I'm building an Azure AI Foundry environment, but Anthropic models aren't available in my region. At the moment, the available regions appear to be Sweden Central and East US 2, and I understand broader availability depends on Microsoft rolling out the newer Foundry experience to more regions.
Could I deploy the Foundry resource in Sweden Central while running my application and Container Apps in my local region? The plan would be to use the Anthropic Claude Agent SDK in Container Apps, call the models through the Sweden-based Foundry resource, and eventually move everything locally once Anthropic becomes available there.
For small proofs of concept, I could use the Foundry Playground and make API calls to the models in Sweden, although that would not provide the full Anthropic harness. I'm also trying to understand the implications for quota, capacity, data residency, data processing, security, and whether Sweden offers the required data-zone guarantees or only global routing.
We are an Azure-only organization, so moving to AWS Bedrock is not ideal. Is this cross-region setup a practical and supported approach, or are there important limitations and gotchas I should know about?
3 Answers
Be careful about assuming that a model being listed in a region means you can immediately use it. Anthropic capacity can be unavailable, and Microsoft may require a paid reservation or commitment before allocating it. A Sweden-based Foundry resource with Container Apps elsewhere may be technically workable, but you’ll need to confirm the model’s routing type, quota, supported SDK/API path, networking, and data-processing terms with Microsoft. The older and newer Foundry experiences also don’t appear to have identical regional behavior, so creating a resource through an older interface is not necessarily evidence that the same deployment is supported in the current model.
Sweden Central may not solve the residency issue. In our case, the available deployment was global rather than data-zone routed, which meant requests could potentially be processed outside Sweden, including in the United States. That was unacceptable for our policies and customers. Capacity was another problem: even when a region appeared available, quota had to be requested and there was no guarantee it would be granted. We eventually used an OpenAI model instead, although we had to deploy in West Europe because North Europe lacked capacity.
We ended up using AWS Bedrock for the Anthropic requirement and kept Azure Foundry for OpenAI models. Unfortunately, that was the cleaner option for us, especially where regional or sovereign hosting requirements mattered.
That would be difficult for us because we’re committed to Azure, but it does suggest the regional Anthropic support in Foundry may not be reliable enough yet.
That matches what I’m seeing. The older interface appears to expose options that aren’t obvious in the newer experience, so I’m hesitant to build around it without confirming the support and residency details.