BIG PICTURE The problem: everyone seems to do everything
Anthropic, OpenAI, Google, Databricks, Azure, Pinecone, LangChain, MCP… the names blur together because each one lives on a different layer of the same stack — and some sit on several layers at once. The trick to seeing clearly is to stop asking "what is this company?" and start asking "which layer is this thing on?". There are really only six layers, plus a set of cross-cutting concerns.
DIAGRAM The six-layer stack (with real players)
Read it top-down: your app calls the orchestration layer, which assembles context and calls a model through an API/gateway; the model was trained by a lab and runs on cloud compute.
LAYERS The stack in words
KEY IDEA One player, many hats
The confusion comes from companies that operate on several layers at once. That's normal — it's vertical integration. Here's how the big names spread across the stack:
🟣 Anthropic (a lab, integrated upward)
- Layer 2 — the lab that trains the models
- Layer 3 — Claude is the model family
- Layer 4 — the Anthropic API (Messages API)
- Layer 5 — MCP (the protocol) + the Agent SDK
- Layer 6 — Claude Code & the Claude apps
🔵 Microsoft / Azure (a cloud, integrated downward)
- Layer 1 — Azure cloud + GPUs
- Layer 4 — Azure AI Foundry hosts many models (OpenAI, Anthropic, Meta, …)
- Layer 5/6 — Copilot Studio, GitHub Copilot
- Note: Microsoft doesn't (mostly) train frontier models — it hosts and packages others'
WHO'S WHO The players, decoded
| Player | Layer(s) | What it actually provides | One-line analogy |
|---|---|---|---|
| Anthropic | 2·3·4·5·6 | Trains Claude; runs the Anthropic API; created MCP; ships Claude Code | A car maker that also runs its own dealerships |
| Claude | 3 | The LLM family (Opus / Sonnet / Haiku) — the "brain" | The engine, not the car |
| OpenAI · GPT | 2·3·4 | Lab + GPT models + the OpenAI API | Anthropic's closest analogue |
| Google · Gemini | 1·2·3·4·6 | Cloud (GCP), lab (DeepMind), Gemini models, Vertex AI, the Gemini app | Full stack, top to bottom |
| Meta · Llama | 2·3 | Trains open-weight Llama models you can self-host | Open-source engine you download |
| Mistral | 2·3 | European lab; open + commercial models | Lean challenger lab |
| AWS Bedrock | 1·4 | A managed gateway to many models (incl. Claude) on AWS | An app store for models |
| Azure AI Foundry | 1·4 | Microsoft's model platform — host, deploy & govern many models | An app store for models (MS edition) |
| Google Vertex AI | 1·4 | GCP's managed model platform (Gemini + partners incl. Claude) | An app store for models (Google edition) |
| Databricks | 1·4·5 | Data + AI platform: your data lakehouse, model serving, RAG, governance | Where your data and the models meet |
| Snowflake | 1·4·5 | Data cloud with hosted models + RAG (Cortex) | Databricks' data-cloud rival |
| Pinecone · Weaviate · pgvector · Milvus | 5 | Vector databases that store & search embeddings for RAG | The searchable long-term memory |
| Voyage AI · Cohere Embed | 3 | Embedding models that turn text into vectors | The translator: words → numbers |
| LangChain · LlamaIndex | 5 | Frameworks that wire prompts, RAG, tools & agents together | The wiring harness |
| MCP (Model Context Protocol) | 5 | An open standard for connecting models to tools & data — not a company | The USB-C port for AI |
| Hugging Face | 3·5 | A hub to find, share & run open models and datasets | GitHub for models |
| NVIDIA | 1 | The GPUs that train and run nearly everything | The picks-and-shovels supplier |
DIAGRAM How one request actually flows
Put the layers in motion. A user asks a question in an app; here's the journey and where each buzzword lives:
(layer 6)"] APP --> ORCH["Orchestrator / Agent
(layer 5)"] ORCH -->|"embed the query"| EMB["Embedding model"] EMB --> VDB[("Vector store
RAG retrieve")] VDB -->|"relevant chunks"| ORCH ORCH -->|"check"| CACHE["Prompt cache"] ORCH -->|"recall"| MEM["Memory"] ORCH -->|"assembled prompt + tools"| API["Model API / Gateway
(layer 4)"] API --> LLM["🧠 LLM · inference
(layer 3)"] LLM -->|"wants a tool"| MCP["Tools via MCP"] MCP --> LLM LLM -->|"answer"| API --> ORCH --> APP --> U classDef s fill:#eef6ff,stroke:#3b7bd0,color:#1f5fa8; class APP,ORCH,API,LLM,MCP s;
- Prompting — the app + orchestrator write the instructions and assemble context.
- RAG — the query is turned into an embedding, matched against a vector store, and the top chunks are pulled into the prompt so the model answers from your data.
- Caching — a stable prompt prefix is reused to cut cost & latency; memory carries facts across turns/sessions.
- Inference — the LLM runs (through an API/gateway) and generates a response.
- Tools / MCP — if it needs live data or actions, it calls tools, often over MCP, then continues.
GLOSSARY The acronyms, fast
| Term | In one sentence |
|---|---|
| LLM | Large Language Model — a neural net trained to predict the next token; the "brain." |
| Foundation model | A large, general model trained on broad data, adaptable to many tasks (LLMs are one kind). |
| Token | A chunk of text (~¾ of a word) — the unit models read, generate, and are billed on. |
| Context window | The max tokens a model can consider at once (input + output). |
| Inference | Running a trained model to produce output (vs. training, which creates it). |
| Embedding | A vector of numbers representing meaning, so similar text sits close together. |
| Vector store | A database that stores embeddings and finds the most similar ones fast. |
| RAG | Retrieval-Augmented Generation — fetch relevant data and feed it to the model at query time. |
| Prompt | The instructions + context you send the model to steer its output. |
| Fine-tuning | Further-training a model on your data to change its behavior (heavier than prompting/RAG). |
| Agent | An LLM that decides and acts in a loop, calling tools until a task is done. |
| MCP | Model Context Protocol — an open standard for plugging tools & data into any model/app. |
| Tool / function calling | Letting the model invoke your code (search, DB, API) and use the result. |
| Prompt caching | Reusing a processed prompt prefix to cut cost & latency on repeated requests. |