Where does your data go when your team pastes the RFP into an AI tool?
Sovereignty is not a setting: it is an architecture that keeps your data, your processing and your models under your jurisdiction.
What is sovereign AI?
Sovereign AI keeps three things under your jurisdiction and your control: your data (subject to its jurisdiction of origin), your processing (transparency, with no exposure to extraterritorial orders) and your models (controlled versions, with no dependence on a provider's roadmap). A private DSLM, dedicated per client, deployed on premises or in a sovereign cloud, satisfies all three.
What are the three planes of sovereignty?
A sovereign DSLM (Domain-Specific Language Model) is judged on three planes, never on one. Well-hosted data is not enough if processing stays opaque, and a brilliant model is worth nothing if you control neither its versions nor its fate.
1. Data sovereignty
Information remains subject to its jurisdiction of origin. An extraterritorial provider takes it outside that perimeter: extraterritorial disclosure laws apply to the provider wherever the servers are, enterprise plans included.
2. Processing sovereignty
Transparency and control over operations: who accesses what, under which law, with no exposure to extraterritorial orders and interference. You cannot audit what you cannot see, and 0 data is used outside the client's environment.
3. Model sovereignty
Continuity and control over your technological fate: the model is versioned, its evolutions are in your hands, and your capability does not evaporate when a provider changes course, price or terms.
For an organization subject to the GDPR, NIS2 or DORA, failing on any one of the three planes is disqualifying in itself, independently of the quality of the model. This is the argument that does not expire as general-purpose models improve: they get smarter every quarter, they do not change jurisdiction. The applicable texts are consolidated on EUR-Lex, and the reference authorities (the CNIL and ANSSI in France, ENISA at European level) publish the guidance that frames these obligations.
What shadow AI use really exposes
The risk is rarely a decision, it is a habit. Under deadline pressure, teams paste specifications, audit responses, security architectures and client data into the assistant that answers fastest. Each paste is a cross-border transfer no one assessed, of content no one classified, under terms of use that discharge the provider of all liability. The remedy is not a ban: it is to give teams a specialized tool that is faster and stays inside the perimeter. This is the stance of the sovereign AI Optivalue.ai.
What are the three sovereign deployment modes?
Three deployment modes, all sovereign, from the most pooled to the most isolated. None takes your data out of its jurisdiction, and hosting is available in 80+ countries.
Shared mode
An instance and models dedicated to your organization, operated in a sovereign environment. Pooling never reaches the level of your data or your model: it stops at the infrastructure.
Private cloud
A sovereign private cloud, isolated in the jurisdiction of your choice, inside your chain of accountability. Setup takes 7 minutes, with no IT project.
On premises
The model is deployed in your own data center, inside your security perimeter. Access rights are inherited from your systems: the model never answers from a document the requester cannot open.
What is the difference between the publisher's certification and the host's?
This is a common confusion, and a costly one in an RFP. A certification covers a precise scope and owner. Adding the host's logos to the publisher's credit is a misreading that your buyer, for their part, will not make.
| Scope | Owner | Certifications |
|---|---|---|
| The publisher | Optivalue.ai | Certified to ISO/IEC 27001 (information security management system) |
| The host | The hosting infrastructure | Additionally certified to ISO 27001, 27017, 27018, 27701, 22301, 42001, 9001, SOC 1, 2 and 3 and HDS, on its own scope |
How to read this: the publisher's ISO/IEC 27001 certification attests to security governance at Optivalue.ai. The infrastructure certifications add up on the host's scope: they do not replace the publisher's and must not be attributed to it. Authoritative sources on security and compliance: the CNIL and ANSSI (France), ENISA (European Union), consolidated texts on EUR-Lex.
What does a sovereign DSLM deployment look like?
- Private per client. Your model is grounded only in your corpus, never pooled across clients.
- On premises, private cloud or dedicated sovereign SaaS. Deployment inside your security perimeter and the jurisdiction of your choice, among 80+ countries.
- No training on your data for the benefit of anyone else, contractually.
- Guaranteed deletion at end of contract, contractually and verifiably.
- Inherited access rights from your systems: the model cannot answer from documents the requester cannot open.
The argument that does not expire. A general-purpose LLM (a consumer generative AI) gets more capable every quarter, but it does not change jurisdiction. Sovereignty is a buying criterion that progress in models does not catch up with. To go further, see how to evaluate a DSLM and the comparison between build or buy.
Frequently asked questions
What is sovereign AI?
Sovereign AI keeps three things under your jurisdiction and your control: your data (subject to its jurisdiction of origin), your processing (transparency, with no exposure to extraterritorial orders) and your models (controlled versions, with no dependence on a provider's roadmap). A private DSLM, dedicated per client, deployed on premises or in a sovereign cloud, satisfies all three.
Why is a general-purpose LLM hosted outside your jurisdiction disqualifying under the GDPR, NIS2 and DORA?
When your teams paste an RFP or an audit response into a service operated by an extraterritorial provider, that data can fall under a foreign jurisdiction regardless of where the servers are, enterprise plans included. For an organization subject to the GDPR, NIS2 or DORA, failing on this point is disqualifying in itself, independently of the quality of the model.
What are the three sovereign deployment modes of a DSLM?
Three modes, all sovereign. Shared mode provides an instance and models dedicated to your organization. Private cloud isolates the whole in the jurisdiction of your choice. On-premises deployment places the model in your own data center, inside your security perimeter. No data is used outside the client's environment.
In how many countries can a DSLM be hosted?
Hosting is available in 80+ countries, which lets you keep each deployment in the jurisdiction imposed by your regulatory or contractual obligations. Data remains subject to its jurisdiction of origin, and setup takes 7 minutes, with no IT project.
What is the difference between the publisher's certification and the host's?
These are two distinct scopes. Optivalue.ai, the publisher, is certified to ISO/IEC 27001 for its information security management system. The hosting infrastructure adds, on its own scope, the ISO 27001, 27017, 27018, 27701, 22301, 42001, 9001, SOC 1, 2 and 3 and HDS certifications. Never present the host's certifications as the publisher's.
Does a private DSLM mean my data trains a third party's model?
No. A properly deployed DSLM is private per client and never pooled: your documents ground only your model, no training is done for the benefit of a third party, and contractual guarantees cover deletion at end of contract. This is the exact opposite of the default setting of consumer generative AI.