How clouds serve frontier models

Frontier (or foundation) models are the large general-purpose models teams build products on. Cloud providers don’t usually “own” every model brand — they often host and sell access to several, behind one account, bill, and permission system.

One mental model (not a catalog)

Cloud neighborhood	Rough idea
AWS Bedrock	Managed API to many foundation models on AWS — details
Azure OpenAI (and related Azure AI)	Microsoft Azure’s managed access to certain OpenAI (and other) models
Google Vertex AI	Google Cloud’s managed ML / generative model platform

Names and catalogs change. The durable idea: your app talks to a provider; the provider routes to a model.

Anthropic’s Claude is a model/product line. It may appear in Anthropic’s own apps or as one option inside a cloud catalog. Claude is not “owned by” Bedrock.

How you get access

1. Console / chat UIs — Click around in a vendor console or consumer chat. Great for learning; less for production apps.
2. API keys + SDK — Your code holds a key (or secret from a vault) and calls an HTTP API. Easy to start; protect keys carefully.
3. Cloud IAM roles — In AWS/Azure/GCP, permissions say which identity may invoke which model. No long-lived key in the app if designed well — think “badge access,” not a password taped to a laptop.

You do not need an account to finish this lesson.

Agentic apps on top

Solid arrows = request path (numbered). Dashed teal path = agentic loop back to the app — same idea as the agentic workflow lesson.

An agentic workflow adds the loop: the app plans, calls tools, observes results, and may call the model again. The cloud provider is usually just the hop that runs the model; your agent logic still lives in your app (or in a coding agent like Claude Code).

Takeaway

Clouds serve frontier models through APIs and permissions. Agentic apps = your loop + provider call + model (+ tools). Learn the map before the console.