Sovereign AI: what the word covers, and how to check a claim
Sovereign AI is artificial intelligence whose every layer — the model, the servers that run it, the company that operates them, what is kept and the law that applies — stays under the control of those who rely on it, not of a foreign company or government. Few services are sovereign on every layer, so a claim is only worth what it says layer by layer. Evanesca is a private AI agent and chat that runs all 16 of its models in Europe, at European operators, and keeps no copy of your messages; this page shows where it holds and where it does not.
Facts checked on 6 Nov 2026
What is sovereign AI?
The phrase is used at three scales, and it does not promise the same thing at each.
- For a state, sovereign AI is industrial policy: computing capacity on its own soil, model makers of its own, and rules it writes itself. In the EU, those rules are the GDPR for personal data and the AI Act for AI systems.
- For a company, it is control of its data and suppliers: which companies process its documents, where, under which law — and the freedom to switch supplier.
- For a person using a chat, it is concrete: who keeps what you type, and which government can order that company to hand it over.
All three come down to one chain — a model, servers, an operator, a law — and one link under a foreign law is enough to bring that law in. That is where sovereign AI differs from private AI: private AI asks who can read your words; sovereign AI asks who can order someone to hand them over.
The layers of a sovereign AI service
- The model. Who made it, and are its weights published? Open weights can be run anywhere, without the maker on the route of your request; a model served only through its maker's API ties you to that maker.
- The servers. Where the model runs: the country and the data centre. Necessary, never enough.
- The operator. Which company runs those servers, and who controls it — the layer that decides. The US CLOUD Act lets US authorities order a provider under US jurisdiction to disclose data in its “possession, custody, or control”, wherever it is stored source. A US cloud's EU region is still run by a US company: “hosted in an EU region” does not make a service sovereign.
- The app in between. The website, the gateway, the agent's loop: who runs them, and do they keep prompts?
- What is kept. What each link stores, for how long, and whether it trains on it. Data never stored cannot be handed over later.
- The law. Which courts and regulator apply — the GDPR in the EU, Swiss law under an EU adequacy decision — and whether a foreign law still reaches a link.
“Under US jurisdiction” is broader than “American”. The US Department of Justice writes that “U.S. jurisdiction is not limited to U.S. corporations, U.S. headquartered companies, or companies owned by U.S. persons” source. So a European company with a US subsidiary is out of the CLOUD Act's direct reach — not out of reach. The sovereignty page explains why Evanesca's own wording stops at “direct”.
How to judge a sovereign AI claim: a checklist
Ask each question of every company in the chain, not only the one named on the website. A one-word claim — “sovereign”, “European”, “GDPR-compliant” — answers none of them.
| Question | A good answer | A warning sign |
|---|---|---|
| Who made the model, and are its weights published? | The maker and the licence are named; the weights can be downloaded | “Our own model”, with no maker, licence or model card |
| Where does it run? | A country and an operator, model by model | “Hosted in Europe”, with no operator named |
| Who controls the operator? | A company with no parent outside Europe, its foreign subsidiaries stated | A US cloud's EU region presented as sovereign |
| What does the app in between keep? | Prompts held in memory only, said in the privacy policy | Conversations stored on the provider's servers by default |
| What does each operator keep, and for how long? | Each operator's own rule, quoted with its exceptions | “We keep no data”, with no word on the operators |
| Is anything used for training? | No — or an objection, with its exceptions stated | Training “to improve the service”, with no way to object |
| Which law applies, and can a foreign one reach a link? | The GDPR or an adequacy decision, and every company outside Europe named | No list of sub-processors |
| Can you check it on a real answer? | The route of each answer is shown | A badge, with nothing behind it |
Evanesca's reply to the last line is the answer passport under every answer: model, operator, country, law and what was kept. The checks each model passed or missed are shown before you pick it.
Where Evanesca stands, layer by layer
Evanesca does not call itself sovereign AI as a whole. Here is each layer, with the part that holds and the part that does not.
| Layer | Evanesca | The limit |
|---|---|---|
| Model | 16 of the 16 models have published weights models | Most of the models are made outside Europe; Mistral AI and the Swiss AI Initiative are among the European makers. How a model was trained stays its maker's choice. |
| Servers | Every model runs in Europe: in France (Scaleway, OVHcloud as a fallback), in Mistral AI's EU region, in Germany (AKI.IO, Verda) or in Switzerland (Infomaniak) sovereignty | Switzerland is outside the EU, under an EU adequacy decision source; Mistral AI does not say in which EU country its region is. |
| Operator | OVHcloud, Scaleway, Mistral AI and Infomaniak are European companies with no US parent, out of the CLOUD Act's direct reach how they were checked | OVHcloud source and Scaleway source each have a US entity, and Mistral AI an office in Palo Alto source. AKI.IO and Verda do not publish their full shareholding, so Evanesca cannot say that they have no US parent. |
| App and relay | The website, the relay to the operator and the agent's loop run on Vercel, in its Paris region; they hold messages in memory only, and write none to a database, a disk or a log privacy | Vercel Inc. is a US company, under US law, CLOUD Act included; its DPA puts its “primary processing facilities” in the United States source. A US company sits on the path, even with no message to keep. |
| What is kept | No prompt and no answer by Evanesca; your history stays in your browser; of your messages, only billing data (tokens and cost) | Each operator applies its own rule — OVHcloud: billing data only; Scaleway: no prompt by default, requests that cause errors or look abusive, up to 14 days; Mistral AI: nothing beyond the time needed to answer (zero data retention on Evanesca's account since 17 November 2026); AKI.IO: states it logs and stores no prompt or output, usage metadata, 90 days; Verda: states it stores no request, context in an in-memory cache, removed within 2 hours; Infomaniak: states it records no query passport |
| Training | Never by Evanesca. Mistral AI lets its customers object to training, and Evanesca's account objects | No exception at the moment. |
| Account, e-mail, payments | Account and billing data — never a message — in a Neon database in Frankfurt; e-mails sent through Resend; payments by Stripe Payments Europe (Ireland) | Neon belongs to Databricks, Inc. source; Resend is Plus Five Five, Inc. source; Stripe transfers data to Stripe, LLC in the United States source. The transfers rest on the EU–US Data Privacy Framework and standard contractual clauses transfers. |
| Web search | Off by default; when on, the queries the model writes and the pages it reads go to Linkup, a French company — never the conversation source | Evanesca has not verified where Linkup's servers run. |
| Law | Evanesca is published by FLEECE AI SASU, a French company: French law and the GDPR apply, and the AI Act's transparency duties AI Act | European, French, German and Swiss authorities can obtain data under their own law, and US authorities can ask through mutual legal assistance. |
How each operator was checked — owners, subsidiaries, its own words on the CLOUD Act — is on the sovereignty page; every company that processes data is listed in the legal notice.
Sovereign, private, compliant: three different promises
The three words are often used as one. A service can keep one promise and break another.
- Private is about who can read your words and what is kept. A service can be private and still run under a foreign law: private AI covers that angle.
- Sovereign is about whose companies and whose law are in the chain. A service can be entirely European and still keep everything you type.
- Compliant is about meeting the rules. A US company can comply with the GDPR — the EU–US Data Privacy Framework lets data go to certified US companies — and still be bound by the CLOUD Act: why hosting your AI in the EU is not enough.
For the European side in general — which makers build models in Europe, which companies run them — see European AI and European inference providers.
When an API is not sovereign enough
With an API, another company runs the model, and sovereignty depends on choosing it well. Some needs go further, and Evanesca does not meet them:
- No company from outside Europe anywhere in the chain, website and payments included: Vercel, Databricks, Resend and Stripe's US entity are in Evanesca's.
- Air-gapped sites, custom weights or very large volumes: your own cluster gives the strongest custody of your data, at a cost this article works out server by server.
- Hardware under your sole control: a small open-weight model on your own computer is slower and less capable, but the data stays on it.
For teams that choose the API route with a contract: a signed art. 28 DPA, the list of sub-processors, SSO and spending limits are on the business page; the Agent mode follows the same route as the models it calls.
Frequently asked questions
What is sovereign AI?
AI whose model, servers, operator, retention and governing law stay under the control of those who rely on it, not of a foreign company or government. For a state, that means compute, model makers and rules of its own; for a company, control of its data and suppliers; for a person, knowing who can be ordered to hand over what they type.
Is AI hosted in an EU region sovereign?
Not by itself. The US CLOUD Act follows the company, not the server: a provider under US jurisdiction must disclose data in its possession, custody or control, wherever it is stored. What decides is who controls the operator — and a European company with a US subsidiary is out of the Act's direct reach, not out of reach.
Is Evanesca sovereign AI?
Partly, and it says which parts. Every model runs in Europe at a European operator, and Evanesca keeps no copy of your messages. But the website and its relay run on Vercel, a US company, in its Paris region, and account data, e-mails and payments go through Databricks (Neon, in Frankfurt), Resend and Stripe. It is not sovereign on every layer.
Can a model made in China or the United States be part of sovereign AI?
Its weights can. When a European operator runs the published weights on its own servers, the maker is not on the route of the request. What stays the maker's is how the model was trained and behaves — which is why Evanesca publishes the checks each model passed.
Read next
- The best ChatGPT alternatives, by what you needThe best ChatGPT alternative depends on your need: privacy, European law, free use, research, code or offline. Claude, Lumo, Duck.ai, Vibe and more.
- Private AI agent: what makes an agent private, and how Evanesca's worksA private AI agent carries out tasks without keeping your data. What to check — where it runs, what it keeps, what it can touch — and Evanesca's Agent mode.
- Private AI: what it means, and how to judge onePrivate AI keeps your words from being stored, read or used for training. The three kinds, a checklist to judge any of them, and the main options compared.
- Is there a European ChatGPT? What it means, and the optionsIs there a European ChatGPT? What to check — EU company, servers in Europe, the CLOUD Act — and how Mistral Vibe, Lumo and Evanesca compare.
- European AI: who makes it, who runs it, and how to use itWhat European AI covers: models made in Europe, European operators and data centres, the AI Act — which ones Evanesca runs, and how to use them today.
- AI for lawyers and law firms: what it is good for, and how to keep client secretsAI for lawyers: drafting, summaries, translation, contract review — and client confidentiality. What the French bar says, a checklist for law firms.