AI development platforms compared: Amazon Bedrock, Microsoft Foundry and Google Vertex AI
Each major cloud now sells one place to pick models, connect company data, build agents and monitor them. The differences are in model choice, integration and where your data already lives.
Building an AI application used to mean assembling a model, a vector database, hosting and monitoring from separate products. The large cloud providers now package these into AI development platforms. For most companies, the choice follows the cloud they already use, but the platforms differ in useful ways.
The three main platforms
| Amazon Bedrock | Microsoft Foundry | Google Vertex AI | |
|---|---|---|---|
| Owner | AWS | Microsoft Azure | Google Cloud |
| Model choice | 100+ foundation models from Amazon, Anthropic, OpenAI, DeepSeek, xAI and others | More than 10,000 models in its catalog from Microsoft, OpenAI, Anthropic, Meta and others | Google's Gemini models plus partner and open models in Model Garden |
| Agents | Agent building and hosting services | Foundry Agent Service | Agent building tools and runtime |
| Customization | Fine-tuning, continued pre-training and distillation | Fine-tuning and deployment of custom models | Tuning, plus full ML training for custom models |
| Fits best with | Data and apps on AWS | Microsoft 365, Azure data and Microsoft identity | BigQuery and Google Workspace |
Microsoft renamed Azure AI Studio and Azure AI Foundry to Microsoft Foundry. AWS describes Bedrock as a fully managed service giving enterprise access to foundation models from leading AI companies.
How to choose
Start with your data. AI applications are only as useful as the company data they can reach, and moving data between clouds adds cost, delay and security work. A company with its data warehouse in BigQuery has a reason to look at Vertex AI first; a Microsoft 365 company has the same reason to look at Foundry.
Then check model access. If a specific model is a requirement, confirm it's available on the platform, in your region, at the version you need. Model catalogs change often.
Then compare the controls: identity integration, private networking, content filtering, logging of prompts and outputs, and terms on how your data is used. Read each service's terms on how prompts and data are stored and whether they can be used for training.
Finally, model cost. Platforms charge per token for hosted models, plus storage, search and agent runtime. A pilot with real usage is the only reliable estimate.
Avoiding lock-in
Keep prompts, evaluation tests and retrieval logic in your own code, and use common interfaces where possible. That makes it practical to switch models, or platforms, as prices and capabilities change.
What IT teams should do
- Start your evaluation on the platform of the cloud that holds most of your data.
- Confirm required models are available in your region before committing.
- Build a set of evaluation tests from real tasks and run them on each model you consider.
- Turn on logging of prompts and outputs, with access controls, from the first pilot.