Microsoft and Anthropic called it general availability. European architects call it a non-starter.

Claude Opus 4.8, Haiku 4.5, and Sonnet 5 now sit inside Microsoft Foundry with Entra ID authentication, role-based access controls, and billing that draws down existing Azure Consumption Commitments. Procurement teams celebrate. Legal teams flinch.

The announcement frames enterprise readiness as a solved problem. Native governance. No new vendor onboarding. Budget already committed. For internal champions who have spent months navigating procurement, this feels like victory.

Then the practitioners asked where the data goes.

Jiri Forman, Head of Transformation at CETIN, posted the question that dominated the LinkedIn thread: Hosted on Azure with all data processing possible within EU? Forty-nine reactions. More engagement than any other comment.

Murat Yasartas, Global AI Team Lead at Sobi, answered: Unfortunately it's not, data zone is only US based for now.

Microsoft's own documentation confirms it. Even under the "Hosted on Azure" option, Anthropic remains the independent data processor for prompts and outputs. Deployment options read "Global" or "DataZone" — but no European data zone exists for Claude models today. Automatic safeguards can flag content for Anthropic Trust and Safety review, meaning customer data can leave the Azure boundary on an exceptions basis even under the Azure-hosted option.

On Reddit, a practitioner working with Dutch financial services was blunt: My client, a big Dutch bank, does not allow the use of Anthropic models through Foundry due to this reason.

Radubogdan added technical precision: Anthropic models 'hosted by Azure' are deployed in Foundry Sweden. Deployment type is 'Global Standard' so inference can happen anywhere.

The Sweden endpoint gives a European address. "Global Standard" means the actual inference routes to US infrastructure. The distinction is not semantic. It is jurisdictional.

Gregor Beuster, a Member of Technical Staff, crystallized the friction point European architects keep hitting: Still a marketplace model with Anthropic being the operator, which will ruffle some feathers in German enterprise. OpenAI models being Azure native are much easier to get approved.

That distinction matters. OpenAI models on Azure are first-party. Microsoft operates inference. Data stays within the Azure trust boundary. EU data zone deployments exist. Claude models on Foundry are third-party marketplace offerings. Anthropic operates inference as the data processor regardless of hosting option. The US CLOUD Act applies.

German enterprises do not ruffle. They reject.

Microsoft knows this. The company has spent years building sovereign cloud constructs for Europe — Azure Germany, Azure EU Regions, the EU Data Boundary program. OpenAI integration benefited from that architecture. Anthropic integration did not.

Why? The marketplace model shifts operational control to the model provider. Anthropic runs the inference stack. Microsoft provides the storefront. That arrangement lets Microsoft expand its model catalog quickly. It also means Microsoft cannot guarantee data residency for models it does not operate.

Anthropic has not built European inference capacity. The company serves US customers from US infrastructure. European demand runs through the same pipes. No regional isolation. No data processing addendum that satisfies GDPR Article 28 for regulated European workloads.

Sonnet 5 arrives at promotional pricing — $2 per million input tokens, $10 per million output tokens through August 31. The discount accelerates adoption in regions where deployment is legal. It does nothing for regions where it is not.

European enterprises face three choices. Deploy Claude through Foundry and accept US data processing — a non-starter for banking, healthcare, public sector, and any organization with contractual or regulatory data residency obligations. Route traffic through Azure Private Link to the Sweden endpoint and pretend "Global Standard" means European inference — a compliance fiction that collapses under audit. Or use OpenAI models on Azure, which offer genuine EU data zones and first-party operational control.

The market signal is clear. Microsoft's model catalog strategy prioritizes breadth over sovereignty. Every third-party model added through Foundry inherits the same limitation: the model provider processes data, the model provider determines jurisdiction, the model provider's legal exposure becomes the customer's legal exposure.

Anthropic could fix this. The company could deploy dedicated European inference capacity, sign EU data processing agreements, accept European regulatory oversight. It has not. The GA announcement suggests no timeline for doing so.

Microsoft could fix this. The company could operate Claude inference itself, as it does for OpenAI models, bringing the models inside the Azure trust boundary. It has not. The marketplace model scales faster.

Until one of them moves, European enterprises have a new model in their catalog they cannot legally use. General availability without geographic availability is not availability. It is marketing.

The procurement win is real. The compliance loss is total. European architects will keep using OpenAI on Azure — not because it is better, but because it is permissible.