Caisse des Dépôts × Mistral AI: a strong political signal. But digital sovereignty is not won on day 1. The real test begins on day 2.
On May 4, 2026, Caisse des Dépôts announced a framework agreement with Mistral AI to structure its uses of AI. 19 subsidiaries pool their purchases, nearly 40,000 generative AI licenses are deployed at start-up, up to 100,000 users targeted over the duration of the contract, and an AI Factory is created to industrialize deployments for the benefit of the Group, public stakeholders and territories. The operation is part of the Digital Horizon 2030 plan, endowed with 18 billion euros.
Many praised a strong act. The observation is widely shared, without reservation on the intention. But the look deserves to be moved. Because the moment of signing is not the one that will decide the value created. The real test begins on day 2.
The real singularity of the operation is not the supplier
Before talking about day 2, let’s clarify what really changes this May 4. The choice of a French player to provide language models is not, in itself, a new signal: Mistral was already working with several major public and private accounts. What is new is the scale (100,000 users ultimately in 19 subsidiaries), the shared purchasing mechanism between Group entities, and the industrialization brick that constitutes the AI Factory backed by GPU computing capabilities. Olivier Sichel says it clearly in the official press release: digital sovereignty “must be supported by orders, particularly public ones”. Public procurement becomes an instrument of industrial policy. This is what deserves attention.
A conviction has been verified for three years in all the organizations that have put generative AI through the daily test: AI that truly transforms does not make noise. It is built in governance, in the orchestration of business flows, in the measurement of real gains. The signing of a framework agreement, as structuring as it may be, is only a starting point. And it is at the start-up that the pitfalls are concentrated.
Three conditions for day 2 to keep its promises
First condition: governance of uses. 100,000 users do not self-regulate. Without shared mapping of authorized use cases, usable data, and critical decisions requiring human supervision, the promise of sovereignty comes up against the reality of shadow AI. However, experience shows that ungoverned deployment in a complex organization produces, from the first months, a usage debt which is then very costly to recover. Controlling outbuildings applies as much to the exterior as to the interior.
Second condition: business orchestration. Licenses by themselves do not produce value. What this produces is the ability to integrate AI into coherent business flows: preparation of risk analysis in a public bank, instruction of financing files, documentary processing in a real estate company, support for users of a public service, automation of audit tasks. The real subject is no longer the isolated use of a co-pilot, but the orchestration of several agents in end-to-end processes. This is where value is concentrated, and this is where organizations still stumble.
Third condition: measurement and acculturation. Without real gain indicators (time saved, quality produced, sustained usage rate, user satisfaction) we will talk about symbols, not transformation. And without acculturation of the 100,000 people concerned, the ROI will collapse. It is the non-technological lever that decides everything: we have seen too many brilliantly equipped AI projects get bogged down for lack of having trained the teams in discerning use. Generative AI transforms the role, not the DNA. We still need to support this change over time.
Ripple effect, or collective delay effect?
Many want to see this agreement as a signal that will inspire other administrations, other communities, other public groups. This is a valid hypothesis. But it is conditional. If the first subsidiaries of the CDC Group demonstrate, within the next twelve months, a controlled deployment and documented gains, the ripple effect will be massive and structuring for the French ecosystem. If they get bogged down in endless pilots or low usage rates, the narrative will be reversed, and Operation CDC × Mistral could become, paradoxically, the opposite argument for those who hesitate to commit. The value of example is never acquired by signature: it is earned in execution.
It is the collective responsibility of all stakeholders involved (Caisse des Dépôts Group, Mistral, integration partners, users) not to confuse today’s momentum with tomorrow’s work. There is, in this agreement, the potential for a precedent that would make digital sovereignty credible, and desirable, for dozens of other public organizations. We still need to demonstrate this concretely.
Day 2 has begun
Day 1 is behind us. The signature, the press release, the first communications, the first conferences: all of this exists, and it is useful. But day 2, that of putting into production, industrialization, orchestration, measurement, has just opened. It is there, and nowhere else, that French digital sovereignty is at stake. Not in the ads. In uses.
