Quotes About This Resource from Chair Randall Schriver and Vice Chair Mike Kuiken:
“The competition with China over artificial intelligence carries real stakes for our security and our economy,” said Chair Randy Schriver. “Clear, common language is the foundation for sound policy on issues this consequential subject, and the Commission produced this glossary to provide it. We will keep it current as the field evolves.”
“AI is evolving fast, and the vocabulary is evolving with it. Terms like chiplets, diffusion, data curation, and open versus closed models now sit at the center of real policy choices, and getting the language right keeps good policy on pace,” said Vice Chair Mike Kuiken. “We built this decoder ring as a working reference for everyone tracking these issues in Congress, and across the public.”
Read the full press release here.
The AI Technology Stack Glossary is a plain-English reference guide designed to help Congress and the broader public better understand the complex technical concepts shaping today’s technology and national security landscape. The glossary explains key terms related to semiconductors, semiconductor manufacturing, and artificial intelligence (AI), providing readers with a practical resource for navigating policy discussions, industry developments, and emerging technologies. The terms in this product have been sorted into relevant categories for ease of understanding, but please note that some terms may fall into more than one category. This product will be updated on a semi-regular basis with additional key terms and concepts.
Last updated on October 15, 2026
Term | Definition |
3-D Chip Stacking | A chipmaking process in which multiple chips are vertically stacked on top of each other, rather than laid out flat side-by-side. This design allows data to move more quickly between chips, improving performance and energy efficiency. It is especially useful for artificial intelligence (AI), high-performance computing, and other applications that require large amounts of data to be processed quickly.
|
Advanced Packaging | Techniques used to combine and connect multiple chips or components within a single package so they can work together efficiently. Advanced packaging can improve chip performance, data transfer speeds, and energy efficiency without relying solely on making individual transistors smaller.
|
AI Accelerator | A computer chip designed specifically to speed up data processing for AI tasks. Unlike a central processing unit (CPU), which handles many different types of computing tasks (see below), an AI accelerator is optimized for the calculations used by AI systems.
|
Application-Specific Integrated Circuit (ASIC) | A computer chip designed to perform one specific task rather than a wide range of tasks. ASICs are built for a particular application, which usually makes them faster and more power-efficient for that job than a general-purpose chip such as a central processing unit (CPU).
|
Central Processing Unit (CPU) | The main processor in a computer that carries out instructions and performs a wide range of general computing tasks. Unlike specialized chips designed for specific workloads, such as AI, a CPU is built to handle many different types of tasks, such as calculation, data management, and logical decisions.
|
Chip-on-Wafer-on-Substrate (CoWoS) | An advanced chip-packaging technology developed by Taiwan Semiconductor Manufacturing Company (TSMC) that places multiple chips, such as AI processors and high-bandwidth memory (HBM), very close together in a single package. This allows the chips to exchange data faster and more efficiently than if they were packaged separately. CoWoS is widely used in advanced AI chips, including GPUs.
|
Chip Packaging | The process of enclosing a finished semiconductor chip in a protective structure that allows it to connect to a circuit board, communicate with other components, and manage power, heat, and performance. Advanced chip packaging can also significantly improve computing performance and energy efficiency by enabling faster data transfer, denser integration of multiple chips, and reduced power loss between components.
|
Chiplet | A small chip built to do one part of a bigger computing job, such as processing, memory, or input/output. Several chiplets can be linked together inside one package so they work like a single larger chip.
|
Deep Ultraviolet (DUV) Lithography | A semiconductor manufacturing technology that uses short wavelengths of light (typically 193–248 nanometers [nm]) to print tiny circuit patterns onto silicon wafers (defined below). DUV lithography is used to produce a wide range of semiconductors, from mature-node chips to advanced logic and memory chips, but it is generally less capable than the newer extreme ultraviolet (EUV) lithography used for leading-edge semiconductor manufacturing. Because DUV and EUV systems rely on fundamentally different optical designs, light sources, and materials, existing DUV machines cannot simply be upgraded or converted into EUV machines.
|
Deposition | A step in semiconductor manufacturing that adds extremely thin layers of material (such as silicon, metals, and insulating compounds) onto the surface of a silicon wafer. These layers are used to build the tiny electronic structures and connections, essential for packing transistors into increasingly smaller chips.
|
Die | An individual piece of a semiconductor wafer containing the circuits that make up a chip. A single wafer typically contains many dies, with the number varying by chip type. These dies are cut apart and packaged for use in electronic devices. For example, a microchip for electronics can yield (see definition below) thousands of dies per 300 millimeters (mm) wafer, but a massive AI accelerator (see definition above) die such as Nvidia H100 yields fewer than 80.
|
Dynamic Random-Access Memory (DRAM) | A type of memory that temporarily stores data a computer or processor needs while it is working. DRAM provides quick access to this data but loses the information it stores when the power is turned off.
|
Electronic Design Automation (EDA) Software | Specialized software used to design and test semiconductor chips before they are manufactured. U.S.-based Synopsys and Cadence, and Germany-based Siemens together control 85 percent of the EDA software market as of 2026.
|
Etching | A step in semiconductor manufacturing that removes selected areas of material from the surface of a silicon wafer. This process helps create the tiny patterns and structures that make up a chip.
|
Extreme Ultraviolet (EUV) Lithography | A semiconductor manufacturing technology that uses extremely short wavelengths of light to print tiny circuit patterns on silicon wafers at nanometer scale. Because EUV light can be absorbed by any form of matter (including air), the manufacturing process must operate in a near-perfect vacuum and be repeated dozens of times with extreme precision. Dutch firm ASML is the world’s sole producer of EUV lithography machines, which are required to manufacture the most advanced chips that power today’s cutting-edge AI systems. High-Numerical Aperture (NA) EUV lithography is a more advanced version of standard EUV lithography. Both use extreme ultraviolet light to print chip features, but High-NA EUV uses a stronger optical setup that can print smaller features more precisely.
|
Fabless Manufacturing | A business model in which a company designs a chip but does not make it in its own factory. Instead, it outsources the production to a specialized manufacturer, such as a foundry (defined below).
|
Fabrication Plant (Fab) | A highly specialized facility where raw silicon is processed into semiconductors. A fab uses advanced manufacturing steps such as photolithography (defined above) to build tiny circuits onto wafers.
|
Field-Programmable Gate Array (FPGA) | A type of chip that can be programmed and reprogrammed after it is manufactured to perform different tasks. Unlike chips designed for one specific purpose, FPGAs can be customized for different applications without manufacturing a new chip, although reprogramming is labor-intensive.
|
Foundry | A company that takes chip designs from a customer and manufactures them into physical silicon wafers. Foundries often work with “fabless” companies that design chips but do not own their own manufacturing plants.
|
Graphics Processing Unit (GPU)
| A type of computer chip that is widely used to train and run AI models. Unlike a traditional computer processor (CPU), a GPU can perform many calculations at the same time, allowing it to process the large amounts of data needed for AI much more quickly and efficiently.
|
High-Bandwidth Memory (HBM) | An innovation in memory designed to move data very quickly between memory and a processor by using 3-D chip stacking (see definition above) to place memory very close to the processor. Compared with conventional memory technologies such as DDR and GDDR, HBM provides substantially higher memory bandwidth—often several times greater—while using less power and occupying less physical space. This allows GPUs and other AI chips to operate at full capacity instead of waiting for data.
|
Hybrid Bonding | An advanced chip-packaging technique that directly connects two chips or wafers using extremely small copper connections. This allows chips to be stacked very close together and exchange data faster and more efficiently.
|
Interconnect | The electrical connections that allow different parts of a chip, or different chips in a system, to communicate and exchange data with one another.
|
Language Processing Unit (LPU) | A specialized processor built by U.S.-based AI chip startup Groq that is designed to run large language models (LLMs) and other AI applications quickly and efficiently. LPUs are designed to rapidly process the calculations needed to generate responses from AI models. In December 2025, Groq entered a non-exclusive licensing agreement with Nvidia to scale its LPU.
|
Memory Bandwidth | The amount of data that can be transferred between a processor and its memory in a given amount of time. Higher memory bandwidth allows a processor to access more data more quickly, which is especially important for AI workloads.
|
Metrology | The process of precisely measuring and inspecting the extremely small features and layers created on a semiconductor wafer during manufacturing. Metrology helps manufacturers identify defects and ensure that chips are being produced correctly and consistently—a critical process in AI chip manufacturing as smaller chip nodes increase chances of defects.
|
Multi-patterning | A semiconductor manufacturing technique that uses multiple rounds of patterning to create chip features that are smaller or more closely spaced than can be produced with a single round. Multi-patterning can allow manufacturers to produce more advanced chips using older lithography equipment, but it adds complexity, time, and cost.
|
NAND | A type of memory used to store data for long periods of time, including when a device is turned off. NAND flash memory is commonly used in solid-state drives (SSDs), smartphones, and other electronic devices.
|
Neural Processing Unit (NPU) | A specialized processor designed to perform the calculations used by artificial intelligence and machine learning applications quickly and efficiently. NPUs are increasingly included in computers, smartphones, and other devices to run AI directly on the device.
|
Photoresist | A light-sensitive material applied to a silicon wafer during semiconductor manufacturing. When exposed to light during lithography, the photoresist helps transfer extremely small patterns onto the wafer that are used to build a chip’s circuits.
|
Process Node | A term used to describe a generation of semiconductor manufacturing technology, commonly identified by names such as 7 nanometer (nm), 5 nm, or 3 nm. In general, newer process nodes allow manufacturers to fit more transistors into a given area, which can improve chip performance and energy efficiency.
Modern node names do not directly represent the physical size of a specific chip feature.
|
Performance Density | A measure of how much computing performance a chip can provide within a given physical area (typically calculated by dividing total processing performance (TPP)—see definition below—by die area). Higher performance density means a chip can deliver more computing power while taking up the same amount of space.
|
System-on-Chip (SoC) | A chip that combines many of a computer or electronic device’s major components—such as processors, memory controllers, graphics, and communications functions—into a single chip. Combining these functions can make devices smaller and more energy efficient.
|
Tensor Processing Unit (TPU) | A chip produced primarily for mathematical operations used in AI systems, especially large matrix operations. Unlike a general-purpose CPU, a TPU is built for narrower tasks, which can make it faster and more efficient for training or running AI models.
|
Through-Silicon Via (TSV) | Tiny vertical holes through a silicon chip that are filled with conductive material, usually copper. These holes carry electrical signals or power from one side of the chip to the other. TSVs allow chips to be stacked and connected vertically, creating shorter connections and more compact packages.
|
Total Processing Performance (TPP) | A measure used in U.S. export controls to estimate the overall computing performance of certain advanced chips. TPP is calculated using the chip’s processing speed and the size of the data it can process in each operation. A higher TPP generally indicates greater computing capability.
|
Wafer | A thin, circular slice of semiconductor materials—usually silicon—that serves as the foundation for manufacturing computer chips. Many individual chips are built on a single wafer before being cut apart and packaged for use in electronic devices.
|
Yield | The percentage of chips produced on a semiconductor wafer that work properly and meet required performance standards. Higher yields mean that more usable chips can be produced from each wafer, lowering manufacturing costs.
|
Term | Definition |
Adversarial Distillation Attack | A term coined by leading U.S. AI labs that refers to attempts to copy or extract the knowledge and/or behavior of an AI model by repeatedly querying it and using its outputs to train a different (often competing) model.
|
Agentic AI | An AI system that can independently plan, make decisions, and carry out a series of actions to complete complex tasks with little human involvement. It can break a task into smaller steps, remember relevant information, and adjust its actions as needed to achieve a specific goal. Some AI systems use a single agent, while others coordinate multiple AI agents working together. An AI agent is a system or program that is capable of nearly autonomously performing specific tasks on behalf of a user or another system.
|
AI Diffusion | Large-scale integration and adoption of AI across existing systems, business operations, and organizations with the goal of improving efficiency, productivity, and innovation.
|
AI Hallucinations | Incorrect, incoherent, or misleading results generated by an AI model, often caused by a variety of factors such as insufficient training data (defined below), incorrect assumptions made by the model that rely on pattern prediction, or biases in the data used to train the model.
AI hallucinations are harmful because they can give users false information that sounds correct, which can lead to mistakes, poor decisions, and misplaced trust in the AI system. The risk is especially serious when AI is used in high-stakes settings such as healthcare, law, finance, customer support, or public policy.
|
AI Inference | The process of using a trained model to generate predictions and perform tasks based on new data beyond those used during the original training phase. To put it simply, while trained models learn more new skills, inference models apply the skill to do a job.
|
AI Slop | Low-quality AI-generated content, often mass-produced with limited concern for accuracy or value. The term can apply to text, images, audio, video, memes, or spam-like posts.
|
AI Training | The process by which a model learns how to perform tasks by exposing it to massive, diverse datasets such as text, image, audio, or video and teaching it to recognize patterns and make predictions. AI training is often the most compute-intensive process of building an AI model, especially for frontier models.
|
Application Program Interface (API) | A software interface that allows different computer programs to exchange data and services, acting as a messenger without knowing the internal functionality of the programs. For example, a company can use an API to connect its website or software to a chatbot model, allowing users to ask questions and receive AI-generated answers inside that service.
|
Base Models | Large AI models that have been trained on broad datasets but have not yet been customized for specific tasks or modified with additional safety features. They often serve as the starting point for creating more specialized AI systems, such as chatbots, coding assistants, or image-generation tools. Base models are also often referred to as foundation models.
|
Benchmarks | Standardized tests used to measure and compare how well AI models perform on specific tasks such as answering questions, writing code, recognizing images, or following instructions.
|
Closed (or Proprietary) Model | A model that is accessible only through the developer’s interface (e.g., ChatGPT) or an approved API. Closed models are often juxtaposed with open models (see definition below) in which developers often publish the model weights and/or allow free downloads and customization.
|
Compute | The processing power used by computers to perform calculations and complete tasks. In AI, compute refers to the computing resources—such as CPUs, GPUs, and other AI chips—used to train and run AI models.
|
Derivative Models | AI models that are modified from base models to better perform specific functions, such as customer service or coding assistance. Some examples of derivative models include Stanford’s Alpaca, Large Model Systems’ (LMSYS) Vicuna, and University of California Berkeley’s Koala, all of which were derived from Meta’s LLaMa model. Another example, Salesforce’s Einstein GPT, was derived from OpenAI’s GPT-3 and GPT-4 and Salesforce’s existing Einstein AI platform to improve customer service for both internal employees and external clients.
|
Embodied AI | AI that can perceive, interact with, and act in the physical world through machines such as robots, vehicles, or drones. Unlike AI systems that operate only in software (such as ChatGPT), embodied AI uses sensors to gather information about its surroundings and can take physical actions in response. Examples of embodied AI include self-driving cars, delivery drones, autonomous delivery robots, and surgical robots assisted by AI.
|
Finetuning | The technical process of adapting a base AI model for specific use cases such as customer service, legal document review, financial forecasting, and more. For example, BloombergGPT is a large language model finetuned on financial data to perform tasks associated with financial analysis.
|
Frontier Models | The most capable AI models available at a given time, achieving top performance across complex tasks like coding, math, and processing multimodal (see definition below) data (e.g., text, images, and audio), often trained with the largest datasets and most compute available.
|
Knowledge Distillation | A machine learning technique to transfer learned knowledge from a “teacher” AI model to a “student” AI model. The technique is not new, having first emerged as a concept in a 2006 academic paper, with the term officially being coined in 2015.
Knowledge distilled could include outputs (i.e., final predictions such as image classification), intermediate layers (i.e., step-by-step problem-solving process), and relationships (i.e., structural layers between different data samples). Knowledge distillation aims to reduce the amount of memory and compute needed to train a model by transferring specialized, diverse, multi-task knowledge from the “teacher” to the “student” model, thereby negating the need for the “student” model to go through the same rigorous amount of training. The “student” model (if left as is), however, usually remains less powerful than the “teacher” model.
|
Large Language Model (LLM) | An AI model that generates text when queried by predicting the next token (see definition below) based on patterns learned from large datasets. This is the type of model most consumers associate with AI. During training of an LLM, input text is split into tokens (see definition below) and encoded so the model can learn and predict possible continuations.
|
Mixture of Experts
| An AI model design that uses only the parts of the model most relevant to a given task or input rather than using the entire model every time. This approach can improve speed and efficiency by reducing the amount of computing power needed while maintaining strong performance.
|
Model Weights | A subset of parameters (see definition below) in a neural network—a brain-inspired technique that teaches models to process data (also defined below)—that determines how models process information, predicts how strongly different inputs influence outputs, and generates outputs.
|
Multimodal | The capacity of an AI model to process and understand multiple types of data ("modalities")—for example, text, still images, audio, and video. A multimodal model can handle more than one type of data on the input or output side or both.
|
Neural Network | A type of AI system made up of interconnected layers that learn patterns and relationships from data. Neural networks—particularly the transformer architecture introduced by Google researchers in 2017—are the foundation of many modern AI systems, including large language models (see definition above) and image-generation models.
|
Open Model
| A model that publicly releases its weights (see definition above). Often referred to as open-weight models, which release only parameters (defined below). Open-weight models also have varying degrees of licensing permissions, which may allow developers to independently understand, audit, recreate, and build upon the model, unlike closed or proprietary models (see definition above).
|
Open-Source Model | A model whose key components—including its code, weights, and training data—are made publicly available so others can use, study, modify, and share it.
|
Parameters | The learned numerical values inside an AI model that determine how it processes inputs (i.e., training data) and generates outputs. Parameters include model weights (see definition above), biases that help models generalize, and other trainable values.
|
Pre-Training | The initial, large-scale stage of training an AI model, in which it learns broad patterns, relationships, and representations from large and diverse datasets. This gives the model general knowledge and capabilities that can later be adapted for many different tasks.
|
Post-Training | The set of training and refinement steps applied after pre-training to make an AI model more useful, reliable, and better aligned with intended behavior. Post-training can describe multiple techniques, including but not limited to finetuning (defined above) and RLFH (defined below).
|
Reinforcement Learning from Human Feedback (RLHF)
| A method of training in which human feedback is used to help an AI model produce better and more preferred outputs.
|
Scaling Laws | The general pattern that AI systems tend to become more capable when they are trained using more computing power, more data, or larger models. Scaling laws help AI developers predict how much an AI system might improve as they give it more resources.
|
Small Language Model | Language models that are trained on fewer parameters and smaller datasets for specialized tasks, requiring less compute and data than LLMs (see definition above). Examples of small language models include Meta’s LLaMa 3.2-1B, Microsoft’s Phi-3.5-Mini-3.8B, and Google’s Gemma3-4B.
|
Software Programming Platforms | The software used for running computing tasks across several AI chips (usually GPUs) at the same time. These platforms help the AI chips handle many calculations in parallel, thereby increasing their speed and efficiency in handling any given task.
|
Tokens | Units of text that AI language models use to process language. A token can be a whole word, part of a word, punctuation mark, or other sequence of characters. AI models read, generate, and count text in tokens rather than words. For example, a word such as “unbelievable” may be broken into smaller components such as “un,” “believe,” and “able.” AI models generate text by predicting likely next tokens in a sequence. Tokens are also commercially and operationally important because AI firms often price services based on token usage, while higher token volumes increase computing demand, energy consumption, and the need for advanced chips and data center capacity.
|
Term | Definition |
Data Curation | The process of collecting, cleaning, selecting, labeling, organizing, and maintaining data so it can be used effectively to train, test, or run AI systems.
|
Data Labeling | The process of adding tags or categories to raw data so an AI model can learn what the data means. Without labeled examples, or if the labels are inconsistent, an AI model’s outputs can be inaccurate.
|
Data Poisoning | The deliberate corruption of training data to cause an AI system to perform poorly or behave incorrectly. It might involve adding false data at the training stage, changing existing data at the inference stage, or biasing the dataset so the model becomes less accurate, less reliable, or easier to exploit.
Data poisoning can cause a model to make incorrect predictions, reveal hidden cybersecurity vulnerabilities, or produce manipulated outcomes. Examples of data poisoning include:
|
Data Provenance | The record that shows the source and history of the data used by a system. It includes information about where the data came from, how it was created, and how it has been changed over time. For AI, data provenance matters because it helps people judge whether data is trustworthy, authentic, and appropriate to use for training a model.
|
Data Sovereignty | The principle that data is subject to the laws and rules of the country where it is stored or controlled. Governments may use data sovereignty policies to require certain data to be stored, processed, or managed within their borders. For example, China’s Cybersecurity Law and Data Security Law require strict data localization.
|
The “Data Wall” | The point at which AI models stop improving much because the available training data (defined below) has been largely used up or no longer adds enough value. To date, AI progress has depended heavily on large amounts of high-quality training data. If that data becomes scarcer, future model improvement could slow unless developers use synthetic data (defined below).
|
Raw Data | Information in its original, unprocessed form before it has been cleaned, labeled, organized, or analyzed for use in an AI system. Raw data can include text, images, audio, video, sensor readings, web pages, spreadsheets, or other source material collected from the real world.
Raw data is the foundation for training and evaluating AI models. However, raw data by itself is usually not ready for use, so it often needs labeling (defined above) and curation (defined above) before a model can learn from it effectively.
|
Synthetic Data | Artificially generated data from AI models and algorithms that is primarily used to train and improve AI systems at scale. The data mimics real-world data so that it can supplement or replace real datasets that are proprietary, difficult to obtain, or becoming finite. The advantages of synthetic data are that it is scalable, cheaper than collecting real-world data, preserves privacy, and can be tailored to specific needs. However, the data often lacks the complexity of real-world data, can amplify existing biases, and risks degrading model performance over time if AI systems are trained repeatedly on AI-generated data.
|
Training Data | The information used to teach an AI model how to recognize patterns, make predictions, or generate outputs. It can include text, images, audio, video, spreadsheets, or other examples the model learns from during training. The quality, quantity, and diversity of training data strongly affect how well an AI system performs.
|
Term | Definition |
AI Ecosystem | Encompasses all elements involved in developing, deploying, and using AI, including technologies, data, talent, governance, funding, and infrastructure (i.e., data centers).
|
AI Infrastructure | The hardware and software needed to develop, run, and manage AI systems. This includes computing chips, data centers, networking equipment, and the software tools that support AI applications and tasks.
AI infrastructure is the underlying foundation that powers the AI tech stack (defined below).
|
AI Tech Stack | The collection of technologies built on top of one another that are necessary to build, train, deploy, and run AI systems. The AI tech stack generally encompasses three interdependent layers: infrastructure, models, and applications.
|
Data Centers | Large facilities housing thousands of specialized chips, servers, and networking systems designed to process, store, and distribute data and handle the massive computing demands of training and running AI models.
|
Hyperscaler | Very large cloud-computing companies that operate massive data centers and provide computing, storage, and networking services at enormous scale. They are built to expand capacity quickly so customers can run large digital services, AI workloads, and other computing tasks without needing to build all the infrastructure themselves.
|
Server Racks | The physical structures (usually metal frames or cabinets) in data centers that house high-performance servers, chips, cables, and cooling equipment.
|