Glossary
49 terms you will meet reading about AI data centers, in plain language. Where a term is a physical part, it links to the page listing the companies that make it. For how the parts connect, see how it fits together and the explainers.
#
- 800 V DC
- A rack power architecture that distributes direct current at about 800 volts instead of converting AC to DC inside every server. It removes conversion stages between the building and the chip and shrinks the copper needed per rack. Who makes it: 800 V DC architecture & solid-state transformers →
A
- ABF substrate · also Ajinomoto Build-up Film
- The high-layer-count package substrate, built on Ajinomoto Build-up Film, that carries a chip and routes it to the board. Large AI packages need very large, very fine-pitch versions, and supply has been tight. Who makes it: ABF / high-layer package substrates →
- AEC (active electrical cable)
- A copper cable with signal-conditioning chips in its connectors so it can run farther at high speed than a passive cable (DAC). It sits between passive copper and optics in reach, cost and power. Who makes it: Copper: DAC / ACC / AEC cables →
- ASIC (custom accelerator)
- A chip designed for one customer's workload rather than sold as a general product. Hyperscalers design their own AI accelerators, usually with a design-services partner, and have them built at a foundry. Who makes it: Hyperscaler custom ASICs →
B
- Behind-the-meter
- Generation located on the customer's side of the utility meter, such as on-site gas turbines or fuel cells. It lets a data center start before a grid connection is available. Who makes it: Gas turbines (onsite / behind-the-meter) & HRSGs →
- BESS · also battery energy storage system
- Battery energy storage system. At a data center it can ride through short outages, smooth the sharp power swings of AI training, or shift when the site draws from the grid. Who makes it: Energy storage: UPS batteries & BESS →
- Busway · also busbar trunking
- Enclosed metal bars that carry heavy current along an overhead or floor-level run, with plug-in points for taps. It replaces cable bundles for feeding rows of racks. Who makes it: Busway / busbar trunking →
C
- CDU (coolant distribution unit)
- The unit that separates the liquid loop touching the servers from the building's water loop. It exchanges heat between the two, pumps and filters the server-side coolant, and controls its temperature and pressure. Who makes it: Coolant distribution units →
- Chiller
- A refrigeration machine that cools water for the facility loop. Liquid-cooled AI halls can run warmer water and lean on chillers less than air-cooled halls do. Who makes it: Chillers & compressors →
- Co-packaged optics (CPO) · also CPO
- Optical engines placed inside or right beside the switch or accelerator package instead of in a pluggable module on the front panel. It shortens the electrical path, which cuts power per bit. Who makes it: Silicon photonics & co-packaged optics →
- Cold plate · also direct-to-chip
- A metal plate with internal channels that sits on a chip, with liquid flowing through it to carry heat away. It is the core part of direct-to-chip liquid cooling. Who makes it: Direct-to-chip cold plates →
- CoWoS · also chip-on-wafer-on-substrate
- TSMC's family of 2.5D advanced packaging, which places a processor and stacks of HBM side by side on a silicon interposer. Capacity at the packaging step has been one of the limits on AI accelerator output. Who makes it: 2.5D/3D advanced packaging →
- CRAH / CRAC
- Computer-room air handler (chilled-water coil) and computer-room air conditioner (built-in compressor). The air-side cooling units of a traditional data hall. Who makes it: CRAH / CRAC / fan walls / in-row units →
- CXL (Compute Express Link)
- A cache-coherent interconnect that runs over PCIe wiring. It lets processors share and expand memory beyond what is soldered or slotted next to them. Who makes it: CXL controllers & memory expanders →
D
- DAC (direct-attach copper)
- A passive copper cable with fixed connectors at each end. It is the cheapest and lowest-power link, limited to a few meters at current speeds. Who makes it: Copper: DAC / ACC / AEC cables →
- DCI (data center interconnect)
- Long-haul links, often coherent optics, that join separate buildings or campuses into one network so a training job can span them. Who makes it: DCI / coherent optics →
- DPU
- Data processing unit: a network card with its own processor cores, used to offload networking, storage and security work from the host CPU. Who makes it: DPUs →
- DRUPS
- Diesel rotary uninterruptible power supply. A flywheel and motor-generator set that bridges power gaps mechanically instead of with batteries, with a diesel engine behind it. Who makes it: UPS (static, modular, rotary/DRUPS) & UPS power semis →
- Dry cooler
- A radiator-like heat exchanger that rejects heat to outside air with fans and no water consumption. Adiabatic versions add water spray on hot days. Who makes it: Dry coolers, adiabatic/evaporative units, cooling towers, fans & motors →
E
- EMS (electronics manufacturing services)
- Contract manufacturers that assemble boards, servers and whole racks for system vendors and hyperscalers, and increasingly integrate and test full racks. Who makes it: EMS, system integration & rack-level test →
G
- Grain-oriented electrical steel · also GOES
- Silicon steel whose crystal grains are aligned to carry magnetic flux efficiently. It is the core material of power transformers, and few mills make it. Who makes it: Grain-oriented electrical steel & amorphous cores →
H
- HBM (high-bandwidth memory)
- Stacks of DRAM dies connected through the silicon with very wide interfaces and packaged next to the accelerator. It supplies the memory bandwidth AI chips need and is made by only a few companies. Who makes it: HBM →
- Heat reuse
- Sending the heat a data center rejects to a district-heating network or a nearby user instead of releasing it to the air. Warmer liquid-cooling return temperatures make it more practical. Who makes it: Heat reuse & district heating interfaces →
- Hyperscaler
- One of the very large cloud operators, such as Amazon, Microsoft, Google or Meta, whose purchases set demand for most of the parts in this catalog. Who makes it: Hyperscalers, neoclouds & model labs →
I
- Immersion cooling
- Submerging servers in a tank of dielectric fluid, either single-phase (the fluid stays liquid) or two-phase (it boils and condenses). Who makes it: Immersion cooling systems →
- Interconnection queue
- The waiting list of generation and large-load projects a grid operator must study before connecting. For new data centers, waiting time here often sets the schedule more than construction does. Who makes it: Transmission, interconnection & grid equipment →
- Interposer
- A thin silicon layer under a processor and its HBM that carries the thousands of short, dense wires between them. Who makes it: Interposers, silicon bridges, RDL →
L
- Liquid cooling
- Moving server heat with a liquid loop (cold plates, manifolds, CDUs) instead of air. It is needed once rack power passes what air can remove. Who makes it: Direct-to-chip cold plates →
M
- Manifold
- The pipe assembly in a rack that splits the coolant feed into branches for each server and gathers the return. Who makes it: Rack manifolds →
- MV/LV transformer
- A step-down transformer that takes medium-voltage distribution to the low voltage used inside a building. Who makes it: MV/LV transformers →
N
- Neocloud
- A cloud provider built around renting GPU capacity, such as CoreWeave, as distinct from the general-purpose hyperscalers. Who makes it: Hyperscalers, neoclouds & model labs →
- NIC / SuperNIC
- The network interface card that connects a server to the network. AI-specific versions are tuned for the high-bandwidth, low-latency traffic between GPUs. Who makes it: NICs / SuperNICs →
- NVLink · also scale-up
- Nvidia's high-bandwidth link between GPUs within a rack, the best-known example of a scale-up fabric. Who makes it: Scale-up fabric →
O
- Optical transceiver
- A pluggable module that converts electrical signals to light and back, so a switch or server can talk over fiber. Sold at 400G, 800G and 1.6T speeds. Who makes it: Optical transceivers →
P
- PDU (power distribution unit)
- A box that splits power from a feed into outlets or branch circuits. Rack PDUs sit in the rack; floor PDUs and RPPs feed rows. Who makes it: Rack PDUs →
- Power shelf
- A chassis of power supply units in a rack that converts incoming power to the DC bus feeding the servers. Who makes it: Power shelves & PSUs →
- PPA (power purchase agreement)
- A long-term contract to buy electricity from a specific plant at agreed terms. Data center operators sign them to secure new or existing generation. Who makes it: Utilities, IPPs & power providers signing data-center deals →
- PUE
- Power usage effectiveness: total facility energy divided by the energy used by the IT equipment. A value of 1.0 would mean no overhead for cooling and power conversion.
Q
- Quick disconnect (QD)
- A coupling that joins liquid lines without spilling, so a server can be pulled from a liquid-cooled rack without draining it. Who makes it: Quick disconnects →
R
- Retimer
- A chip that cleans up and re-transmits a high-speed signal so it can travel farther across a board or cable. Who makes it: PCIe/CXL switches & retimers →
S
- Scale-out network
- The network between racks, usually Ethernet or InfiniBand over optics, that ties thousands of accelerators into one cluster. Who makes it: Scale-out switch ASICs →
- Scale-up fabric
- The very-high-bandwidth network inside a rack that makes many accelerators behave like one, currently carried mostly over copper. Who makes it: Scale-up fabric →
- Silicon photonics
- Building optical components on silicon wafers with chip-making processes, which lowers the cost of optical transceivers and enables co-packaged optics. Who makes it: Silicon photonics & co-packaged optics →
- SMR (small modular reactor)
- A nuclear reactor designed to be built in factory-made modules, with a smaller output than conventional plants. Several data center operators have signed agreements to buy their power. Who makes it: SMRs, advanced reactors, components, fuel →
- Solid-state transformer · also SST
- A power-electronics replacement for a conventional transformer that converts voltage and AC/DC without bulky iron cores. Central to 800 V DC designs. Who makes it: 800 V DC architecture & solid-state transformers →
- Switchgear
- Breakers, switches and protection that control and isolate electrical circuits. It comes in high-voltage, medium-voltage and low-voltage forms along the power chain. Who makes it: MV switchgear, E-houses, power skids →
T
- TIM (thermal interface material)
- The paste, pad, or metal layer between a chip and its lid or cold plate that conducts heat across the gap. Who makes it: Chip-level thermal: lids, heat spreaders, TIMs →
U
- UPS
- Uninterruptible power supply. Batteries or a flywheel and power electronics that carry the load through a grid disturbance until generators start. Who makes it: UPS (static, modular, rotary/DRUPS) & UPS power semis →
V
- VRM (voltage regulator module) · also power stage
- The board-level power stage that steps the rack's DC bus down to the roughly one volt a GPU core runs at, at very high current. Who makes it: Board-level power: VRMs, power stages, PoL, power semis →