Sovereignty
Sovereignty: a model’s nationality matters less than you think
Sovereign AI means keeping the ability to choose and switch models: where processing runs, the law that applies and RAG count for more than the flag.
The word sovereignty has a bad reputation in tech. People associate it with inward-looking rhetoric, with national solutions chosen on principle, sometimes at the expense of quality. Does being sovereign mean using only models trained at home? At Oxebo, we defend another reading. For an organisation, being sovereign means keeping the ability to choose, and to change its mind.
This ability plays out on four levels: data, and the law it lives under; infrastructure, meaning the place and the operator running the processing; models, and the ability to switch them; and finally competence, which is often forgotten, meaning understanding the models you use and knowing how to evaluate them on your own documents. The country where a model was trained does not appear on this list.
Three signals pushing towards sovereignty
The first signal is geopolitical. On 10 June 2025, before the French Senate’s committee of inquiry on public procurement, Microsoft France’s director of public and legal affairs answered that he could not guarantee, under oath, that French citizens’ data entrusted to Microsoft through the public purchasing agency Ugap would never be transferred to US authorities without the agreement of the French authorities, adding that “this has never happened” (our translation). He was honestly describing a legal reality: an American company remains subject to American law, wherever its servers are. In June 2026, Anthropic suspended access to two of its models for nearly three weeks to comply with US export controls, according to the company’s statement, before restoring it from 1 July, according to Al Jazeera. An administrative decision can cut access to a model overnight, whatever the goodwill of its publisher.
The second signal is economic. A model’s price is its publisher’s decision. On 29 September 2026, at its DevDay conference, OpenAI launched a $500-a-month plan and announced a cut in the usage included in its $200 plan, according to The Next Web, while presenting a model it describes as close to its flagship at a fifth of the price per token. Volatility works both ways: an organisation tied to a single supplier absorbs the increases and misses the decreases.
The third signal comes from European law. The Data Act has applied since 12 September 2025 and, from 12 January 2027, its Article 29 prohibits cloud providers from charging any fee for switching provider. The law itself pushes towards reversibility.
Pragmatic and open
We use many open-weight models, which can be downloaded, hosted and audited. Yet we remain pragmatic about open source. Some uses justify a cutting-edge proprietary model, and an organisation is perfectly entitled to choose one. The decisive criterion lies elsewhere: being able to switch models without rebuilding everything else.
At Oxebo, we run open-weight models of American and Chinese origin, such as Gemma, GLM and DeepSeek, on Scaleway’s French infrastructure. Sovereignty is about where processing runs, which law applies to the data, and the ability to switch. The flag of the lab that trained the model comes second.
This position calls for an honest caveat. Every model carries the biases of its training, whatever its origin. That is why competence is part of sovereignty: you evaluate each model on your own documents and your own questions, and open weights are precisely what make that evaluation possible.
RAG, the architecture of sovereignty
What makes a model replaceable? With RAG, before generating an answer, the model retrieves the relevant passages from an indexed corpus, then writes from what it found: knowledge lives in the corpus, not in the model.
When knowledge lives in the sources, the model becomes a replaceable component.
An organisation whose knowledge has been absorbed into a model retrained on its data, or into a supplier’s proprietary memory, ties itself to that model. An organisation that owns its corpus, its index and its evaluation sets can switch models in a few days, test the new one on its own questions, and compare the results before switching.
It is also our conviction that when quality is built into search, a more compact, well-served model can be enough for many uses. It is tested use case by use case. Our choice to work with such models takes on a strategic dimension here: it reduces dependence on a handful of frontier models and their publishers. Everything starts with the source, and that is also what keeps the organisation free.