| LLM (Large Language Model) | An AI system trained on vast amounts of text that can understand and generate human language; the engine behind tools like ChatGPT, Claude, and Gemini. |
| Foundation Model | A very large AI model trained on broad data that can be adapted to many different tasks — like a generalist consultant who can then specialize in any field. |
| AI Agent | Software that uses an AI model to take real actions on your behalf — browsing the web, sending emails, running calculations — and makes its own decisions about what to do next, rather than just answering questions. |
| Agentic | Describing AI behavior that takes autonomous actions, not just generates text. An “agentic workflow” is one where AI handles multiple steps in sequence without a human approving each one. |
| Chatbot | An AI that can hold a conversation and remember what you said earlier, but can only produce text — it cannot take actions in the world or do things on your behalf. |
| RAG (Retrieval-Augmented Generation) | A technique where the AI looks up relevant documents in real time before answering, rather than relying solely on what it learned during training — like giving the AI a reference library to consult before responding. |
| Context Window | How much information the AI can hold in mind at once — like working memory. A larger context window means the AI can read and reason over longer documents or longer conversations in one go. Measured in tokens. |
| Token | The basic unit of text an AI processes — roughly three-quarters of a word. “The quick brown fox” is about five tokens. Pricing, speed, and context limits are all measured in tokens. |
| Prompt | The instructions or question you give to an AI to shape its response. Writing prompts skillfully — “prompt engineering” — can dramatically change the quality of AI output. |
| Hallucination | When an AI confidently states something that is factually wrong. A known limitation — AI can “make things up” that sound plausible, which is especially dangerous when agents take real actions based on those invented facts. |
| Grounding | Connecting an AI’s responses to verified, real data sources to reduce hallucinations. A “grounded” AI answer is backed by actual documents or databases, not just the model’s memory. |
| Inference | The act of running an AI model to get a response. When you send a question to ChatGPT and it replies, that is inference. Distinct from training, which is teaching the model — inference is using it. |
| Training | The process of teaching an AI model by exposing it to enormous amounts of data. Done once (or periodically) by AI labs; extremely expensive. Using the trained model afterward is called inference. |
| Fine-tuning | Taking a general-purpose AI model and further training it on specific data to make it better at a particular task — like hiring a generalist and then giving them specialized on-the-job training in your industry. |
| Reasoning Model | An AI model that “thinks” for several seconds or minutes before answering — working through a problem step by step internally before giving a response. Newer models like GPT-o1 and Claude 3.7 do this, enabling more complex multi-step work. |
| Multimodal | An AI that can work with multiple types of input — text, images, audio, and video — not just written words. |
| Embedding | A way of converting text into numbers so that similar concepts end up mathematically close together, allowing AI to search by meaning rather than exact keyword matches. |
| Vector Database | A specialized database that stores embeddings and enables fast search by meaning rather than keywords. Powers the retrieval step in RAG systems — think of it as a filing system organized by concept rather than alphabetically. |
| Semantic Search | Search that finds results by meaning, not exact word match — “best way to cut costs” would return results about “cost optimization” even if those exact words weren’t used. |
| Orchestration | Coordinating multiple AI agents or steps in a workflow so they work together in the right order. The “project manager” layer that decides which AI does what, when, and how to handle results. |
| Orchestrator | The coordinating component in a multi-agent system that delegates work to specialized sub-agents, collects their results, and synthesizes a final answer — like a senior analyst directing a team. |
| Tool Use / Function Calling | The ability for an AI to call external tools — looking up a stock price, running a calculation, querying a database, sending an email — rather than just generating text. This is what makes agents different from chatbots. |
| MCP (Model Context Protocol) | A universal standard that lets AI agents connect to external tools and data sources in a consistent way — like a universal USB standard for AI integrations, so any agent can plug into any tool without custom wiring. |
| ReAct Pattern | The standard loop most AI agents follow: Reason about what to do next, Act by calling a tool, then Observe the result — and repeat until the task is done. Named from “Reason + Act.” |
| Multi-Agent System | A setup where multiple specialized AI agents work in parallel or in sequence — one researches, one writes, one reviews — coordinated by an orchestrator. Can be faster but also more complex and harder to control. |
| Prompt Caching | A technique that saves frequently reused parts of a prompt so they don’t need to be reprocessed every time, cutting costs by ~90% and response time by ~75% on repeated context. |
| Context Engineering | The practice of carefully selecting and structuring what information goes into an AI’s context window to maximize usefulness and minimize wasted tokens. Like briefing a consultant with exactly the right documents before a meeting. |
| API (Application Programming Interface) | A standardized way for software systems to communicate with each other. When a company “accesses AI via API,” they’re calling an AI service programmatically — like a bank wire rather than going in person. |
| SDK (Software Development Kit) | A toolkit of pre-built code components that developers use to build applications on top of a platform, like LEGO bricks for software. An “AI SDK” is a toolkit for building AI-powered products. |
| LangChain / LangGraph | The most widely used open-source toolkit for building AI agents. LangChain chains together AI calls; LangGraph adds a structured runtime for more complex, multi-step agent workflows. |
| CrewAI | An agent-building framework where you define AI agents by role — “researcher,” “writer,” “reviewer” — and they collaborate, like assigning roles to a team. Lower technical barrier than most alternatives. |
| LLMOps | Short for “Large Language Model Operations” — the engineering discipline of deploying, monitoring, evaluating, and maintaining AI models in production, analogous to how software teams manage and monitor applications. |
| Eval / Evals | Short for “evaluations” — tests that measure how well an AI agent performs its job. Unlike traditional software tests with right/wrong answers, evals for AI are often scored by another AI model, making them expensive and imprecise. |
| Observability | The ability to see inside a running AI system — tracking what it did, why, and where it went wrong. In agentic AI, this means logging every reasoning step and tool call so failures can be diagnosed. |
| Hyperscaler | The three dominant cloud providers — Microsoft Azure, Amazon AWS, and Google Cloud — that own the massive data center infrastructure most AI runs on. When they spend $600B on AI infrastructure, that is the market betting on how big AI gets. |
| GPU (Graphics Processing Unit) | The specialized chip that AI training and inference runs on. Originally designed for video games, now essential for AI computation. Extremely expensive and in high demand; NVIDIA makes the dominant ones. |
| TPU (Tensor Processing Unit) | Google’s custom-built chip for AI computation, designed specifically for AI workloads rather than adapted from gaming chips. Gives Google a cost and speed advantage — they are the only major AI lab with their own silicon at scale. |
| LPU (Language Processing Unit) | Groq’s custom chip designed specifically for fast AI inference (generating responses), distinct from GPUs. Claims much faster response times for certain AI tasks. |
| Wafer-Scale | A chip design approach where the entire silicon wafer becomes a single massive chip, rather than cutting it into smaller individual chips. Cerebras uses this for extremely fast AI computation — think one giant processor instead of thousands of small ones. |
| Compute | Shorthand for the GPUs, servers, and data center capacity needed to run AI. “Compute costs” means the cost of running AI workloads. Currently the single biggest bottleneck and capital expenditure in AI. |
| Capex (Capital Expenditure) | Money spent by a company on long-lived physical assets — in AI, primarily data centers, chips, and servers. When hyperscalers announce $200B capex, that is money committed to building AI infrastructure. |
| Inference Cost | The cost of actually running an AI model to get a response. At scale, this dominates total AI spending. Has fallen roughly 95% since 2023 as hardware and software improved. |
| Latency | How fast the AI responds. Low latency means fast responses; high latency means slow. Critical for real-time applications — voice agents need sub-one-second responses to feel natural. |
| Serverless | A cloud computing model where you pay only for the exact computation you use, with no need to manage or reserve servers. The provider scales capacity up and down automatically. Think “pay per use” vs. “pay to reserve a parking spot.” |
| On-Premise (On-Prem) | Running AI on servers the company owns and operates, rather than using cloud services. More control and privacy, but much higher upfront cost and maintenance burden. |
| Open-Weights Model | An AI model whose internal parameters are publicly released — anyone can download and run it without paying the developer. Llama (Meta) and Mistral are the leading examples. Contrasted with closed models like GPT-4 or Claude. |
| Headless Browser | A web browser that runs in a computer without a visible screen — used by AI agents to navigate websites, fill forms, and click buttons programmatically, the same way a human would but invisible and automated. |
| Code Execution Sandbox | A secure, isolated environment where an AI agent can write and run code safely, without being able to damage the host computer or access unauthorized data. Like a quarantine room for AI-generated programs. |
| MoE (Mixture of Experts) | An AI architecture where only a fraction of the model’s components activate for any given task — like a firm where different specialists handle different questions rather than everyone working on everything. More efficient than using the whole model every time. |
| Diffusion LLM | A newer type of AI model architecture that generates text in parallel rather than one word at a time, achieving much faster speeds (600+ tokens per second vs. typical 50–100). Mercury 2 is an example. |
| SWE-Bench | An industry benchmark that measures how well AI coding agents can solve real software engineering problems — fixing bugs in actual open-source code repositories. A score of 80%+ means the AI can handle the majority of typical software engineering tasks. |
| Anthropic | The AI safety company that built Claude. Founded 2021 by former OpenAI researchers; principal Amazon partner. Now the highest-revenue frontier AI lab as of April 2026. |
| OpenAI | The company behind ChatGPT and GPT-4/5. Backed by Microsoft. The most widely recognized AI brand; 900M weekly active ChatGPT users as of 2026. |
| xAI | Elon Musk’s AI lab; maker of the Grok model. Merged with X Corp (Twitter) in early 2026. Access to X’s real-time data is a key strategic differentiator. |
| Google DeepMind | Google’s consolidated AI research and product unit; maker of the Gemini model family. The only major AI lab that designs its own chips (TPUs). Not directly investable — exposure is through Alphabet (GOOGL). |
| Meta | The company behind Facebook, Instagram, and WhatsApp; also the dominant publisher of open-weights AI models (Llama family). Shifting toward closed models in 2026 with “Muse Spark.” |
| Mistral AI | French AI company producing both open and commercial models; primary investment thesis as a European sovereign-AI alternative to US labs. Backed by ASML. |
| Claude | Anthropic’s AI model family (Haiku, Sonnet, Opus). Known for strong coding performance and safety focus; the default model in many AI coding tools in 2026. |
| Gemini | Google’s family of AI models, integrated into Google Workspace (Gmail, Docs, etc.) and Google Cloud. Reached 750M monthly active users via AI Overviews in Google Search. |
| Copilot | Microsoft’s AI assistant brand, integrated into Office 365, GitHub (for coding), and Windows. Powered by OpenAI models. Over 15 million enterprise seats sold. |
| Copilot Studio | Microsoft’s platform for building custom AI agents that integrate with Microsoft 365 data and workflows. Used by 230,000+ organizations as of 2026 — one of the largest enterprise agent deployments. |
| Agentforce (Salesforce) | Salesforce’s AI agent platform embedded in its CRM products. Hit $800M ARR by early 2026 — the fastest product launch in Salesforce’s 26-year history. Represents the “incumbent bundling” threat to standalone AI startups. |
| Vertex AI | Google Cloud’s platform for building and deploying AI models and agents, competing with Microsoft Azure’s AI services. |
| Databricks | A data and AI platform that helps enterprises store, process, and build AI on top of their own data. Not an AI model lab — more like the infrastructure that enterprise data teams use. Private, valued at $134B in 2026. |
| CoreWeave (CRWV) | A GPU cloud company that rents AI compute capacity to frontier labs and enterprises. The only public-market pure-play on AI infrastructure (besides NVIDIA). Largest customer is Microsoft. |
| Cerebras Systems | An AI chip company that builds “wafer-scale” processors — uniquely large chips for very fast AI computation. Filed for IPO in April 2026; has a $20B+ compute contract with OpenAI. |
| Pinecone | The first major vector database company; powered many early RAG applications. Losing market share to newer architectures; reportedly exploring a sale at $2B+ (2025). |
| Data Flywheel | A self-reinforcing cycle where more customers generate more data, which improves the AI product, which attracts more customers. Companies with proprietary data flywheels are much harder to copy than those relying on generic AI models. |
| Knowledge Graph | A structured map of relationships between concepts, documents, or data points in a company. Glean builds these across enterprise SaaS tools to enable AI search that understands organizational context. |
| Ambient Scribe | An AI system that listens to a conversation (such as a doctor-patient appointment) and automatically generates structured written documentation without anyone typing. Eliminates a major administrative burden for clinicians. |
| CLM (Contract Lifecycle Management) | Software that manages legal contracts from creation through signing, renewal, and expiration. Ironclad is the leading AI-native CLM platform. |
| Prior Auth (Prior Authorization) | A healthcare administrative process requiring insurers to approve certain treatments before they are delivered. Highly paper-intensive; AI companies like Tennr automate the processing of these requests. |
| Outcome-Based Pricing | A pricing model where the customer pays per successful result — per resolved customer service ticket, per legal case won — rather than per seat or per month. Changes the economic relationship between AI vendor and buyer. |
| Services-as-Software | The thesis (popularized by a16z) that AI agents will capture revenue from human services industries (consulting, legal, accounting) not just from software budgets — potentially a 10x larger market than traditional SaaS. |
| Per-Seat SaaS | The traditional software pricing model: a fixed monthly fee per user who accesses the software. Being disrupted in AI by outcome-based pricing, where charges are tied to results rather than access. |
| BDR / SDR (Business/Sales Development Representative) | Entry-level salespeople whose job is to find and qualify new leads before passing them to senior sales staff. “AI BDR” companies sell AI agents that attempt to automate this role. |
| GTM (Go-to-Market) | The strategy and activities a company uses to reach its target customers and generate revenue — sales, marketing, partnerships, pricing. “GTM agents” automate parts of this process. |
| ITSM (IT Service Management) | The processes organizations use to design, deliver, manage, and improve their IT services — helpdesk ticketing, incident management, etc. ServiceNow is the dominant ITSM platform. |
| HRIS (Human Resources Information System) | Software that manages employee data, payroll, benefits, performance reviews, and other HR processes. Workday is the dominant HRIS platform. |
| CRM (Customer Relationship Management) | Software that tracks a company’s interactions with customers and prospects — contact records, sales pipeline, deal history. Salesforce is the market leader. |
| Acqui-hire | When a large company “acquires” a startup primarily to recruit its talented employees, often shutting down the startup’s product. The startup’s investors may get their money back, but there is no upside beyond return of capital. |
| Reverse-Acquihire | A deal where a large company hires away the founders and key staff of a startup — without formally acquiring it — leaving investors with a much smaller residual business. Adept (absorbed by Amazon) and Inflection (absorbed by Microsoft) are recent examples. |
| Secondary Market | The market for buying and selling shares in private companies from existing shareholders (employees, early investors), rather than directly from the company. Active AI secondary market volume hit $103B in H1 2025. |
| SPV (Special Purpose Vehicle) | A legal entity created for a single investment — often used to pool multiple investors’ capital into one position in a private company. Common mechanism for family office AI exposure. |
| LP (Limited Partner) | An investor in a venture capital fund who provides capital but does not make investment decisions. As an LP, you get access to the fund’s portfolio companies and sometimes direct co-investment opportunities. |
| Co-investment | When an LP invests directly alongside a VC fund in a specific company, in addition to their fund commitment. Often at lower fees and better terms than the fund itself. |
| Crossover Fund | An investment fund (like Coatue, Tiger Global) that invests in both private companies and public stocks. They can follow a company from late-stage private all the way through its IPO. |
| EU AI Act | The European Union’s comprehensive regulation of AI systems, phasing in from 2025 through 2027. Classifies AI systems by risk level and imposes compliance requirements — relevant for any AI company selling into the EU. |
| Alignment | Ensuring AI systems behave in ways consistent with human values and intentions. A core research area — particularly important as AI becomes more autonomous and capable of taking consequential actions. |
| Guardrails | Controls placed on an AI system to prevent harmful, inappropriate, or off-topic outputs. Can be built into the model itself or applied as a filtering layer around it. |
| AI-Native | A product or company built from the ground up around AI capabilities, as opposed to having AI added on top of an existing non-AI system. Harvey (legal AI) is AI-native; Microsoft Word with Copilot added is not. |
| Vertical AI | AI products built for a specific industry — legal AI, medical AI, finance AI — rather than general-purpose tools. Often more defensible due to specialized data and deep workflow integration. |
| Horizontal AI | AI products built for one function used across many industries — such as coding assistants, customer support agents, or enterprise search — rather than targeting a specific sector. |
| Open Source | Software whose underlying code is publicly available for anyone to use, modify, and distribute. Open-source AI models (like Llama) can be run without paying the developer, setting a price floor that commercial models must compete against. |
| Free Cash Flow (FCF) | Cash generated by a business after accounting for capital expenditures — a measure of how much real cash a company produces. Amazon expects $17B negative FCF in 2026 due to AI infrastructure spending. |