Reference

Computing terms and definitions

Computing terms: AI and machine learning, retrieval and data systems, processors and memory, chip manufacturing and lithography, photonics, quantum computing, data centers, networks and industrial systems, and the rules that govern them. These are general reference definitions, not tied to any one report.

309 terms · Updated 21 Oct 2026

Filter terms309 of 309 terms shown

#

3D NAND (vertical NAND, V-NAND, layer count): NAND flash built by stacking hundreds of layers of memory cells vertically. Adding layers raises density without shrinking cells.
3GPP (3rd Generation Partnership Project)In Aerospace: The partnership of standards bodies that writes the technical specifications for mobile networks such as 4G LTE and 5G, issued in numbered releases.

A

Actinic inspection: Inspecting EUV masks with light of the same 13.5 nm wavelength used to print, which reveals defects other light cannot.
Advanced packaging (heterogeneous integration, 2.5D packaging, 3D packaging): Methods for connecting chips densely inside one package, such as interposers, 2.5D and 3D stacking and hybrid bonding. It has become a main lever for AI chip performance alongside transistor scaling.
AI accelerator (accelerator, custom silicon): A chip designed specifically to run AI workloads faster or more efficiently than general-purpose processors, such as GPUs, TPUs and NPUs.
AI agent (agentic AI, AI agents): An AI system that plans and carries out multi-step tasks by calling tools, software or other services, instead of only answering a single prompt.
Analog computing (analog compute-in-memory): Computing with continuously varying signals, such as currents or light intensities, instead of digital bits. It can be efficient for matrix math but is limited by noise and precision.
Anomaly detection: Automatically flagging data points or patterns that differ from normal, such as a leak, intrusion or failing machine.
API (application programming interface, APIs): Application programming interface: a defined way for one piece of software to request data or services from another. AI models and cloud services are sold largely through APIs.
Approximate nearest neighbor search (ANN, ANN search, approximate nearest-neighbor): Finding the stored vectors closest to a query quickly by accepting a small chance of missing the exact best match. It makes large-scale vector search practical.
ASIC (application-specific integrated circuit, ASICs): Application-specific integrated circuit, a chip designed for one task. It beats general processors on speed and energy for that task, as bitcoin miners and AI accelerators show, but cannot be repurposed.
ATEX (ATEX directive): The EU rules for equipment and workplaces in explosive atmospheres. Sensors used in such areas must be certified to them.

B

Back end of line (BEOL, front end of line, FEOL): The later chip-making steps that build the layers of metal wiring connecting transistors. Front end of line builds the transistors themselves.
Behind-the-meter and front-of-meter (behind-the-meter, front-of-the-meter, BTM, FTM, utility-scale storage)In Batteries: Behind-the-meter storage sits on a customer's side of the utility meter, such as home or factory batteries. Front-of-meter storage connects directly to the grid and sells services to it.
Bill of materials (BOM)In Batteries: The list and cost of every material and part in a product. In batteries the cathode, anode, electrolyte, separator and foils make up most of it.
Bitcoin mining (bitcoin miner, crypto mining): Using specialized computers to compete to add blocks to the bitcoin ledger, earning new coins and fees. Profit depends on bitcoin's price, electricity cost and the efficiency of the machines.
Bits per cell (SLC, MLC, TLC, QLC, penta-level cell): How many bits each flash memory cell stores. More bits per cell lower cost but reduce speed and endurance. See the table
BM25: A classic keyword search ranking formula that scores documents by how often query words appear, adjusted for document length. It remains a strong baseline beside vector search.
Brain-computer interface (BCI, brain-machine interface): A system that reads brain signals and turns them into commands for a computer or machine. Non-invasive types use scalp electrodes; invasive types implant electrodes in the brain.
Bureau of Industry and Security (BIS)In Aerospace: The US Commerce Department agency that administers the EAR and keeps lists of restricted foreign parties, such as the Entity List.

C

Capacity factorIn Batteries: A plant's actual output over a period divided by what it would produce running at full power the whole time. Solar typically reaches ≈10–25%; storage owners use it to describe utilization.
CBRS (Citizens Broadband Radio Service): Citizens Broadband Radio Service, a shared US band at 3.55–3.7 GHz that companies can use for private LTE and 5G networks.
Chemically amplified resist (CAR): A photoresist in which each absorbed photon releases an acid that triggers many chemical reactions, raising sensitivity. It is standard for deep ultraviolet and early EUV lithography.
Chiplet (chiplets): A small die designed to be combined with others in one package, instead of building one large chip. Chiplets improve yield and let different parts use different processes.
CHIPS and Science Act (CHIPS Act): The 2022 US law that funds domestic semiconductor manufacturing and research with ≈$52 billion and created the Section 48D investment tax credit.
Chunking (semantic chunking): Splitting documents into smaller passages before embedding them for retrieval. Chunk size and boundaries strongly affect retrieval quality.
Cloud computing: Renting computing, storage and software over the internet from providers instead of owning hardware.
CMMC (Cybersecurity Maturity Model Certification)In Aerospace: The US Defense Department program requiring contractors to meet, and prove, cybersecurity standards for the information they handle, at Levels 1 to 3. It began phasing into contracts on 10 November 2025.
CMOS (complementary metal-oxide-semiconductor): Complementary metal-oxide-semiconductor, the transistor technology used in almost all digital chips. Its low idle power made modern processors possible.
Co-location (behind-the-meter co-location): Placing a large load, such as a data center, directly at a power plant to draw power behind the meter. Regulators scrutinize it because it can shift costs to other grid users.
Co-packaged optics (CPO): Placing optical engines in the same package as a switch or processor chip, instead of in pluggable modules at the front panel. It cuts power and raises bandwidth density.
Coherent optics (coherent DSP): Optical links that encode data in both the amplitude and phase of light, recovered with a digital signal processor. They carry the most data over long distances.
Colocation (colo): Renting space, power and cooling in a data center for customer-owned servers.
Combinatorial optimization (combinatorial optimisation): Finding the best option among a huge but finite set of combinations, such as delivery routes, schedules or chip layouts. Many such problems are NP-hard, so solvers rely on heuristics.
Commerce Control List (CCL)In Aerospace: The EAR list of controlled dual-use items, each identified by an export control classification number that determines which destinations need a license.
Computational lithography (OPC, optical proximity correction, ILT, inverse lithography technology, SMO, source-mask optimization, curvilinear masks): Using simulation and optimization to adjust mask shapes and light sources so the printed pattern matches the design. It includes optical proximity correction, inverse lithography and source-mask optimization.
Computer vision (machine vision): AI that interprets images and video, for tasks such as detecting objects, reading text or inspecting parts.
Context window (context length, long context): The maximum amount of text, in tokens, a language model can consider at once, including the prompt and its answer. Longer windows cost more compute and models often use the middle poorly.
Corpus poisoning (index poisoning, data poisoning): Planting false or malicious documents in a retrieval collection so that an AI system retrieves and repeats them.
Cost of ownership (CoO, cost per wafer pass): The full cost of running a manufacturing tool over its life, including price, throughput, uptime, consumables and yield, usually expressed per wafer pass.
CPU (central processing unit): Central processing unit, the general-purpose processor that runs operating systems and most software. It handles varied tasks well but few operations in parallel.
Critical dimension (CD, CDU, critical dimension uniformity): The width of the smallest feature in a pattern that must be controlled tightly, such as a transistor gate. Critical dimension uniformity measures how consistent it is across the wafer.
Critical IT load (critical IT capacity, IT load, IT MW): The power used by a data center's computing equipment itself, excluding cooling and other overhead. Data center capacity is usually quoted in megawatts of IT load.
Cryogenic computing (cryogenic CMOS, cryo-CMOS): Running computer hardware at very low temperatures, from liquid nitrogen's 77 K to near absolute zero, to cut resistance and leakage. Cooling costs can cancel much of the gain.
Crypto-agility (cryptographic agility): Designing systems so their cryptographic algorithms can be swapped quickly, for example to post-quantum methods, without rebuilding the system.
Curtailment (curtailed electricity)In Batteries: Deliberately reducing output from wind or solar plants because the grid cannot use or carry the power. Curtailed electricity is cheap or free to store.

D

DARPA (Defense Advanced Research Projects Agency)In Aerospace: The US Defense Department agency that funds high-risk, high-payoff research, often through short, focused programs with industry and universities.
Data center (datacenter, data centre): A facility housing servers, storage and networking, with the power and cooling to run them. AI data centers draw hundreds of megawatts.
Data converter (DAC, ADC, digital-to-analog converter, analog-to-digital converter): A circuit that converts between analog and digital signals. Digital-to-analog and analog-to-digital converters can dominate the power of analog and photonic computers.
Data historian (process historian): Software that records time-stamped readings from plant sensors and controllers for analysis and compliance.
Data lake (data lakes): A large store that holds raw data in its original formats, from sensor logs to documents, for later analysis.
Data residency (data localization, data localisation): A requirement that data be stored and processed within a particular country or region, often for legal or security reasons.
Data sovereignty: The principle that data is subject to the laws of the country where it is stored or whose citizens it concerns. It drives demand for local clouds and on-premises AI.
Decoherence (coherence time, T1, T2): The loss of a qubit's quantum state through interaction with its surroundings. Coherence times, called T1 and T2, measure how long a qubit keeps its state.
Deep learning: Machine learning with neural networks of many layers, trained on large datasets. It drives modern speech, vision and language systems and needs large amounts of compute.
Deep ultraviolet lithography (DUV, ArF, KrF, ArF immersion, 193i, i-line): Lithography using 248 nm krypton fluoride or 193 nm argon fluoride lasers. Immersion versions place water between lens and wafer to sharpen resolution; with multipatterning they reach ≈7 nm-class nodes.
Demand response (demand-response, curtailment credits): Paying large electricity users to cut consumption when the grid is strained. Data centers and bitcoin miners earn credits by curtailing.
Dense and sparse retrieval (dense retrieval, sparse retrieval, hybrid retrieval, hybrid search): Dense retrieval matches queries to documents by embedding similarity; sparse retrieval matches on keywords, as BM25 does. Hybrid retrieval combines both.
Die yield (wafer yield, yield ramp): The share of chips on a wafer that work. It falls as chips get larger and as processes get newer, which is why chiplets help.
Diffractive neural network (D²NN, diffractive optical neural network): An optical system of patterned layers that diffract light so the output pattern computes a neural network's result as light passes through.
Digital Omnibus on AI (Regulation (EU) 2026/1744, AI omnibus): Regulation (EU) 2026/1744, in force from July 2026, which amends the EU AI Act. It postpones high-risk obligations to 2 December 2027 for Annex III systems and to 2 August 2028 for Annex I systems.
Digital twin (digital twins, urban digital twin): A live digital model of a physical asset, process or system, kept up to date with sensor data and used to simulate, monitor and predict its behavior.
Dilution refrigerator: A cryogenic cooler that reaches a few thousandths of a degree above absolute zero by mixing two helium isotopes. Superconducting quantum computers run inside one.
Direct-to-device (direct-to-cell, D2D, direct-to-handset, direct-to-phone)In Aerospace: Satellite service that connects directly to ordinary mobile phones, without a dish or special terminal. It needs very large antennas in orbit and spectrum shared with or leased from mobile operators.
Directed self-assembly (DSA, block copolymer, BCP): Using block copolymers, molecules made of two incompatible parts, that naturally arrange into fine regular patterns guided by a coarser printed template. It can multiply pitch or repair patterns.
Distributed control system (DCS): A plant-wide control system that runs many control loops through controllers spread across a facility, common in chemical and power plants.
Domain generalization (domain-generalisation): A model's ability to perform well on data from settings it was not trained on, such as new scanners, sites or materials.
DRAM (dynamic random-access memory): Dynamic random-access memory, the main working memory in computers. Each bit is stored as charge on a tiny capacitor that must be refreshed many times per second.
Dual-use (dual-use technology, dual-use goods)In Aerospace: Describes goods, software and technology with both civilian and military uses, such as drones, machine tools and some chips. They are controlled under regimes such as the EAR and the Wassenaar Arrangement.
Dynamic positioning (DP): A computer-controlled system that holds a ship in place using its thrusters, essential during cable repair and offshore work.

E

EAR (Export Administration Regulations)In Aerospace: The US Commerce Department rules that control exports of commercial and dual-use items, software and technology. Items are classified on the Commerce Control List.
EBITDAIn Finance: Earnings before interest, taxes, depreciation and amortisation, a rough measure of operating earnings before financing and non-cash charges. It is not defined under GAAP and ignores capital spending and working capital needs.
ECCN (export control classification number)In Aerospace: A five-character code that places an item on the Commerce Control List, such as 2B001 for certain machine tools. It determines which license requirements apply.
Edge computing (edge AI, edge analytics): Processing data near where it is produced, such as on a device or local gateway, instead of in a distant cloud. It cuts delay and bandwidth needs.
Edge placement error (EPE): The total error in where a feature's edge lands, combining critical dimension, overlay and random variation. It is the budget that ultimately limits patterning.
Effective number of bits (ENOB): The real precision of an analog or mixed-signal system after noise is counted. Photonic and analog computers often reach only a few effective bits.
Electroencephalography (EEG): Recording the brain's electrical activity with electrodes on the scalp. It is the main signal source for non-invasive brain-computer interfaces.
Electron-beam lithography (e-beam lithography, EBL, direct write, maskless lithography, variable-shaped beam, VSB): Writing patterns directly with a focused electron beam, without a mask. It is very fine but slow, so it is used mostly to make masks and for research.
Electronic design automation (EDA): Software used to design, simulate and verify chips, from logic design to physical layout. A few vendors dominate it.
Electronic warfare (EW)In Aerospace: Military use of the electromagnetic spectrum to attack, protect or exploit, including jamming, spoofing and intercepting enemy signals.
Embedding (vector embedding, embedding model): A list of numbers that represents the meaning of text, images or other data, so that similar items sit close together. Embeddings power semantic search and retrieval.
Energy per bit (pJ/bit, fJ/bit, energy per operation): The energy needed to move or store one bit of data, in picojoules (pJ) or femtojoules (fJ). It is the key efficiency measure for interconnects and memory.
Entanglement (quantum entanglement): A quantum link between particles whose measurement results stay correlated however far apart they are. It is a key resource for quantum computing and communication.
Entity List (BIS Entity List): A US Commerce Department list of foreign companies and organizations that need a license to receive US-controlled items, which is usually denied. Many Chinese chipmakers are on it.
EPA (U.S. Environmental Protection Agency)In Sustainability: The federal agency that runs the Renewable Fuel Standard and its lifecycle analysis, including ILUC.
ERCOT (Electric Reliability Council of Texas): The Electric Reliability Council of Texas, which runs most of the Texas power grid and its wholesale market. The grid is largely isolated from the rest of the US.
Error-correcting code (ECC memory): Extra data stored with information so that errors can be detected and fixed automatically. Memory and storage rely on it as cells shrink.
EU AI Act (AI Act, Regulation (EU) 2024/1689, Artificial Intelligence Act): Regulation (EU) 2024/1689, the EU law regulating AI by risk level. Bans on prohibited practices applied from February 2025 and general-purpose model rules from August 2025; a 2026 amendment moved most high-risk obligations to December 2027 and August 2028.
EU Chips Act (European Chips Act, Regulation (EU) 2023/1781, Chips Act 2.0): Regulation (EU) 2023/1781, in force since September 2023, which supports chip research, pilot lines and new fabs and aims for 20% of the global semiconductor market by 2030. A Chips Act 2.0 was proposed in June 2026.
EuroHPC Joint Undertaking (EuroHPC, AI Factory, AI factories): The EU body that funds and buys supercomputers, AI factories and quantum computers together with member states. The machines are hosted in national centers.
EUV light source (laser-produced plasma, LPP, EUV source): The system that produces 13.5 nm light for EUV scanners, usually by hitting tiny tin droplets with a powerful laser to make glowing plasma. Source power sets scanner throughput.
EUV lithography (EUV, extreme ultraviolet, EUVL): Lithography using extreme ultraviolet light at 13.5 nm, which needs vacuum and mirrors instead of lenses. It prints the finest features in leading chips; ASML is the only supplier of EUV scanners.
Exothermic reactionIn Sustainability: A chemical reaction that releases heat. Hydrotreating is strongly exothermic, and plants recover that heat to cut their energy use.

F

Fab (fabrication plant, semiconductor fab): A semiconductor fabrication plant. Leading-edge fabs cost tens of billions of US dollars and take years to build and ramp.
Fault-tolerant quantum computing (fault tolerance, fault-tolerant): Quantum computing in which error correction keeps errors under control however long the computation runs. It is the goal for running algorithms such as Shor's.
FinFET (fin field-effect transistor): A transistor in which the channel is a thin vertical fin wrapped by the gate on three sides, giving better control than flat transistors. It has been standard since the 22 nm–16 nm era and is giving way to gate-all-around designs.
Finite-difference time-domain (FDTD)In Aerospace: A computer method for simulating how electromagnetic waves travel and interact with objects, by stepping Maxwell's equations through space and time.
Fog computing (fog layer): A layer of computing between devices and the cloud, using local gateways and servers to filter, process and store data. It suits remote sites with weak connectivity.
Foreign direct product rule (FDPR)In Aerospace: An EAR provision that extends US export controls to foreign-made items produced with certain US technology, software or equipment.
Foundry (contract chipmaker, pure-play foundry): A company that manufactures chips designed by other companies. The model lets designers avoid owning fabs.
Four coincident peak (4CP): A Texas transmission charge based on a customer's demand during the single highest 15-minute grid peak of each month from June to September. Large users cut load at expected peaks to lower the charge.
FPGA (field-programmable gate array, FPGAs): Field-programmable gate array, a chip whose logic can be reconfigured after manufacture. It is used for prototyping, signal processing and low-volume specialized tasks.
Free cash flowIn Finance: Cash generated by operations minus capital expenditures. It is the cash left over that can repay debt, pay dividends, buy back shares or be reinvested.
Free cooling (seawater cooling): Cooling a data center with cold outside air or water instead of mechanical chillers, saving energy.
Free-electron laser (FEL, EUV-FEL): A laser that makes light from a beam of fast electrons in a magnetic field. It could in future supply EUV light for many scanners at once.
Frequency comb (microcomb, optical frequency comb): A light source that emits many evenly spaced wavelengths at once. Chip-scale microcombs could supply many data channels from one laser.
Frontier model (frontier AI): One of the most capable AI models available at a given time, usually from a small number of labs with the largest compute budgets.

G

Gate fidelity (fidelity, two-qubit gate fidelity, gate error): How closely a quantum operation matches the ideal, as a percentage; the error rate is one minus fidelity. Two-qubit gate fidelity is the most watched hardware benchmark.
Gate-model quantum computer (gate-based quantum computing, universal quantum computer): A universal quantum computer that runs algorithms as sequences of quantum gates. It can in principle run any quantum algorithm, unlike an annealer.
GDPR (General Data Protection Regulation): The EU General Data Protection Regulation, which governs how personal data is collected, used and transferred. It includes a right to erasure and limits on transfers outside the EU.
General-purpose AI model (GPAI, GPAI model): Under the EU AI Act, a model trained on broad data that can perform many tasks, such as a large language model. Its providers face transparency and copyright duties from August 2025.
GNSS (global navigation satellite system, satellite navigation)In Aerospace: The general term for satellite navigation systems, including GPS, Galileo, GLONASS and BeiDou. Receivers compute position from the timing of signals from several satellites.
GPU (graphics processing unit, GPUs): Graphics processing unit, a processor with thousands of small cores that run the same operation in parallel. Originally for graphics, it is now the main chip for AI training and inference.
Grover's algorithm (Grover): A quantum search algorithm that finds an item among N options in about √N steps instead of N. It weakens, but does not break, symmetric encryption.

H

Half-pitch (metal pitch): Half the distance from one line to the next in a repeating pattern, the standard measure of lithographic resolution.
Hallucination: When an AI model states something false or invented as if it were fact, such as a fake citation. Retrieval and verification steps aim to reduce it.
Halving (bitcoin halving, block subsidy, block reward): The scheduled cut in half of bitcoin's block reward, about every four years. The April 2024 halving reduced it to 3.125 BTC per block.
Hard disk drive (HDD, HAMR, heat-assisted magnetic recording): A storage drive that records data magnetically on spinning platters. It remains the cheapest storage per terabyte for large, less active data.
Hash rate (hashrate, EH/s, J/TH, TH/s): The computing power devoted to bitcoin mining, in hashes per second; network totals are in exahashes per second (EH/s). Machine efficiency is measured in joules per terahash (J/TH).
Heat pumpIn Batteries: A device that moves heat from a cooler place to a warmer one using electricity, delivering several units of heat for each unit of power used.
High-bandwidth memory (HBM, HBM3E, HBM4): DRAM dies stacked vertically and placed beside a processor on an interposer, giving very high bandwidth. It is essential for AI accelerators.
High-NA EUV (High-NA, 0.55 NA, Hyper-NA): EUV scanners with a numerical aperture of 0.55 instead of 0.33, which print finer features in one exposure. They use anamorphic optics and expose half the field size, so large chips are stitched.
High-performance computing (HPC, supercomputing, supercomputer): Using supercomputers or clusters of many processors working in parallel to solve large scientific and engineering problems.
High-volume manufacturing (HVM): Full-scale production of a chip or process, after development and pilot stages.
HIPAA (Health Insurance Portability and Accountability Act, protected health information, PHI): The US Health Insurance Portability and Accountability Act, which protects patient health information and governs how healthcare providers and their vendors handle it.
HNSW (hierarchical navigable small world): Hierarchical navigable small world, a graph-based index for approximate nearest neighbor search. It is fast and accurate but memory-hungry.
Hybrid bonding: Joining chips or wafers face to face with direct copper-to-copper and oxide bonds, without solder bumps. It allows much denser connections for stacked memory and logic.
Hybrid quantum-classical computing (hybrid quantum-classical): Running workloads that split tasks between quantum processors and classical computers, often in supercomputing centers. The integration software and scheduling are major challenges.
Hyperscaler (hyperscale): One of the largest cloud companies that run vast data center fleets, such as Amazon, Microsoft and Google.

I

IEA (International Energy Agency)In Sustainability: The Paris-based intergovernmental energy analysis body.
IEC 62443: The international standards series for cybersecurity of industrial automation and control systems.
Industrial Internet of Things (IIoT, industrial IoT): Connected sensors, machines and software used in factories, plants and utilities to monitor and control operations.
Industrial wireless standards (WirelessHART, ISA100.11a): Wireless protocols designed for plant instrumentation, notably WirelessHART and ISA100.11a, which form reliable mesh networks of sensors.
Inference (AI inference): Running a trained model to produce outputs, such as answers or predictions. Its share of AI compute grows as models are put to wider use.
Insertion lossIn Aerospace: The signal power lost when a component is placed in a transmission path, given in decibels.
Interconnection queue (grid interconnection, large-load interconnection): The list of projects waiting for permission and studies to connect to the power grid. Long queues delay new data centers and power plants.
Internet of Things (IoT, AIoT): Everyday and industrial objects fitted with sensors and network connections so they can report data and be controlled remotely.
Interposer (silicon interposer, photonic interposer, optical interposer): A thin layer, often silicon, on which several chips are mounted side by side and wired together with very fine connections. It links GPUs to high-bandwidth memory.
InverterIn Batteries: A power electronic device that converts direct current (DC) from batteries or solar panels into alternating current (AC). Inverter losses and size are part of every storage system's cost.
IoT gateway (edge gateway, network gateway): A device that links local sensors or networks to a wider network or the cloud, often filtering and converting the data.
Ising model: A physics model of interacting spins that point up or down. Finding its lowest-energy state is equivalent to many optimization problems, which is why annealers use it.
ITAR (International Traffic in Arms Regulations)In Aerospace: The US State Department rules that control exports of defense articles, services and technical data on the US Munitions List. Sharing controlled data with a foreign person, even inside the US, can count as an export.
ITU (International Telecommunication Union, ITU-R)In Aerospace: The UN agency that coordinates use of radio spectrum and satellite orbits worldwide. Satellite operators file with it to secure frequencies and orbital positions.

J

Josephson junction: Two superconductors separated by a very thin barrier through which current tunnels. It is the nonlinear element in superconducting qubits and superconducting logic.

K

Knowledge graph (knowledge graphs): A database of entities and the relationships between them, such as companies, people and contracts. Graph-based retrieval uses it to answer questions that span many documents.

L

Landauer limit (Landauer's principle): The minimum energy needed to erase one bit of information, about kT ln 2, ≈3 × 10⁻²¹ J at room temperature. Today's chips use vastly more than this per operation.
Large language model (LLM, LLMs): A neural network trained on very large amounts of text to predict the next token, which lets it answer questions, write and reason in language. Chat assistants are built on them.
Laser source (VCSEL, vertical-cavity surface-emitting laser, DFB laser, distributed feedback laser): The light source for optical links. VCSELs emit from the chip surface and suit short links; distributed feedback lasers give a single clean wavelength for longer links.
Line-edge roughness (LER, line-width roughness, LWR): Random variation along the edges of printed lines. At small sizes it can cause shorts or open circuits and limits resist performance.
Liquid cooling (direct-to-chip cooling, immersion cooling): Cooling servers by circulating liquid to cold plates on the chips, or by immersing servers in a non-conductive fluid. It handles far higher heat loads than air.
Logical qubit (logical qubits, logical error rate): An error-corrected qubit built from many physical qubits. Its logical error rate must fall as more physical qubits are added for error correction to work.
LoRa and LoRaWAN (LoRa, LoRaWAN): LoRa is a long-range, low-power radio technique; LoRaWAN is the network protocol built on it. They suit battery sensors and off-grid mesh networks.
Low Earth orbit (LEO)In Aerospace: Orbits from ≈160 km up to 2,000 km altitude, where satellites circle Earth in ≈90–130 minutes. The short distance means little signal delay, but each satellite sees a small area, so coverage needs many.
LPWAN (low-power wide-area network, NB-IoT): Low-power wide-area network, a class of wireless networks that send small amounts of data over kilometres on very little power, such as LoRaWAN and NB-IoT.

M

Machine learning (ML): Software that learns patterns from data instead of following hand-written rules. Most modern AI, from image recognition to language models, is built this way.
Manufacturing yield (yield, first-pass yield)In Batteries: The share of cells or electrodes made that pass quality tests. New lines often start with low yields; reaching ≈90% or more is a key test of a factory ramp.
Mask blank (EUV mask blank, Mo/Si multilayer): The unpatterned substrate from which a photomask is made. EUV blanks carry dozens of alternating molybdenum and silicon layers that reflect EUV light.
Matrix multiplication (GEMM, general matrix multiplication): Multiplying grids of numbers, the operation that dominates neural network training and inference. Accelerators are designed to do it as fast and efficiently as possible.
Memory wall: The growing gap between how fast processors can compute and how fast memory can supply them with data. It often limits AI performance more than raw compute does.
Memristor: A circuit element whose resistance depends on the current that has flowed through it and is retained without power. Arrays of memristors can store weights and compute in place.
Mesh network (mesh networking): A network in which devices relay data for each other, so messages can hop around failures without a central tower or router.
Metaheuristic (metaheuristics, genetic algorithm, evolutionary algorithm, simulated annealing): A general strategy for finding good, though not guaranteed best, answers to hard optimization problems, such as genetic algorithms or simulated annealing.
Metal-oxide resist (MOR, dry resist): An EUV photoresist built from metal-oxide clusters, often tin-based, which absorbs EUV light more strongly and resists etching better than organic resists. Some versions are deposited dry from vapor.
MetasurfaceIn Aerospace: A thin layer of engineered, repeating structures smaller than the wavelength, which shapes radio waves or light in ways ordinary materials cannot.
Metrology (CD-SEM, scatterometry, wafer inspection): Measuring features and layers on wafers to control the process, with tools such as CD-SEMs, which use electron microscopes, and optical scatterometry.
Micro-CT (microtomography, X-ray microtomography, phase-contrast microtomography): X-ray computed tomography at micrometre resolution, which builds a 3D image of an object's interior from many X-ray views. Phase-contrast micro-CT can reveal materials with almost no density difference.
Microcontroller (MCU, microcontrollers): A small, low-power computer on a single chip with its own memory and input-output pins, used to control devices such as sensors and appliances.
MicrogridIn Batteries: A local grid with its own generation and storage that can run connected to the main grid or alone, in island mode. Ports, campuses and remote sites use microgrids for resilience.
Microring resonator (microring, micro-ring resonator, MRR): A tiny ring-shaped waveguide that strongly affects light only at certain wavelengths. It is compact and wavelength-selective but sensitive to temperature.
Model drift (data drift): The gradual loss of a model's accuracy as real-world data shifts away from the data it was trained on. It is managed by monitoring and retraining.
MTEB (Massive Text Embedding Benchmark): The Massive Text Embedding Benchmark, a public leaderboard comparing embedding models across retrieval, clustering and classification tasks.
Multi-beam mask writer (MBMW, multibeam mask writer): An electron-beam tool that writes a photomask with hundreds of thousands of beams at once, making complex curved mask shapes practical.
Multi-column electron-beam lithography (MEBL, multi-column e-beam): E-beam lithography using many independent electron columns in parallel to raise throughput enough for some direct writing on wafers.
Multipatterning (multi-patterning, double patterning, SADP, SAQP, LELE, self-aligned double patterning, self-aligned quadruple patterning): Printing a dense layer in two or more exposure and etch steps so each step prints sparser features. It extends older lithography to finer pitches at the cost of more steps, time and overlay risk.
Multiply-accumulate (MAC, MACs, multiply-accumulate operation): The basic operation of neural network computing: multiply two numbers and add the result to a running total. Hardware efficiency is often quoted as energy per MAC.

N

NAND flash (NAND, flash memory): Non-volatile memory that keeps data without power, used in SSDs, phones and memory cards. It is cheaper per bit than DRAM but slower and wears out with writes.
Nanoimprint lithography (NIL, nanoimprint, J-FIL): Pressing a patterned template into liquid resist on a wafer and curing it, like stamping, instead of projecting light. It is cheap and fine-grained but struggles with defects and overlay.
National Quantum Initiative Act: The 2018 US law that set up a coordinated federal program for quantum research across NIST, the National Science Foundation and the Department of Energy.
Network difficulty (mining difficulty, bitcoin difficulty): A bitcoin setting that adjusts about every two weeks so blocks keep arriving every ≈10 minutes. Rising difficulty means each machine earns less.
Network segmentation: Dividing a network into separate zones so an intrusion in one cannot easily spread to others, such as keeping plant control systems apart from office IT.
Neural network (artificial neural network, CNN, convolutional neural network, recurrent neural network): A model made of layers of simple connected units whose weights are adjusted during training. Variants include convolutional networks for images and recurrent networks for sequences.
Neuromorphic computing (neuromorphic chip): Hardware designed to mimic the brain's structure, often with spiking neurons and memory placed next to processing, to cut the energy cost of AI.
NISQ (noisy intermediate-scale quantum): Noisy intermediate-scale quantum: the current era of quantum computers with tens to hundreds of qubits but no full error correction, which limits how long a computation can run.
NIST PQC standards (FIPS 203, FIPS 204, FIPS 205, ML-KEM, ML-DSA, SLH-DSA, HQC): The US post-quantum cryptography standards: FIPS 203 (ML-KEM) for key exchange, FIPS 204 (ML-DSA) and FIPS 205 (SLH-DSA) for digital signatures, all published in August 2024. NIST chose HQC as a backup key-exchange method in March 2025.
Non-terrestrial network (NTN)In Aerospace: The 3GPP term for parts of a mobile network carried by satellites or high-altitude platforms instead of ground towers. Standards from Release 17 onward let phones and base stations work through them.
Non-volatile memory (NVM, MRAM, ReRAM, resistive RAM, magnetoresistive RAM): Memory that keeps data when power is off, such as flash, MRAM, ReRAM and phase-change memory.
NP-hard: A class of problems for which no known method finds the exact best answer quickly as the problem grows. Routing and scheduling problems are typical examples.
NPU (neural processing unit): Neural processing unit, a processor block designed for AI tasks, often built into phones, laptops and other devices.
Numerical aperture (NA): A measure of the range of angles a lens or mirror system collects light from. Higher numerical aperture gives finer resolution but less depth of focus.
NVMe (Non-Volatile Memory Express): Non-Volatile Memory Express, the standard protocol for connecting SSDs over the PCIe bus, giving far lower delay than older SATA connections.

O

Open Compute Project (OCP): An industry group that publishes open hardware designs for data center servers, storage, racks and power.
Operational technology (OT): The hardware and software that monitor and control physical equipment, such as pumps, valves and production lines, as distinct from office IT.
Operations research: The use of mathematical models and optimization to make better decisions in logistics, scheduling, networks and resource allocation.
Optical circuit switch (OCS): A switch that redirects light beams between fibers without converting them to electrical signals, saving power in large AI clusters.
Optical I/O (optical I/O chiplet, optical interconnect): Optical input-output connections built close to or onto a chip, carrying data off-chip as light instead of electrical signals.
Optical modulator (modulator, Mach-Zehnder modulator, MZM, MZI, Mach-Zehnder interferometer, microring modulator, MRM, electro-absorption modulator, EAM): A device that encodes data onto light by changing its intensity or phase. Common types are Mach-Zehnder interferometers, microring modulators and electro-absorption modulators.
Orbital data center (orbital compute, space-based data center)In Aerospace: A proposed computing facility in orbit, powered by solar arrays and cooled by radiating heat into space. Backers cite abundant solar power; launch cost, heat rejection and maintenance are the main obstacles.
OSAT (outsourced semiconductor assembly and test): Outsourced semiconductor assembly and test: companies that package and test chips for chipmakers.
Overlay: How accurately each new layer lines up with the layers already on the wafer, measured in nanometres. Overlay errors break connections between layers.
OWASP Top 10 for LLM applications (OWASP, Open Worldwide Application Security Project): A list of the most serious security risks for applications built on large language models, published by the Open Worldwide Application Security Project. Prompt injection ranks first.

P

PAM4 (PAM-4, pulse amplitude modulation): Four-level pulse amplitude modulation, which sends two bits per symbol by using four signal levels. It doubles data rate over simple on-off signalling.
PCIe (PCI Express): PCI Express, the standard high-speed connection between a processor and devices such as GPUs, SSDs and network cards. Each generation roughly doubles bandwidth per lane.
Pellicle (EUV pellicle, carbon nanotube pellicle): A thin transparent membrane stretched over a photomask to keep particles out of focus. EUV pellicles must survive intense heat while letting most light through.
Performance metrics for compute (FLOPS, PFLOP/s, TOPS, TOPS/W, exaflop, exascale): Common measures of processor capability. FLOPS count floating-point operations per second; TOPS count trillion operations per second, often low precision; TOPS/W measures energy efficiency.
Personally identifiable information (PII): Data that can identify a person, such as a name, ID number or location. Privacy laws restrict how it is stored and shared.
Phase-change memory (PCM memory, phase-change chalcogenide, GST, Ge-Sb-Te): Memory that stores bits by switching a chalcogenide glass between crystalline and amorphous states, which differ in resistance and optical properties. It is used in electronic memory and photonic computing research.
Phased array (phased-array antenna, electronically scanned array, AESA)In Aerospace: An antenna made of many small elements whose signal timing is adjusted electronically to steer the beam without moving parts. Radars, satellite terminals and satellites use it.
Photodetector (photodiode, APD, avalanche photodiode): A device that converts light into an electrical signal. Avalanche photodiodes add internal gain for weak signals.
Photolithography (lithography, optical lithography): Printing chip patterns by shining light through a patterned mask onto light-sensitive resist on a wafer. It is the step that sets the smallest features a fab can make.
Photomask (mask, reticle): A plate carrying a chip layer's pattern that is projected onto the wafer during lithography. A full chip needs dozens of masks, each costing a great deal at advanced nodes.
Photonic computing (optical computing, photonic neural network, PNN, photonic tensor core): Doing computation, usually matrix multiplication for AI, with light instead of electrons. It promises speed and efficiency but converting data between light and electronics eats much of the gain.
Photonic integrated circuit (PIC, PICs, photonic chip): A chip that routes and processes light, combining many optical components on one die.
Photonic materials (indium phosphide, InP, silicon nitride, SiN, III-V): The main material platforms for photonic chips. Silicon is cheap and CMOS-compatible but cannot emit light; indium phosphide makes lasers; silicon nitride gives low-loss waveguides.
Photoresist (EUV resist): A light-sensitive coating on the wafer that changes chemically where exposed, so the pattern can be developed and etched. Resist chemistry limits resolution and dose.
Pilot line (pilot plant, demonstration line)In Batteries: A small production line used to prove a cell design and process before full-scale manufacturing. Pilot output is far too small to show whether costs will fall at scale.
Pluggable transceiver (pluggable optics, optical transceiver, 800G, 1.6T, linear-drive pluggable optics, LPO): A removable module that converts electrical signals to optical and back, plugged into the front of a switch or server. Speeds such as 800G and 1.6T refer to its data rate.
Post-quantum cryptography (PQC, quantum-safe cryptography): Encryption and signature methods designed to resist attack by quantum computers. NIST published its first standards, FIPS 203, 204 and 205, in August 2024.
Power usage effectiveness (PUE): Total data center power divided by the power used by the computing equipment. A value of 1.0 would be perfect; efficient modern sites reach ≈1.1–1.2.
Predictive maintenance: Using sensor data and analytics to predict when equipment will fail so it can be fixed just before, reducing downtime.
Private 5G (private LTE, private cellular): A dedicated cellular network owned or operated for one site, such as a factory, port or farm, giving reliable coverage and control.
Process design kit (PDK, PDKs): The set of files and rules a foundry gives chip designers to describe its manufacturing process, so designs can be built correctly there.
Process node (technology node, leading-edge node, mature node): A name for a generation of chip manufacturing technology, such as 3 nm or 2 nm. The number no longer matches any physical feature; it signals relative density and performance.
Process safety rules (PSM, Process Safety Management, RMP, Risk Management Program): The main US rules for hazardous chemical plants: OSHA's Process Safety Management standard protects workers, and EPA's Risk Management Program protects nearby communities.
Processing-in-memory (PIM, near-memory computing, compute-in-memory, in-memory computing): Placing compute logic inside or right next to memory so data does not have to travel to a separate processor. It targets the energy and delay of moving data.
Programmable logic controller (PLC, PLCs): A rugged industrial computer that controls machines and processes in real time.
Prompt injection (indirect prompt injection): An attack that hides instructions in text an AI model reads, such as a web page or document, to make it ignore its rules or leak data.
Public-key cryptography (RSA, elliptic-curve cryptography): Encryption that uses a public key to encrypt or verify and a private key to decrypt or sign. RSA and elliptic-curve cryptography, the current standards, are vulnerable to Shor's algorithm.

Q

qLDPC codes (quantum low-density parity-check codes, qLDPC): Quantum low-density parity-check codes, error-correcting codes that could encode logical qubits with far fewer physical qubits than the surface code, but need longer-range connections.
Quantization (int8 quantization, product quantization): Storing a model's numbers at lower precision, such as 8-bit integers instead of 32-bit floats, to cut memory and speed up computation at a small accuracy cost.
Quantum advantage (quantum supremacy, quantum utility): When a quantum computer solves a useful problem faster, cheaper or better than the best classical methods. Quantum supremacy refers to beating classical computers on any task, useful or not.
Quantum annealing (quantum annealer): A type of quantum computing that solves optimization problems by letting a system of qubits settle into its lowest-energy state. It is limited to certain problem types and is not a universal quantum computer.
Quantum error correction (QEC, error correction): Encoding one logical qubit across many physical qubits so errors can be detected and fixed during computation. It is required for large, reliable quantum computers.
Quantum gate (two-qubit gate, CZ gate, quantum circuit, circuit depth): A basic operation on one or more qubits, analogous to a logic gate in classical computing. Algorithms are built as circuits of gates.
Quantum key distribution (QKD): Sharing encryption keys using quantum states of light, so any eavesdropping disturbs the signal and is detected. It needs special hardware and distance-limited links.
Quantum processing unit (QPU): The chip or system that holds and controls qubits, the quantum counterpart of a CPU or GPU.
Quantum volume: A single-number benchmark of a quantum computer's capability that combines qubit count, fidelity and connectivity. Higher is better.
Quantum-inspired algorithm (quantum-inspired): A classical algorithm that borrows ideas from quantum computing, such as annealing or tensor networks, and runs on ordinary hardware.
Qubit (qubits, quantum bit, physical qubit): The basic unit of quantum information, which can be in a blend of 0 and 1 at once. Physical qubits are error-prone; logical qubits combine many physical ones to correct errors.
Qubit modalities (trapped-ion, trapped ion, neutral-atom, neutral atom, photonic qubit, spin qubit): The main physical ways to build qubits: superconducting circuits, trapped ions held by electric fields, neutral atoms held by lasers, photons, and electron spins in semiconductors. Each trades speed, fidelity and scalability.
QUBO (quadratic unconstrained binary optimization): Quadratic unconstrained binary optimization, a standard way to write optimization problems as yes-no variables with pairwise costs. It is the input format for quantum annealers and many solvers.

R

Rack density (kW per rack): The power drawn by one rack of servers, in kilowatts. AI racks draw several times more than traditional racks, which pushes data centers to liquid cooling.
Radiation hardening (radiation-hardened, rad-hard, radiation-tolerant, single-event effects, SEE, total ionizing dose, TID): Designing or screening chips to survive radiation in space or near reactors. Single-event effects are glitches from individual particles; total ionizing dose is damage that builds up over time.
RAGAS (Retrieval-Augmented Generation Assessment): Retrieval-augmented generation assessment, an open-source framework that scores RAG systems on measures such as faithfulness to sources and answer relevance.
Randomized benchmarking: A standard method for measuring quantum gate error rates by running random gate sequences and seeing how fast results degrade.
Rayleigh criterion (k1, k1 factor, Rayleigh equation): The formula for the smallest printable half-pitch: k1 × wavelength ÷ numerical aperture. Shorter wavelength, higher numerical aperture or lower k1 all shrink features; k1 cannot go below 0.25 in one exposure.
Reaction wheel (momentum wheel)In Aerospace: A flywheel inside a spacecraft that is spun faster or slower to turn the spacecraft the opposite way, letting it point without using propellant. Most are driven by permanent-magnet motors.
Reciprocal rank fusion (RRF): A simple way to merge ranked lists from different search methods by summing scores based on each item's rank.
Reconfigurable intelligent surface (RIS)In Aerospace: A metasurface whose elements can be switched electronically to redirect radio signals, for example around obstacles in a wireless network.
Reinforcement learning (RL): Training a system by trial and error, rewarding actions that lead to good outcomes. It powers game-playing systems and is used to refine language models.
Remotely operated vehicle (ROV, work-class ROV): An uncrewed underwater robot controlled from a ship through a tether, used for cable repair, inspection and seabed work.
Reranking (reranker, cross-encoder): A second retrieval step in which a more accurate model reorders the top search results before they reach the language model.
Reservoir computing: A machine learning approach that feeds inputs into a fixed, complex dynamic system and trains only a simple readout layer. Physical systems such as optics can serve as the reservoir.
Retrieval metrics (nDCG, nDCG@10, top-k, retrieval recall): Measures of search quality. Recall is the share of relevant items found; precision is the share of results that are relevant; nDCG also rewards ranking relevant items near the top.
Retrieval-augmented generation (RAG, GraphRAG, agentic RAG): A method that searches a document collection for relevant passages and gives them to a language model to ground its answer. It lets models use private or recent information and cite sources.
Reversible computing: Computing designed so that operations can be undone without erasing information, which in principle avoids the Landauer energy cost.
RLS trade-off (RLS triangle, resolution, line-edge roughness and sensitivity): The tension between resolution, line-edge roughness and sensitivity in photoresists: improving one usually worsens another.
ROADM (reconfigurable optical add-drop multiplexer): Reconfigurable optical add-drop multiplexer, a network node that adds, drops or routes individual wavelengths remotely without converting them to electrical signals.

S

SCADA (supervisory control and data acquisition): Supervisory control and data acquisition, software that gathers data from remote equipment and lets operators monitor and control it.
Section 48D (48D, advanced manufacturing investment credit): The US advanced manufacturing investment credit for building semiconductor fabs and tool plants, created by the CHIPS Act. The 2025 OBBBA law raised it from 25% to 35% for property placed in service after 2025.
Secure boot: A startup process that checks a device's software is genuine and unaltered before running it.
Self-play: A reinforcement learning method in which a system improves by competing against copies of itself, without human examples.
SerDes (serializer/deserializer): Serializer/deserializer, the circuit that converts parallel data into a fast serial stream for transmission and back. Lane speeds such as 200 Gb/s set chip-to-chip bandwidth.
Seveso III Directive (Seveso III, Directive 2012/18/EU): Directive 2012/18/EU, the EU law on preventing and limiting major accidents at sites that store dangerous chemicals.
Shor's algorithm (Shor): A quantum algorithm that factors large numbers efficiently, which would break RSA and elliptic-curve encryption once large fault-tolerant quantum computers exist.
Shot noise (photon shot noise, Poisson noise): Random variation in the number of photons or electrons arriving, which grows relatively larger as doses shrink. It drives stochastic defects in lithography and noise in optical detectors.
Silicon photonics (SiPh): Building optical components, such as waveguides, modulators and detectors, on silicon wafers using chip-industry processes. It makes optical links cheap enough for data centers.
Silicon-on-insulator (SOI, FD-SOI, fully depleted SOI, buried oxide): A wafer with a thin silicon layer on top of an insulating oxide layer. It reduces leakage in chips and is the standard substrate for silicon photonics.
Smart Cut: A layer-transfer process that splits a thin crystal layer from a donor wafer using implanted hydrogen and bonds it to another wafer. It is the main way silicon-on-insulator wafers are made.
Software as a service (SaaS): Software delivered over the internet on subscription instead of installed and owned. Customers pay per user or per use.
Solid-state drive (SSD, SSDs): A storage drive built from flash memory with no moving parts. It is much faster than a hard drive and is the main storage in data centers for active data.
Spiking neural network (SNN): A neural network whose units communicate with discrete timed pulses, like biological neurons. It suits low-power neuromorphic hardware.
SRAM (static random-access memory, cache memory): Static random-access memory, a fast memory that holds data without refreshing but takes much more chip area per bit than DRAM. Processors use it for caches.
SSVEP (steady-state visual evoked potential): Steady-state visual evoked potential, a brain signal produced when a person looks at a light flickering at a set frequency. Brain-computer interfaces use it to detect which target a user is looking at.
Stochastic defects (stochastics, stochastic failures): Random printing failures, such as missing contacts or bridged lines, caused by the small number of photons and molecules involved at EUV scales.
Storage tiers (nearline storage, cold storage, hot storage): Classes of storage matched to how often data is used: hot data on fast SSDs, warm or nearline data on cheaper drives, and cold data on archival media.
Store-and-forward (store and forward): Holding data locally until a connection is available, then sending it. Remote sites use it to ride out network gaps.
Submarine cable (undersea cable, subsea cable, submarine fiber): A fiber-optic cable laid on the seabed to carry internet traffic between continents. Repeaters along the route boost the signal; landing stations connect it to land networks.
Submarine cable components (repeater, branching unit, cable landing station, SLTE, submarine line terminal equipment, wet plant, dry plant): The main parts of an undersea cable system. Wet plant is everything in the sea, such as repeaters and branching units; dry plant is the shore equipment, including the submarine line terminal equipment (SLTE).
Superconducting logic (single flux quantum, SFQ, RSFQ, rapid single flux quantum, ERSFQ, RQL, reciprocal quantum logic, AQFP, adiabatic quantum flux parametron): Digital circuits built from superconductors and Josephson junctions that switch with very little energy. Families include RSFQ, energy-efficient ERSFQ, RQL and the adiabatic AQFP.
Superconducting qubit (transmon, transmons, superconducting quantum computer): A qubit made from a superconducting circuit cooled to millikelvin temperatures. Transmons, the most common type, use Josephson junctions; gates are fast but coherence is short.
Superposition: A quantum state that combines several possible values at once, with each outcome appearing with some probability when measured.
Surface code: The leading quantum error-correcting code, which arranges qubits in a grid and needs only neighboring connections. It needs many physical qubits per logical qubit.
SWAP overhead (SWAP gate): Extra operations needed to move quantum information between qubits that are not directly connected. Limited connectivity raises it and adds errors.
System-on-chip (SoC): A single chip that combines processors, memory controllers, graphics, AI blocks and input-output functions. Phones, drones and embedded devices run on them.

T

Technology readiness level (TRL)In Aerospace: A 1–9 scale, developed at NASA, for how mature a technology is, from basic principles observed to proven in operation. Agencies and investors use it to judge development risk. See the table
Texas SB 6 (Senate Bill 6): A 2025 Texas law setting rules for loads of 75 MW or more, such as data centers and crypto miners. It requires remote curtailment equipment, interconnection study fees and approval of co-location deals.
Thermal design power (TDP): The maximum sustained heat a chip is designed to produce, in watts, which its cooling must remove.
Thermal management (battery thermal management, BTMS)In Batteries: Keeping battery cells in their best temperature range, typically ≈15–35 °C, with air, liquid or immersion cooling and heating. It affects life, safety, power and charging speed.
Thermo-optic and electro-optic effects (thermo-optic effect, electro-optic effect): Ways to tune optical devices. The thermo-optic effect changes refractive index with heat; the electro-optic effect changes it with an electric field, which is much faster.
Thin-film lithium niobate (TFLN, lithium niobate on insulator, LNOI): A thin layer of lithium niobate crystal bonded on insulator, used for very fast, low-power optical modulators.
Token (tokens): The unit of text a language model reads and writes, typically a word or part of a word. Model usage and pricing are counted in tokens.
TPU (tensor processing unit): Tensor processing unit, Google's custom AI accelerator for training and inference, offered mainly through Google Cloud.
Training (model training, pre-training): The compute-heavy process of adjusting a model's weights on data until it performs a task well. Training a frontier model can occupy tens of thousands of GPUs for months.
Transformer (transformer model, attention mechanism): The neural network design behind most large language models, built around an attention mechanism that relates every token to every other. Its compute cost grows with the square of input length.
Tunable coupler (tunable couplers): A circuit element between qubits whose coupling can be switched on and off, reducing unwanted interactions in superconducting processors.
TV white space: Unused television broadcast frequencies that can carry long-range wireless broadband in rural areas.

U

Uptime Institute tiers (Tier III, Tier IV, Tier III data center): A widely used rating of data center reliability, from Tier I, with basic capacity and no redundancy, to Tier IV, which is fault tolerant. Tier III sites can be maintained without shutting down.

V

ValuationIn Finance: An estimate of what a company or asset is worth, based on what it owns today and the cash it is expected to generate in future. Common methods are discounted cash flow, comparisons with similar companies, and the value of net assets.
Variational quantum algorithms (VQE, variational quantum eigensolver, QAOA, quantum approximate optimization algorithm): Hybrid algorithms in which a quantum computer evaluates a parameterized circuit and a classical computer adjusts the parameters. VQE targets chemistry problems; QAOA targets optimization.
Vector database (vector store, vector search): A database built to store embeddings and find the ones most similar to a query quickly, even across billions of entries.
Virtual unwrapping: Digitally flattening the layers of a rolled or folded object, such as a carbonized scroll, from a 3D scan so its surface can be read without opening it.

W

Wafer (silicon wafer, 300 mm wafer): A thin disc of pure silicon, usually 300 mm across, on which many chips are built at once before being cut apart.
Wafers per hour (wph): Lithography tool throughput, the number of wafers exposed per hour. It drives cost per layer as much as tool price does.
Wassenaar Arrangement (Wassenaar)In Aerospace: A multilateral export-control regime in which participating states agree on lists of conventional arms and dual-use goods to control, then apply them in national rules.
Waveguide (optical waveguide): A structure that confines and guides light along a path on a chip or in a fiber, like a wire for light.
Wavelength-division multiplexing (WDM, dense wavelength-division multiplexing, DWDM): Sending several data streams on different wavelengths of light down one waveguide or fiber at once.
Wear leveling (wear-leveling): Spreading writes evenly across a flash drive's cells so no area wears out early.
Write endurance (program/erase cycles, write/erase cycles): How many times flash memory can be written and erased before it wears out, often given as program/erase cycles or drive writes per day.

Z

Zero-trust architecture (zero trust, zero-trust): A security approach in which no user or device is trusted by default, even inside the network, and every access is verified.