Back to blog

Sovereign AI: verifying the scope of managed inference

Article created on November 18, 2026 · Source publication November 17, 2026 · Sources · Google Cloud, S3NS

On November 17, Google Cloud announced that Gemini models are coming to Google Cloud Dedicated. In the same announcement, Google confirmed that managed inference is generally available on PREMI3NS, the offering operated by S3NS in France. The two announcements cover different service scopes. When procuring sovereign AI, start by identifying the model, the service and the controls that apply to your business workflow.

1. Distinguish the model, service and qualification

S3NS describes a catalog of open-weight models, including models from Mistral and Google, a registry that lets customers import their own models and an online inference service. Google separately announced that Gemini models are coming to Dedicated. S3NS says its managed service runs on SecNumCloud 3.2-qualified infrastructure and that the process to extend qualification to the managed AI layer is underway. Ask for the current qualification scope for the service and version you are buying.

2. What changes for Belgian and French organizations

An SME building a document assistant can compare managed operation with a local model using a synthetic corpus. An IT department or public administration that requires qualified services can request a data flow diagram and responsibility matrix covering support, administrative access and subcontractors before procurement. For a company operating in both countries, the location of compute is one factor to assess alongside the requirements of each processing activity and contract.

3. Underside’s analysis · map RAG and business actions

Map document storage, embedding generation, indexing, inference and logging separately. In a RAG system connected to Odoo Enterprise, enforce read permissions before retrieval and keep write approvals in the ERP. If Apple Enterprise devices access the assistant or run a local model, include caches and exports in the data flow review and apply the same data classes as in the cloud.

4. Test your exit plan before going live

Use fictional records to compare responses and citations, verify that access is denied where required and measure latency under a representative load. Simulate loss of the model endpoint. Verify that automation pauses writes, preserves pending records and resumes after approval without duplicates. Record the model version, what you can export and how to rebuild the index in the fallback environment.

Include the current qualification scope, data flow diagram and fallback test results in the procurement file. Assign an owner to each control.

Scope a sovereign AI deployment

Read the official Google Cloud announcement · Read the official S3NS announcement