Gen AI Fundamentals › The AI Universe

🌌 The AI Universe — Who Does What

One map of the whole landscape · concepts + the real players that fill each slot

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.

🔑 Mental model: a Gen AI system is a stack. At the bottom is raw compute; at the top is the app a human uses. Every product name you've heard slots into one (or more) of these layers.

DIAGRAM The six-layer stack (with real players)

flowchart TB subgraph L6["6 · APPLICATIONS (what a human uses)"] A1["Claude Code · Cursor"]:::real A2["ChatGPT · Copilot"]:::real A3["Your custom app / agent"]:::real end subgraph L5["5 · ORCHESTRATION & CONTEXT (the glue)"] O1["Prompting"]:::c O2["RAG + Vector Store"]:::c O3["Memory · Caching"]:::c O4["Agents · Tools · MCP"]:::c end subgraph L4["4 · MODEL ACCESS (APIs & gateways)"] G1["Anthropic API"]:::real G2["AWS Bedrock"]:::real G3["Azure AI Foundry"]:::real G4["Google Vertex AI"]:::real end subgraph L3["3 · FOUNDATION MODELS (the LLMs)"] M1["Claude"]:::real M2["GPT · Gemini · Llama"]:::real M3["Embedding models"]:::c end subgraph L2["2 · MODEL PROVIDERS / LABS (who trains them)"] P1["Anthropic · OpenAI"]:::real P2["Google · Meta · Mistral"]:::real end subgraph L1["1 · COMPUTE & CLOUD (where it all runs)"] C1["AWS · Azure · GCP"]:::real C2["Databricks · NVIDIA GPUs"]:::real end L6 --> L5 --> L4 --> L3 L2 -.trains.-> L3 L1 -.hosts.-> L3 L1 -.hosts.-> L4 classDef c fill:#f3f0ff,stroke:#764ba2,color:#4a3b7a; classDef real fill:#eef6ff,stroke:#3b7bd0,color:#1f5fa8;

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

6 · Applicationswhat a human touches
chatbotscopilotsagents Claude CodeCursorChatGPTGitHub Copilot
5 · Orchestration & Contextthe glue / "AI engineering"
promptingRAGvector storememorycachingtool useMCP LangChainLlamaIndexClaude Agent SDKPinecone
4 · Model AccessAPIs & cloud gateways
REST APIsgateways Anthropic APIAWS BedrockAzure AI FoundryGoogle Vertex AIDatabricks
3 · Foundation Modelsthe actual model weights
LLMsembedding modelsmultimodal ClaudeGPTGeminiLlamaMistralVoyage (embeddings)
2 · Providers / Labswho trains the models
frontier labs AnthropicOpenAIGoogle DeepMindMetaMistralxAI
1 · Compute & Cloudwhere it physically runs
GPUs / TPUshyperscalersdata platforms AWSAzureGCPDatabricksNVIDIA

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'
So "Claude" (a model, layer 3) can be reached through the Anthropic API (layer 4) or through AWS Bedrock, Google Vertex AI, Azure AI Foundry, or Databricks — all layer-4 gateways. Same model, different front door. That's why the same name shows up in many places.

WHO'S WHO The players, decoded

PlayerLayer(s)What it actually providesOne-line analogy
Anthropic2·3·4·5·6Trains Claude; runs the Anthropic API; created MCP; ships Claude CodeA car maker that also runs its own dealerships
Claude3The LLM family (Opus / Sonnet / Haiku) — the "brain"The engine, not the car
OpenAI · GPT2·3·4Lab + GPT models + the OpenAI APIAnthropic's closest analogue
Google · Gemini1·2·3·4·6Cloud (GCP), lab (DeepMind), Gemini models, Vertex AI, the Gemini appFull stack, top to bottom
Meta · Llama2·3Trains open-weight Llama models you can self-hostOpen-source engine you download
Mistral2·3European lab; open + commercial modelsLean challenger lab
AWS Bedrock1·4A managed gateway to many models (incl. Claude) on AWSAn app store for models
Azure AI Foundry1·4Microsoft's model platform — host, deploy & govern many modelsAn app store for models (MS edition)
Google Vertex AI1·4GCP's managed model platform (Gemini + partners incl. Claude)An app store for models (Google edition)
Databricks1·4·5Data + AI platform: your data lakehouse, model serving, RAG, governanceWhere your data and the models meet
Snowflake1·4·5Data cloud with hosted models + RAG (Cortex)Databricks' data-cloud rival
Pinecone · Weaviate · pgvector · Milvus5Vector databases that store & search embeddings for RAGThe searchable long-term memory
Voyage AI · Cohere Embed3Embedding models that turn text into vectorsThe translator: words → numbers
LangChain · LlamaIndex5Frameworks that wire prompts, RAG, tools & agents togetherThe wiring harness
MCP (Model Context Protocol)5An open standard for connecting models to tools & data — not a companyThe USB-C port for AI
Hugging Face3·5A hub to find, share & run open models and datasetsGitHub for models
NVIDIA1The GPUs that train and run nearly everythingThe 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:

flowchart LR U["👤 User"] --> APP["App / UI
(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;
  1. Prompting — the app + orchestrator write the instructions and assemble context.
  2. 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.
  3. Caching — a stable prompt prefix is reused to cut cost & latency; memory carries facts across turns/sessions.
  4. Inference — the LLM runs (through an API/gateway) and generates a response.
  5. Tools / MCP — if it needs live data or actions, it calls tools, often over MCP, then continues.

GLOSSARY The acronyms, fast

TermIn one sentence
LLMLarge Language Model — a neural net trained to predict the next token; the "brain."
Foundation modelA large, general model trained on broad data, adaptable to many tasks (LLMs are one kind).
TokenA chunk of text (~¾ of a word) — the unit models read, generate, and are billed on.
Context windowThe max tokens a model can consider at once (input + output).
InferenceRunning a trained model to produce output (vs. training, which creates it).
EmbeddingA vector of numbers representing meaning, so similar text sits close together.
Vector storeA database that stores embeddings and finds the most similar ones fast.
RAGRetrieval-Augmented Generation — fetch relevant data and feed it to the model at query time.
PromptThe instructions + context you send the model to steer its output.
Fine-tuningFurther-training a model on your data to change its behavior (heavier than prompting/RAG).
AgentAn LLM that decides and acts in a loop, calling tools until a task is done.
MCPModel Context Protocol — an open standard for plugging tools & data into any model/app.
Tool / function callingLetting the model invoke your code (search, DB, API) and use the result.
Prompt cachingReusing a processed prompt prefix to cut cost & latency on repeated requests.
🔧 Want the products, not just the layers? Two companion pages go tool-by-tool through the cross-cutting concerns this map only names: Observability & Eval — What Each Tool Offers (capability matrices for open source, commercial and every cloud provider) and AI Gateway — The Control Point (the layer-4 front door, compared properly).
➡️ Next: see how these pieces assemble into a real product on the Anatomy of a Gen AI Application page — then you're ready for the CCA-F study material.