedicated to IT equipment (racks, aisles) as opposed to "gray space," which houses mechanical and electrical support equipment (chillers, UPS, switchgear). In high-density AI facilities the white-space-to-gray-space ratio often shifts toward more gray space than legacy data centers, because the mechanical and electrical plant needed to support extreme rack densities is proportionally larger.

Quick Reference Table — 55 AI Data Center Terms

Term	Category	Key Point
Air-Cooled Rack	Cooling	Practical up to ~20–30 kW/rack
AI Factory	Industry term	Facility purpose-built for AI compute
Ambient Temperature	ASHRAE TC 9.9	Dry-bulb temp at IT equipment intake
Availability	Reliability	Uptime % expressed in "nines"
Busway	Electrical	Overhead prefab power distribution
BESS	Electrical / NFPA 855	Battery storage for ride-through/peak shaving
CDU	Liquid cooling	Separates facility loop from TCS loop
Cold Plate	Liquid cooling	Mounted on chip; carries coolant
CRAC / CRAH	HVAC	DX vs chilled-water air cooling units
Chilled Water Plant	HVAC	Central chillers, towers, pumps
Colocation	Business model	Multi-tenant leased data center
Direct-to-Chip (D2C)	Liquid cooling	Coolant piped to cold plates on hottest parts
Demand Response	Grid interconnection	Load reduction for grid stress events
Dry Cooler	Heat rejection	No water evaporation; lower water use
Economizer	ASHRAE	Free cooling using outdoor conditions
Effective Power / IT Load	Capacity planning	Power delivered to IT equipment only
Facility Water System (FWS)	Liquid cooling	Building loop feeding CDUs
Fault Ride-Through	IEEE 1547 / grid code	Stays connected through grid disturbance
GPU Pod / SuperPOD	Cluster architecture	Tightly coupled GPU group with fast fabric
Grid Interconnection Queue	Utility planning	Backlog of connection requests
Hot/Cold Aisle Containment	Airflow management	Separates supply and exhaust air
Hyperscale Data Center	Classification	5,000+ servers, single large operator
Immersion Cooling	Liquid cooling	Servers submerged in dielectric fluid
Interconnection Study	Utility grid	Determines grid upgrades needed
Inference	AI workload	Running a trained model on new data
kW per Rack	Capacity metric	40–130+ kW/rack common for GPU racks
Liquid-to-Air (L2A)	Hybrid cooling	Heat rejected to room air, not facility loop
Liquid-to-Liquid (L2L)	Facility cooling	Heat passed to building water loop
MW Critical Load	Capacity metric	Power backed by UPS/generation
Modular Data Center	Construction	Factory-built power/cooling skids
N / N+1 / 2N / 2N+1	Uptime Institute	Redundancy notation for critical systems
Overhead Busway	Electrical	Keeps floor clear for liquid cooling piping
PDU	Electrical	Distributes and meters power to racks
PUE	The Green Grid	Total facility power ÷ IT power
Rear-Door Heat Exchanger	Hybrid cooling	Liquid-cooled rack rear door
Redundancy Tiers I–IV	Uptime Institute	Concurrent maintainability / fault tolerance
Stranded Power	Capacity planning	Capacity unusable due to other constraint
Substation	Electrical infrastructure	Steps utility voltage down on-site
TCS	Liquid cooling	Tightly controlled loop to cold plates
TDP	Semiconductor spec	Max sustained heat output of a chip
Training	AI workload	Synchronized power swings across cluster
Two-Phase Cooling	Liquid cooling	Uses boiling/condensing phase change
UPS	Electrical protection	Bridges utility loss to generator start
Uptime Institute	Certification body	Defines Tier I–IV classification
WUE	The Green Grid	Water use ÷ IT power consumption
White Space	Planning term	IT floor area vs. gray (M&E) space