The Cloud AI Platforms

The idea in one minute#

Most enterprises do not call model makers directly. They reach models through the cloud they already use, because the contract, the security approval and the private network are already in place. Three platforms cover most of the market: Amazon Bedrock on AWS, Microsoft Foundry on Azure, and the Gemini Enterprise Agent Platform on Google Cloud.

Underneath different names, all three sell the same seven capabilities: a catalogue of models, retrieval over your data, a place to run agents, a way to connect tools, guardrails, evaluation and tracing, and governance of who may use what. Learn the seven once and you can read any vendor’s diagram.

The choice between them is seldom about features. It is decided by where the customer’s data and identity already live, which models they need, and what they have already signed. Everything dated here was checked on 13 October 2026; this is the fastest-changing page in the course.

A picture#

flowchart TB
  subgraph CAP["What every platform provides"]
    direction LR
    C1[":i-layers: <b>Model catalogue</b>"]
    C2[":i-search: <b>Retrieval</b>"]
    C3[":i-bot: <b>Agent runtime</b>"]
    C4[":modelcontextprotocol: <b>Tool gateway</b>"]
    C5[":i-shield-check: <b>Guardrails</b>"]
    C6[":i-flask-conical: <b>Evaluation, tracing</b>"]
    C7[":i-fingerprint: <b>Identity, governance</b>"]
  end
  AWS[":aws: <b>Amazon Bedrock</b><br/><small>+ AgentCore</small>"]
  AZ[":azure: <b>Microsoft Foundry</b><br/><small>+ Copilot Studio</small>"]
  GCP[":googlecloud: <b>Gemini Enterprise<br/>Agent Platform</b><br/><small>formerly Vertex AI</small>"]
  CAP --> AWS
  CAP --> AZ
  CAP --> GCP
  AWS --> ENT[":i-building-2: <b>The enterprise's landing zone</b><br/><small>identity, network, data, policy</small>"]
  AZ --> ENT
  GCP --> ENT
  class C1,C2,C3,C6 compute
  class C4 io
  class C5,C7 warn
  class AWS,AZ,GCP queue
  class ENT memory

How it really works#

The seven capabilities, named on each platform#

The names in this table are the ones to recognise in a customer conversation, not to memorise. Renaming is constant: Azure AI Foundry became Microsoft Foundry at the start of 2026, and Google announced at Cloud Next in April 2026 that Vertex AI was becoming part of the Gemini Enterprise Agent Platform. The capability rows did not move.

What each is known for#

All three run agents built with open frameworks and speak MCP, so an agent’s tools and much of its code can move. What does not move easily is identity, networking, data location and the commercial agreement — which is why those decide the choice.

Five other places AI is bought#

The cloud platforms are not the whole market. A complete options list includes:

How to choose#

Work down this list and stop at the first answer that decides it.

1. Where is the data, and may it move? AI goes to the data, not the reverse.
2. Where is identity? The platform that already knows the users and their permissions saves the hardest integration.
3. Which models are required? If a specific model is a hard requirement, check which platforms carry it — and in which regions.
4. What is already contracted? Committed cloud spend and existing data-processing agreements often outweigh any feature.
5. What can the team operate? Skills in the building beat capabilities on a slide.
6. Only then: features. Compare the seven capabilities for the gaps that matter to this use case.

Three ways to pay for model capacity#

Every platform offers the same three, under different names.

Deployment location is a second axis. A global deployment lets the provider route to any region — cheapest and most available. A regional or data-zone deployment keeps processing inside a boundary and usually costs more or has lower limits. That is a residency decision before it is a cost decision; see Capacity, Quotas and Resilience.

A neutral way to compare in front of a customer#

Use a one-page grid and fill it in together. It keeps the conversation on requirements.

Requirement                      Weight   Platform A   Platform B   Evidence
Data stays in required region      must        yes          yes     region list, date checked
Required model available           must        yes           no     catalogue, region
Works with existing identity       high         5            3      who manages SSO today
Existing commercial commitment     high         4            2      procurement
Team can operate it                high         4            3      skills survey
Retrieval over our sources         med          4            4      prototype
Agent runtime with approvals       med          3            4      prototype
Evaluation and tracing             med          4            4      prototype

“Must” rows are pass or fail. Only then are the scores added. This is the same constraints-first logic as Choosing a Model, one level up.

Remember this#

• Three platforms: Amazon Bedrock, Microsoft Foundry, Gemini Enterprise Agent Platform.
• Seven capabilities on each: models, retrieval, agent runtime, tool gateway, guardrails, evaluation, governance.
• Data, identity, required models and existing contracts decide. Features come last.
• On-demand, provisioned, batch: three ways to pay on every platform.
• Names change yearly. Learn the capability, look up the name.

Try it#

1. Pick one platform. In its documentation, find the current name of each of the seven capabilities and the date the page was last updated.
2. Fill in the comparison grid for a real or imagined customer whose email and identity are with one vendor and whose data warehouse is with another. Which rows decide it?
3. Find the regions in which one specific model is offered on two different platforms. Are they the same?

Check yourself#

1. Name the seven capabilities every platform provides.
2. Why do data location and identity usually decide the platform before features do?
3. When is provisioned capacity the right way to pay, and when is batch?
4. A customer’s tools are exposed over MCP. What does that make easier if they change platform, and what stays hard?

Sources#

• Amazon Bedrock AgentCore
• What is Microsoft Foundry? (page dated 13 August 2026)
• Gemini Enterprise Agent Platform overview
• Microsoft Cloud Adoption Framework: AI strategy — the four adoption models