Key takeaways
- Data residency (where the data sits) and jurisdiction (whose law can compel access) are different questions, and an EU region does not answer the second one for a company with a non-EU parent.
- Read the deployment type, not the marketing: on Microsoft's platform, for example, Global deployments may be processed in any region, while Data Zone deployments are processed within the EU zone.
- The five realistic options are EU-based providers, hyperscaler EU data zones, hyperscaler sovereign clouds, self-hosted open-weight models and gateways that route to someone else's models.
- For health, legal, HR and public-sector data, redact personal details before the model and prefer an EU-based provider or your own servers.
Residency, sovereignty and jurisdiction are not the same thing
EU hosting describes a location, not a complete compliance assessment. Check where inference, storage, logs, backups and support access take place. Then check the operator, sub-processors and contracts. An EU endpoint can still route requests elsewhere, and a provider’s “EU Data Zone” can have a wider boundary than EU member states.
Data residency and legal jurisdiction are different checks. Identify the contracting operator and its parent company, then assess applicable access laws, sub-processors and safeguards with your legal team. A data-centre address alone does not answer those questions.
Five ways to run AI inside the EU
| Option | Where data is processed | Open question | Often a fit for |
|---|---|---|---|
| EU-based providers with their own inference APIs (for example IONOS AI Model Hub, or EU model makers such as Mistral AI) | In the provider's EU data centres | Check each provider's documents for region, sub-processors and training use. Model choice may be narrower | Regulated work where jurisdiction matters |
| US hyperscaler, EU data zone (for example Azure OpenAI Data Zone for the EU) | Within the configured zone, whose boundary must be checked | The zone's exact boundary, the parent company's jurisdiction, and which setting is the default | General business use with a solid contract |
| US hyperscaler, sovereign cloud (for example the AWS European Sovereign Cloud) | In a separate EU-only region, with operations in the EU | Whether the structure removes foreign-law exposure is argued, not settled. Model choice may be limited | Organisations that want hyperscaler tooling with stronger separation |
| Self-hosted open-weight models in your own EU cloud or on-premises | On infrastructure you control | You run, patch and scale it | The most sensitive data, public bodies, hospitals |
| Gateways and aggregators | Depends on the model behind them | A gateway with an EU endpoint may still forward your prompt to a model hosted elsewhere | Prototyping, with the routing checked |
How can you test an EU-hosted claim?
Ask these questions in writing and keep the answers with your DPIA. The details matter more than the label.
- Is inference, not just storage, performed inside the EU? Microsoft's documentation, for instance, says data at rest stays in the chosen geography while Global deployments may process prompts in any region.
- Which exact countries make up the zone? Microsoft's EU Data Zone follows its EU Data Boundary, which can include EFTA countries such as Norway and Switzerland, and Microsoft can add regions without prior notice.
- Is the EU setting the default, or must you choose it? On Microsoft's platform the documentation recommends starting with Global Standard, so an EU deployment is a deliberate choice.
- Are prompts and outputs logged, and for how long? IONOS states that content is not written to logs or persistent storage and that only request metadata is kept.
- Is your content used to train or improve any model, and is that written in the contract?
- Who operates the service, who owns the operator and which country's law can reach it?
- Which sub-processors and support staff can see your data, and from where?
- Do optional features, such as connectors, tools or preview services, send data elsewhere?
- What do the DPA, the audit rights and the exit terms say?
What changed in 2026
In September 2026, AWS announced open-weight models on Amazon Bedrock in its European Sovereign Cloud. Availability adds an option, not a guarantee for your workload. Confirm the exact model, region, service terms and data flows before choosing it.
Which option for which data?
| Type of data | Our default recommendation |
|---|---|
| Public or marketing content | Any provider with a DPA and a clear training policy |
| Ordinary customer or employee personal data | An EU data zone or an EU-based provider, with redaction before the model and audit logs |
| Health, legal, HR and public-sector data | An EU-based provider or self-hosted open-weight models, with redaction, audit logs and human review |
| Anything where a leak would be a serious incident | Self-hosted on your own servers or on-premises |
Redact personal details that the task does not need. Redaction limits exposure, but it does not automatically make the remaining text anonymous or remove your other GDPR duties.
How does this connect to the AI Act?
The AI Act does not require EU hosting. It asks you to know your role, classify the use and keep logs, documentation and human oversight where the risk calls for them. Where the platform runs matters mainly through GDPR and through how easily you can obtain logs and documentation from the provider. For the model layer, the provider of a general-purpose model carries its own obligations, which have applied since 2 August 2025, so ask your vendor for the documentation they owe you.
Send us your data types, required countries and expected workload. We can compare EU-hosted and self-hosted deployment options for your project.
Compare AI hosting optionsFAQ
Is an EU region of a US cloud provider GDPR compliant?
It can be, with the right contract and settings, because data staying in the EU simplifies transfer rules. It does not remove the jurisdiction question for a US parent company, so for the most sensitive data many organisations prefer EU-based operators or their own servers.
Can we run an LLM entirely on our own servers?
Yes. Open-weight models can run in your own EU cloud account or on-premises. You take on operations and capacity planning, and in return the data never leaves your control.
Does the AI Act require us to host AI in the EU?
No. The Act applies wherever a system's output is used in the EU. Hosting location matters mainly for GDPR and for access to the logs and documentation you need.
Sources
- Microsoft Learn, Understanding deployment types in Microsoft Foundry Models
- IONOS Cloud documentation, AI Model Hub: data handling
- AWS Security Blog, Run open weight models on Amazon Bedrock in the AWS European Sovereign Cloud, 17 September 2026
- European Data Protection Board, Opinion 28/2024 on AI models
- Regulation (EU) 2016/679 (GDPR), Chapter V on transfers, Official Journal of the EU
About the author
Artem Synytskyi · Co-founder · AI
Artem runs AI at ESADEX, building LLM, RAG and speech-to-text systems for regulated environments.
