Amazon Bedrock, Google Vertex AI and Azure AI Foundry

Comparing the three managed GenAI platforms on models, RAG, agents, guardrails, evaluation, fine-tuning, security and enterprise integration.

[!VERSION] These platforms change monthly. Feature names below reflect late 2025 to 2026; always confirm in current documentation. Microsoft has been consolidating branding under Microsoft Foundry.

Capability Amazon Bedrock Google Vertex AI Azure AI Foundry
First-party models Amazon Nova family Gemini family, Imagen, Veo (Microsoft Phi small models)
Third-party/open models Anthropic Claude, Meta Llama, Mistral, Cohere, others Model Garden: Claude, Llama, Mistral, Gemma and more OpenAI models (Azure OpenAI), plus catalog of Meta, Mistral, DeepSeek, Cohere, xAI and others; Anthropic Claude models announced for Foundry in late 2025
Managed RAG Knowledge Bases RAG Engine, Vertex AI Search Azure AI Search integration, Foundry knowledge features
Agents Agents for Bedrock, AgentCore (runtime, memory, gateway, identity, observability) Agent Builder, Agent Engine, Agent Development Kit Foundry Agent Service
Guardrails Bedrock Guardrails (content filters, denied topics, PII, contextual grounding checks) Model Armor, safety settings Azure AI Content Safety, Prompt Shields, groundedness detection
Evaluation Model and RAG evaluation Gen AI evaluation service Foundry evaluations
Fine-tuning For selected models Supervised tuning for Gemini and open models For selected OpenAI and open models
Capacity On-demand, batch, provisioned throughput Pay-as-you-go, provisioned throughput Standard, global/data-zone deployments, provisioned throughput units
Enterprise fit AWS-centric estates; strong model choice Data/analytics-centric (BigQuery), Google Workspace Microsoft-centric estates (Entra ID, M365, Teams)

Selection guidance

  1. Start where your data, identity and security tooling already live (data gravity and governance).
  2. Check model availability in required regions and compliance programs.
  3. Evaluate the specific models on your own golden dataset; platform differences matter less than model fit.
  4. Keep an LLM gateway and provider-neutral evals to avoid lock-in.

Interview questions and answers

A bank on Azure wants to use Claude and Gemini models too. How do you architect it?

Use Azure as the primary platform with Entra ID and private networking; use models available in the Foundry catalog where they meet requirements, and reach others through their managed platforms (Bedrock or Vertex AI) via an LLM gateway with consistent logging, guardrails and data handling agreements, after verifying data residency and contractual terms for each provider.