AWS Bedrock vs Azure OpenAI
Azure OpenAI is now part of Microsoft Foundry. That rename matters, because half the comparisons you will read describe products under names their vendors have retired.
The short answer
Choose Amazon Bedrock when your systems and data already live in AWS, and choose Microsoft Foundry when your company runs on Microsoft and your identity and compliance work is already there.
The old reason to prefer one was model choice. That reason has gone. Bedrock documents models from Amazon, Anthropic, Meta, Mistral, OpenAI and others. Microsoft says Foundry carries more than 1,900 models from Microsoft, OpenAI, Anthropic, Meta and others.
What is left is gravity. Data has weight, contracts have inertia, and the platform your team already knows will cost you less in mistakes.
Side by side
From each vendor's own documentation, checked August 2026. Model lists move monthly, so treat those rows as a snapshot.
| Dimension | Amazon Bedrock | Microsoft Foundry |
|---|---|---|
| Current name | Amazon Bedrock | Microsoft Foundry, formerly Azure OpenAI |
| Model vendors | Amazon, Anthropic, Meta, Mistral, OpenAI, others | Microsoft, OpenAI, Anthropic, Meta, others |
| Training on your data | Inputs and outputs are not used to train models | Not used to train without your permission or instruction |
| Model providers seeing your data | Not shared with model providers | Not available to OpenAI or other model providers |
| Human review of prompts | Not described this way in the pages we checked | Possible where automated review is not confident |
| Turning that off | Not applicable in the same form | Eligible customers may apply and must be approved |
| Keeping data in one place | Varies by model and region, with routing options | Data at rest stays in the chosen Azure geography |
| Reserved capacity | Available for steady workloads | Provisioned throughput units guarantee capacity |
- IdentityWho may call the model
- NetworkWhether traffic leaves your estate
- RegionWhere processing may happen
- ModelThe part everyone argues about
- AuditWhat you show a regulator later
Four of these five are cloud decisions you already made. Only one is about the model.
Amazon Bedrock
Where it wins
- AWS states plainly that inputs and outputs are not used to train its models or third-party models, and are not shared with model providers.
- Guardrails are a documented product feature, covering content filters, denied topics and checks that an answer is grounded in your source.
- Managed knowledge bases cover the retrieval pattern without you assembling one.
- If your data already sits in S3 and your permissions already live in AWS, nothing new has to be approved.
Where it hurts
- Bedrock Agents is now closed to new customers, and AWS points new builds at AgentCore instead. Guides written last year will send you down a dead end.
- Model availability varies by model and region, so a model you tested may not be available where you must run.
- The naming around knowledge bases and agents has changed more than once, which makes older internal documentation unreliable.
- Teams with no AWS footprint inherit a large platform to learn for one feature.
Microsoft Foundry
Where it wins
- Data at rest stays in the Azure geography you pick, and the data zone options cover the US, the EU and Asia Pacific.
- Microsoft states prompts and completions are not available to OpenAI or other model providers, and that models are stateless.
- An existing Azure agreement usually means no new vendor review, which can save more calendar time than any technical feature.
- Provisioned throughput gives a guaranteed level of capacity for workloads that cannot tolerate a queue.
Where it hurts
- Abuse monitoring may involve human review where automated checks are not confident, which surprises teams handling sensitive text.
- Switching that off is an application against Limited Access criteria, and approval is not guaranteed. Some models have stricter criteria.
- The product has been renamed more than once, so your internal runbooks and any older guide will use stale names.
- The training-data promise carries the qualifier "without your permission or instruction", which is worth reading before quoting it to a regulator.
How to choose
- Choose Bedrock if your documents, permissions and logging already live in AWS. Moving data across clouds to reach a model is a cost with no upside.
- Choose Foundry if your company runs on Microsoft identity and your compliance evidence is already Azure-shaped.
- Choose Foundry if data at rest must demonstrably stay inside one geography and you need that written down.
- Choose Bedrock if you want retrieval and safety filtering as managed features rather than as a build.
- Choose either through an abstraction layer if a switch is plausible within two years. Both are one provider behind your own interface.
- Choose neither if the workload is small and experimental. Calling a provider directly is simpler, and you can move it into a cloud later once volume justifies the paperwork.
Which platform fits your estate?
These questions are about your company, not about the models.
Where does the data the model will read live today?
Who signs off a new supplier?
How strict is your data residency rule?
What stage is the work at?
Every outcome
- Amazon Bedrock
- Your data and permissions already live in AWS, and you want managed retrieval and guardrails nearby.
- Microsoft Foundry
- Your company runs on Microsoft identity, and residency needs to be provable to an auditor.
- Call the provider directly
- The workload is small or still being proven. Add cloud plumbing when the volume earns it.
AI systems we run on cloud infrastructure
Questions architects ask
01Is Azure OpenAI still called Azure OpenAI?
Not in Microsoft's current documentation, which uses Microsoft Foundry. Models are described as Foundry Models sold by Azure. Microsoft documents upgrading an Azure OpenAI resource to a Foundry one while keeping the endpoint and API keys. Expect older guides, and maybe your own runbooks, to use dead names.
02Can we run the same model on both clouds?
Often yes, and it is a reasonable hedge for a model family available in both places. The catch is that identical model names do not guarantee identical behaviour once regional versions and safety filtering differ. Run your evaluation set on each deployment rather than assuming the results transfer.
03Should we still use Bedrock Agents?
Not for a new build. AWS documents that Bedrock Agents, now called Agents Classic, is no longer open to new customers and points new work at AgentCore. Existing customers can continue as normal. Any tutorial recommending Agents Classic for a fresh project is out of date.
04Does either option keep our data off the public internet?
Both offer private networking within their own cloud, which is a configuration job rather than a default. The more common leak is architectural: a logging pipeline, an error tracker or an analytics tool that copies prompts somewhere nobody reviewed. Audit what your own application does with the text before auditing the cloud provider.
05Which one is cheaper to run?
Cost depends on model, region, traffic shape and whether you reserve capacity, so a general answer would mislead you. Both offer reserved throughput for steady load and pay-per-use for spiky load. Model your own traffic. Then add the time your team will spend learning a platform they do not know, because that is the bill nobody puts in the spreadsheet.
