Which AI providers keep your data in the EU

An EU Data Boundary service for tenants with an EU or EFTA sign-up location. Microsoft states that EU traffic stays inside the boundary.

The catch Anthropic models offered inside Copilot sit outside the EU Data Boundary. Tenants that bought Multi-Geo are out of scope even with an EU address. LLM calls route to the nearest data centre and can go to other regions when capacity is tight.

Regional Azure services you deploy into an EU Data Boundary region are in scope for the boundary. Non-regional services have to be configured separately.

The catch The commitment covers customer data and pseudonymised personal data, and Microsoft documents limited circumstances in which data still moves outside the boundary. Read that transfer list before you rely on it.

Data residency is configured per project. Europe (EEA and Switzerland) supports regional storage and regional processing.

The catch Abuse monitoring logs are generated for all API usage and kept up to 30 days by default. You need approval for Zero Data Retention or Modified Abuse Monitoring to change that.

Two settings govern geography. Inference geo controls where the model runs, workspace geo controls where data sits at rest.

The catch Neither setting offers Europe. EU residency for Claude models exists only through AWS Bedrock or Google Gemini Enterprise Agent Platform (formerly Vertex) EU endpoints, so the answer depends on which door you come through.

Regional endpoints such as europe-west4 keep processing in that region. The EU multi-region endpoint keeps processing inside EU locations.

The catch The global endpoint does not support data residency. If the region is left out of the call, the guarantee goes with it. ML processing commitments now reach well beyond the US and EU, into Brazil, Canada, the UK, Australia, India, Japan, Singapore and South Korea, so check the model's own location table rather than assuming it stops at Europe.

Bedrock runs in EU regions, and each region holds its own model deployment accounts operated by the Bedrock service team.

The catch Cross-region inference profiles move traffic between regions inside a geography. Check the profile you are actually calling is an EU one, because the model choice does not decide this for you.

Mistral states data is hosted in the European Union by default. Depending on the feature, it can be temporarily transferred to sub-processors outside the EU, under Article 46 GDPR safeguards.

The catch Inputs and outputs are kept for 30 rolling days for abuse monitoring unless Zero Data Retention is switched on, and ZDR is not available for the stateful products such as chat, agents, libraries and batch files.

Serverless models are hosted in a data centre in Paris operated by OPCORE. Personal data may be stored in the Paris region.

The catch Aggregated and anonymised usage data is kept up to 6 months, and request content can be stored temporarily when traffic causes errors or security issues.

OVHcloud states the service is hosted and run in the European Union.

The catch The model catalogue is open weight models. If you need a frontier proprietary model, this route does not carry one.

Inference endpoints and managed vector databases run in ISO 27001 certified IONOS data centres in Germany.

The catch Open weight models only, and IONOS positions itself as the AI system provider while model documentation obligations stay with the model developer.

Data is stored and processed exclusively in Europe. STACKIT belongs to the Schwarz Group and states it is independent of external investors.

The catch A smaller model catalogue than the hyperscalers, so check the specific models you need are on it before you plan around this.

Installs into an environment you choose using Helm and Kubernetes: on premises, or on a cloud provider you pick, including StackIT, AWS, Google Cloud, or Azure.

The catch Aleph Alpha does not offer PhariaAI as a SaaS product themselves. This is a stack to deploy rather than an endpoint to call, so the real cost is the operations around it.

Llama, Mistral, Qwen and the other open weight families run on a European VPS or on hardware in your office. Nothing leaves the machine you rent or own.

The catch There is no vendor to hold to a commitment and no support line. You own the updates, the evaluation and the uptime, and quality sits below the frontier models.