EU-hosted AI models run inference on compute inside the EU. In 2026, practical options include Connic Models for deployed agents, Scaleway Generative APIs, OVHcloud AI Endpoints, and open-weight models on infrastructure operated by the buyer. Mistral’s EU endpoint uses a broader EU-and-EFTA boundary, so buyers that require EU-member-state-only processing should confirm its scope contractually. Processing location, retention, operator, portability, fallbacks, and the rest of the application data path all matter.
What does EU-hosted mean for an AI model?
The model’s creator does not determine where a prompt is processed. An American model can run on a server in Paris; a European model can run on a global endpoint. Procurement needs the data path for the exact endpoint used in production.
EU-hosted AI model providers and deployment options compared
This table was verified against provider documentation on 17 August 2026. It compares the production boundary rather than ranking model quality, which changes faster and must be tested on the intended workload.
| Option | EU boundary | Retention | Model choice | Interface | Best fit |
|---|---|---|---|---|---|
| Connic Models | Every connic/* call uses EU inference capacity. The catalog labels upstream retention exceptions. | No prompt or output content retained for billing. Narrow security, error, and labeled upstream-provider exceptions can apply. | Managed catalog spanning multiple creators | Model ID in agent YAML | Teams that need models inside a deployed, observable agent runtime |
| Mistral Regional Inference | The api.eu.mistral.ai endpoint processes requests in EU and EFTA data centers. Its control plane is not regional. | ZDR is an approval-gated organization control on paid plans and applies only to supported stateless API calls. | Models available on Mistral’s EU endpoint | Mistral SDK or HTTP API | Direct Mistral API users who can accept its regional limits |
| Scaleway Generative APIs | Serverless and dedicated inference run in European data centers. The public API is currently in Paris, fr-par. | Content ZDR by default. Exceptional harmful or error requests may be held for up to two weeks; anonymized usage metadata may remain for up to six months. | Curated serverless catalog or dedicated deployment | OpenAI-compatible API | Applications that need a direct European model API |
| OVHcloud AI Endpoints | AI Endpoints runs on OVHcloud’s European infrastructure and advertises zero data retention beyond billing data. | The provider states that it retains only data required for billing. | Broad catalog of open-weight models | OpenAI-compatible API | Teams that want a France-hosted, European-operated API for open-weight models |
| Self-hosted open weights | The operator selects the EU region, network, logs, and retention policy. | Set by the operator's logging, storage, support, and incident-response controls. | Any model whose license and hardware needs fit | vLLM, Ollama, or another serving layer | Requirements that include infrastructure control or private-network operation |
Sources: inspect the Connic model boundary, Mistral regional inference, Mistral privacy controls, Mistral ZDR scope, Scaleway data policy, and OVHcloud AI Endpoints capabilities.
A connic/* model selected in agent YAML deploys with traces, retries, budgets, and evaluation already connected.
Explore Connic ModelsWhich option fits each workload?
What did the Commission-requested cloud and AI study find?
The study, published on 11 August 2026, identifies two interrelated problems: limited and geographically concentrated computing capacity in the EU, and dependence on cloud and AI services supplied by non-European providers. It also examines lock-in practices along the AI computing stack and risks from third-country laws with extraterritorial effects. European Commission study summary (2026).
The report uses a 2025 capacity baseline despite its 2026 publication date. Germany, France, the Netherlands, and Ireland account for 63% of estimated EU27 capacity among the data centre operators, cloud providers, and telecom sites it inventories. Enterprise and public-sector facilities and high-performance computing data centres are outside that inventory. study methodology and source years.
A documented or contractually bounded EU endpoint addresses where processing occurs. Procurement teams should separately review operator control, exposure to third-country law, and portability.
How to verify an EU model endpoint before procurement
Ask for answers tied to the exact product, operator, and endpoint. A general cloud-region page cannot prove where one model API sends a prompt or how easily the deployment can move.
Record the production base URL, region selector, model ID, operator, and every fallback route.
Trace prompts, outputs, caches, safety copies, logs, embeddings, tool payloads, and evaluation data through each processor.
Read the retention rule and its exceptions; confirm whether zero retention is default or enabled per account.
Check the contracting entity, ownership and control, subprocessors, support-access locations, backups, failover regions, and any third-country law that may compel access.
Document the exit path: portable APIs and formats, export of model and runtime configuration and traces, and a tested primary or fallback route through another operator.
Put the data, operator, and fallback boundaries in the DPA or order form and add a review date for model, endpoint, or service changes.
The same review must cover the surrounding agent platform. Retrieval stores, traces, tool calls, judges, and guardrails can carry the same customer data as the model prompt. The complete agent data path comparison describes the surrounding infrastructure that must also be assessed.
Residency narrows the shortlist; tests choose the model
Once the data boundary is acceptable, evaluate the models on representative cases. Compare task success, structured-output validity, tool-call accuracy, latency, and total token cost. A benchmark score cannot show whether a refund agent chose the correct tool or produced the schema expected by the application.
Our platform lets a Project mix managed and BYOK models as primary and fallback choices. The same test suite can run against each candidate, followed by A/B testing on live traffic when an offline test is not enough. The EU boundary should be identical on both sides of the test.