Sovereignty is not a conference theme, it is a question of survival for the company from the moment its AI model can be suddenly disconnected from the outside.
Friday June 12, in a few hours, the most advanced AI model on the market disappeared. Not a failure, not a bug. The decision taken by the Trump administration to prohibit any foreign national from accessing Anthropic’s Fable 5 and Mythos 5 models sends us a clear message: sovereignty is not a conference theme, it is a question of survival for the company from the moment its AI model can be suddenly disconnected from the outside. This decision is a wake-up call about Europe’s dependence on foreign LLMs and should encourage business and political leaders to implement “survivatech” strategies.
Last week, Paris celebrated ten years of VivaTech. But when access to AI can decide a company’s survival, it’s no time to celebrate. No other industry would accept such exposure. A refinery that only sources its supplies from a single supplier, without any stock, would not survive the first shortage. Yet this is how most companies run their AI today.
We must first get out of the confusion: “sovereignty” has become a selling point, boxes to check in a call for tenders to ensure a sovereign cloud or a local LLM provider… In reality, this does not protect against anything. A company can very well host everything locally and have no operational sovereignty over its AI, if it is someone else who built it, who understands it and holds the keys. Another source of confusion: opposing the United States and Europe, performance and independence. This opposition is an illusion. AI laboratories produce extraordinary models, hyperscalers have built essential infrastructures, sovereign clouds meet real needs. They are partners, not adversaries. The choice businesses face is not which side to take. It’s about preserving their freedom of choice.
In reality, sovereignty operates on three levels: infrastructures (where are they, who operates them?), AI systems (who built them, what power of control and governance do the company’s employees have over them?) and economic value (who captures it, the company or its AI supplier?). These three levels of sovereignty are naturally dependent on each other: if we are not free in our infrastructure, we are not free in the construction of AI models. And if we cannot build them freely, we cannot appropriate their value, we can only consume it.
Today, competitive advantage is based on frugality and ownership of intelligence. In a context of high processing costs and energy dependence, an AI model at the right scale, deployed where it can be controlled, is more effective than a very large rental model. If one praises one’s intelligence, one runs the risk of the landlord changing the locks during the night and being the last to know. The intelligence that AI produces for the company (reasoning, logic, expertise, etc.) must remain its property.
The final destination of an AI strategy is optionality. An AI model must be an exchangeable resource, not a load-bearing wall… Concretely, the discipline consists of two stages: prototyping on cutting-edge models, deploying in open source. We design with the best models of the moment, often closed, which stay three to six months ahead, then we switch production to models that we own and govern. Optionality is now a necessity. Technological sovereignty and financial sovereignty are not separate battles: we need both to achieve the ultimate goal, intellectual independence, which allows you to be free to choose your models, your suppliers, your infrastructure at each stage of AI deployment.
This is what a business can do alone. The most difficult thing is collective. Countries that feel they are lagging behind cannot, each on their own, build a cutting-edge laboratory: the gap in resources is too great. They must join forces and pool their resources to build models that belong to them. The only license that makes this shared ownership possible is open source. An AI “for good” belongs to everyone and cannot be disconnected by anyone. The countries that build it are masters of their survival the day the next model is disconnected from the outside.
