Sovereignty guide
Sovereign AI: what hosting in Europe protects, and what it does not
Updated: · General information, not legal advice.
In short
Hosting an AI system in a data centre located in Europe does not remove it from US law: the CLOUD Act allows US authorities to require data from a provider subject to their jurisdiction, wherever the data is stored. The sovereignty of an AI system is therefore judged on four criteria: the jurisdiction of its providers, where data and computation are located, external dependencies (model APIs in particular) and control over keys and logs.
GDPR applies in every case: legal basis, impact assessment, processor contracts and safeguards for transfers remain the organisation's responsibility, whoever the host is.
What the CLOUD Act says
| Point | Content | Source |
|---|---|---|
| Enactment | Clarifying Lawful Overseas Use of Data Act, enacted on 23 March 2018 as part of the federal appropriations act | US Department of Justice |
| Scope | A provider of electronic communication or remote computing services must disclose data in its possession, custody or control, whether stored in the United States or abroad | 18 U.S.C. § 2713 |
| Challenge | The provider may move to quash a request in certain cases, in particular when the customer is not a US person and disclosure would violate the laws of a country that has concluded an agreement with the United States. This route therefore depends on whether such an agreement exists with the country concerned | 18 U.S.C. § 2703(h) |
| Customer notice | A court may order the provider not to inform its customer of the request for a set period | 18 U.S.C. § 2705(b) |
On the European side, Article 48 GDPR provides that a judgment or decision of a third-country court or authority requiring the transfer of personal data is only recognised or enforceable if based on an international agreement, such as a mutual legal assistance treaty (GDPR, Article 48). A provider subject to both laws can therefore face conflicting obligations.
Transfers of personal data to certified US companies are also covered by the adequacy decision of 10 July 2023 on the EU-US Data Privacy Framework (Implementing Decision (EU) 2023/1795). That decision covers commercial transfers; it does not remove the reach of the CLOUD Act.
The 4 sovereignty criteria for an AI system
| Criterion | Question to ask | Why it matters for AI |
|---|---|---|
| Provider jurisdiction | Which law applies to the host, the model vendor and their parent companies? | The CLOUD Act follows the provider, not the server |
| Location | Where are data, search indexes and logs stored, and where does computation take place? | Prompts and logs often contain more sensitive information than the source data |
| External dependencies | Does the system call a model API, a search service or telemetry outside the EU? | A single external call is enough for data to leave |
| Control over keys and logs | Who holds the encryption keys and who can read the decision logs? | The logs required by the AI Act for high-risk systems are themselves sensitive data |
For cloud services, France has an official benchmark: the ANSSI SecNumCloud qualification, whose current version adds criteria for immunity from non-European laws (EU headquarters, majority European ownership, limits on using non-EU third parties, operational autonomy). The qualification covers a specific offering, not a whole provider (ANSSI, SecNumCloud FAQ).
Comparing deployment options
| Option | Provider jurisdiction | External dependency | Operating effort | Suitable for |
|---|---|---|---|---|
| On-premises (the organisation's servers) | The organisation's own | None if the model is local | High: hardware, updates, security | The most sensitive data, isolated environments |
| Private cloud run by a European provider | European law, to be checked for each provider and its subcontracting chain | Low if the model is hosted in that cloud | Medium | Sensitive data, variable capacity needs |
| European region of a US provider | European and US law | Depends on the architecture | Low | Data whose exposure to US law is accepted after analysis |
| Model API hosted outside the EU | The vendor's country | Total for the data sent | Very low | Non-sensitive data, prototypes |
| Hybrid | Depends on the split | Limited and listed | Medium | Sensitive data processed locally, non-sensitive tasks delegated |
The choice is made per data type and per process, not for the whole organisation. The options KOREV offers are detailed on the Deployment page.
GDPR: what remains your responsibility, whoever the host
- Legal basis and purpose of each processing operation (Articles 5 and 6).
- Data protection impact assessment when processing is likely to result in a high risk (Article 35).
- Processor contract with each provider processing data on your behalf (Article 28).
- Safeguards for transfers outside the European Economic Area, including by your providers' providers (Articles 44 to 49).
- Decisions based solely on automated processing: the right to obtain human intervention in the cases provided for (Article 22).
- Minimisation and retention of data, including in prompts and logs.
Source: Regulation (EU) 2016/679.
What KOREV AI offers and does not claim
- Choice of deployment: European cloud, hybrid or on-premises, depending on the environment defined by the organisation. The same architecture works at each level of independence.
- A model that can run locally: KOREV Atlas is built on Olmo 3 32B, an open model released by the Allen Institute for AI (Ai2), so it can run without a mandatory call to an external API. For on-site execution, a compact inference node (for example an MSI EdgeXpert) can host the local models and the KOREV Evidence agents. See the deployment options.
- Listed dependencies: in hybrid mode, the accepted external dependencies are listed explicitly for each project.
- Client data: KOREV does not use its clients' data to train public models.
- No certification claimed: KOREV AI holds no SecNumCloud qualification, HDS certification or ISO 27001 certification to date. If one of these is required, it must be provided by the host chosen for the project.
Source: Trust Center.
Checklist: assessing the sovereignty of an AI system
- List every provider in the system: host, model vendor, vector database, web search, telemetry, support.
- For each, identify the jurisdiction of the contracting entity and of its parent company.
- Map the flows: where documents, prompts, answers, logs and backups go.
- Identify every call to an API outside the EU and decide whether it is acceptable for that type of data.
- Check who holds the encryption keys and who can access the logs.
- If SecNumCloud is required, check that the specific offering used (GPU compute, storage, API) is within the qualified scope.
- Update the record of processing activities, the impact assessment and the processor contracts.