Forgeron3
/ SovereigntyAug 11, 20259 min read

Why choose sovereign AI in 2026

The word “sovereign” is everywhere. Often wrongly. Here are seven concrete criteria to tell real sovereignty from a marketing claim — and why it changes everything for your data.

F3
The Forgeron3 teamMarseille & Paris

The real issue, in two sentences

Sovereign AI is AI you keep legal, operational, and technical control over. The opposite is AI whose operator can, at some point, be compelled to hand your data to a foreign state — without you knowing.

For most SMBs, the risk is theoretical. For accounting firms, local governments, and industrial businesses with trade secrets, it’s concrete.

1. Physical hosting: where the servers are

The first criterion, the most obvious, the one most often lied about. Ask: “where are the servers on which my data is stored and processed?”. The right answer is precise (one or two named datacenters, in France, operated by a French or European actor).

A vague answer (“Europe,” “OVH or equivalent,” “partner datacenters”) should trigger an alert. Ask for the name of the datacenter and the operator’s jurisdiction.

2. Applicable jurisdiction in the contract

The contract must be governed by French law, before French courts. A “law of the State of Delaware” clause on a “sovereign” provider is a red flag.

Also check the parent company’s nationality. A French company can have a US parent, which is enough to expose data to the Cloud Act (see point 7).

3. The AI model used: open source or black box?

The AI engine can be:

  • Open source (Mistral, Llama, Qwen and their variants), hosted by a French operator.
  • Proprietary (OpenAI, Anthropic, Google), accessed via API.

Genuinely sovereign providers run open source models on their own infrastructure. Providers who merely wrap an OpenAI API aren’t sovereign — your data passes through their infrastructure.

The trick question”Are your models hosted in-house or accessed via a third-party API?”. If the answer hesitates, you have a black box — not sovereign AI.

4. Use of data for training

The contractual commitment to expect: “your data is never, under any circumstances, used to train or improve the model”.

A vaguer formulation (“we don’t reuse your data without your consent”) opens the door to an opt-out hidden in the T&Cs.

5. Access and audit logs

You must be able, at any time, to see who has accessed what: the provider’s internal operators, client-side users, maintenance staff. Without these logs, you can’t respond to an audit or a CNIL request.

A serious provider supplies exportable logs, with signed timestamps and configurable retention.

6. Reversibility: what happens when you leave

Three questions to ask before signing:

  1. Within how many days can I retrieve all my data in a usable format?
  2. Within how many days is my data fully deleted from servers and backups?
  3. Is a destruction certificate provided?

Acceptable answers: 7 days for export, 30 days for full deletion, systematic certificate.

7. The Cloud Act trap and extra-territorial jurisdictions

The Cloud Act (US, 2018) allows US authorities to require any US company, or subsidiary, to hand over data even if it’s stored in Europe. No provider with US ownership can escape it, even with datacenters in France.

Same on the Chinese side (2017 cybersecurity law). For real sovereignty, you need a 100% French or European chain, at every step — ownership, infrastructure, models, operators.

To go further on the questions to ask your provider, see The essential questions to ask your AI provider. And our security & GDPR page details our commitments.

Audit a provider

Twenty minutes to walk through the seven criteria above against your current or prospective provider. We tell you where the red flags are.

Book a demo