Hyperscaler AI Platforms: AWS Bedrock vs Google Vertex AI vs Azure AI Foundry
Every hyperscaler now offers a full GenAI platform. The choice usually inherits from where your data and workloads already live — but the platforms differ meaningfully on model catalog, enterprise controls, and roadmap emphasis. This is the honest comparison.
Last reviewed:
What we're comparing
- AWS Bedrock
Amazon's managed GenAI platform. Broad third-party model catalog (Anthropic Claude, Meta Llama, Mistral, Cohere, Amazon Titan/Nova). Deep IAM integration.
- Google Vertex AI
Google Cloud's ML + GenAI platform. Native access to Gemini + third-party models (Anthropic Claude, Meta Llama, Mistral). Strong on grounded-search integration.
- Azure AI Foundry
Microsoft's unified AI platform (successor to Azure OpenAI Service + Azure AI Studio). Deep OpenAI partnership + Meta Llama + Mistral + others. Copilot Studio for agent building.
Side-by-side
| Attribute | AWS Bedrock | Google Vertex AI | Azure AI Foundry |
|---|---|---|---|
| First-party frontier model | Amazon Nova (mid-tier) | Gemini 2.5 Pro (frontier) | GPT-5 via OpenAI partnership |
| Third-party model catalog breadth | Best-in-class (Claude, Llama, Mistral, Cohere, +more) | Good (Claude, Llama, Mistral) | Good (Llama, Mistral, DeepSeek, others) |
| Enterprise IAM integration | Best-in-class (deep AWS IAM) | Good (Google Cloud IAM) | Best-in-class (Entra ID) |
| Managed retrieval | Knowledge Bases (mature) | Vertex AI Search (mature) | Azure AI Search (mature) |
| Guardrails | Bedrock Guardrails (native) | Vertex safety filters | Azure AI Content Safety |
| Agent building | Bedrock Agents | Agent Builder (Vertex AI Agents) | Copilot Studio + Foundry Agent Service |
| Regions available | Widest (~30 regions) | Wide | Widest |
| Data-residency guarantees | Strong (regional isolation) | Strong | Best-in-class (contractual sovereignty options) |
| Best-fit org | AWS-native workloads; broad third-party model catalog needed | Google Cloud shops; multimodal + grounded search core to use case | Microsoft-native workloads; OpenAI models required; regulated industries needing sovereignty |
When to use which
- Use AWS Bedrock
Your data lives in AWS, you want the widest third-party model choice on one platform, and you have committed IAM integration.
- Use Google Vertex AI
You are Google-Cloud-native, Gemini fits your workload, or grounded search + multimodal are central.
- Use Azure AI Foundry
You are Microsoft-native, OpenAI models are required, or you need EU/regional data-sovereignty commitments that Azure documents deeply.
FAQs
Can we run multi-cloud AI?
Yes but the operational tax is real. Route by capability (Gemini on Vertex for grounded search, Claude on Bedrock for coding, GPT-5 on Foundry for OpenAI-specific features) and use an LLM gateway to abstract. Multi-cloud makes sense above ~$5M/year total GenAI spend; below that, single-platform wins on operational simplicity.
How do costs compare?
Very similar at the API-call level; the third-party models cost the same on any hyperscaler that offers them. Where costs diverge: managed retrieval per-query fees, egress costs, and enterprise-tier support/SLA pricing. Model at your projected volumes.
What about vendor lock-in?
High on the platform primitives (Bedrock Agents, Vertex Agent Builder, Foundry Agent Service are not portable). Medium on managed retrieval (portable in concept, painful in practice). Low on model calls (the LLM API surface is standardizing). Design for portability at the agent-framework layer where lock-in is highest.