The Shared Shape

Lesson 1 of 3 in Bedrock, Azure AI Foundry, Vertex AI.

Amazon Bedrock, Azure AI Foundry, and Vertex AI were built by different companies, on different infrastructure, with different model partnerships — and they have converged on the same shape. Every one of them now offers a Model catalog, serverless pay-per-Token inference, Provisioned throughput you commit to in advance, Fine-tuning and customization, managed Retrieval-augmented generation (RAG) with Grounding, configurable content safety, and private networking into your own network. Seven capabilities, three brand vocabularies.

The convergence is not a coincidence. Enterprises everywhere asked the same questions — which models can I run, what does a call cost, can I guarantee Throughput, can I teach it my domain, can I ground it in my data, can I stop it saying things, can I keep traffic off the public internet — and the physics of serving large models constrains the possible answers. So the platforms grew the same organs, under different names.

That makes the shape itself the most durable thing you can learn. Product names churn; the seven slots do not. Once you carry the shape in your head, a new platform, a rebrand, or a rival cloud’s pitch all become the same exercise: which slot is this, and what is different about this cell? (One boundary note: each cloud also sells an agent platform — Bedrock AgentCore, Foundry Agent Service, Vertex’s agent stack. Those orchestrate models; our sister AI Agent Academy covers them. This site owns the model-serving layer underneath.)

The shared shape: seven capabilities, three brand vocabularies. Every cell condensed from the platform documentation pages read on 2026-09-16 — the full, sourced version of this table is the Cloud Model Matrix tool.
CapabilityAmazon BedrockAzure AI FoundryVertex AI

Model catalog

Model Catalog + Amazon Bedrock Marketplace

Model catalog (“Foundry Models”)

Model Garden

Serverless inference

On-demand pay-per-token, with Standard / Priority / Flex service tiers

Serverless deployments: standard, provisioned, and batch categories

Standard / Priority / Flex PayGo tiers + Batch

Provisioned capacity

Provisioned Throughput in Model Units (MUs)

Provisioned deployment types in provisioned throughput units (PTUs)

Provisioned Throughput in GSUs

Fine-tuning & customization

Supervised + reinforcement fine-tuning, distillation, Custom Model Import

SFT, DPO, RFT — serverless or managed compute

Supervised, preference, and RL tuning (adapter-based)

Managed RAG / grounding

Knowledge Bases

Azure OpenAI On Your Data (deprecated, retiring 2026)

RAG Engine + grounding sources

Guardrails / content safety

Amazon Bedrock Guardrails

Content filtering via Azure AI Content Safety

Safety and content filters

Private networking

AWS PrivateLink VPC endpoints

Azure Private Link private endpoints

Private Service Connect (PSC)

Two more facets cut across all seven slots and deserve their own columns in any real comparison: regions and cross-region serving (every platform now documents both geography-bounded and globally routed options — Cross-region inference is a first-class mechanism, not an accident) and quotas and scaling (every serverless tier is quota-limited, and every quota system has its own units and increase process). The Cloud Model Matrix tool carries all nine facets with a per-cell source link and verification date — this module teaches you to read it.

Standing up a model workload — the same walk on any platform

  1. Pick a model in the catalog

    Bedrock Model Catalog (plus Bedrock Marketplace), the Foundry model catalog, or Vertex Model Garden. Same slot, three doors.

  2. Baseline traffic predictable enough to commit?

    The provisioned-vs-serverless call is a cost-modeling question — covered in depth in the Cost Modeling module.

  3. Buy provisioned capacity

    MUs on Bedrock, PTUs on Foundry, GSUs on Vertex. Each unit buys a documented processing level for a term — and the terms differ a lot.

  4. Serve serverless, pay per token

    Quota-limited on every platform: token-usage quotas on Bedrock, TPM/RPM per deployment on Azure, per-project-per-region-per-base-model request metrics on Vertex.

  5. Customize if the base model falls short

    Fine-tuning and its relatives. Watch the constraint sentences: e.g., a model customized in Bedrock must run on Provisioned Throughput.

  6. Ground it in your data

    Knowledge Bases (Bedrock), Azure OpenAI On Your Data (deprecated, retiring 2026), RAG Engine (Vertex) — or your own RAG pipeline.

  7. Guard inputs and outputs

    Bedrock Guardrails, Azure AI Content Safety filters, Vertex safety and content filters — different configurability, same slot.

  8. Wire it privately, place it geographically

    PrivateLink / Private Link / Private Service Connect for the network path; geographic vs. global serving mechanisms for where processing happens.

  9. Production endpoint

    A managed endpoint you never patch — that is what you are paying the platform premium for.

Key terms: Model catalog, Managed endpoint, Provisioned throughput, Private endpoint, Cross-region inference, Quota

Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.