France’s €655M AI Plan Builds an Exit Option

France’s €655 million AI plan puts Mistral AI in front of citizens and ChapsVision inside the DGSI, where it is planned to replace Palantir. A chatbot and an intelligence migration appear to have little in common. The June plan paired them because both test how much control France can exercise over the technology its institutions use.

The June 16 plan groups citizen interfaces, controlled data environments and intelligence workflows. That pairing defines sovereign AI as the practical ability to change suppliers without abandoning critical operations. Whether France can do that remains unproven.

DGSI is the test of whether France can remove an incumbent

By naming Palantir, ChapsVision and a sensitive workflow, DGSI has made its planned transition more revealing than another funding round for a domestic model developer. Funding a French supplier does not show that an agency can remove foreign software after employees, data and decisions have accumulated around it. DGSI’s migration would.

To complete the transition, DGSI would have to reconnect data sources, reproduce permissions, preserve analyst workflows and carry security controls into the new environment without interrupting the work. It would also have to transfer operating procedures and integration knowledge that the June disclosure does not describe.

The transition remains planned. The June 16 report provides no migration costs, performance results or evidence about the difficulty of displacing Palantir. Until DGSI completes it, the plan establishes intent rather than a record of replacement.

The same June report says the French government plans to move computers from Windows to Linux, although it gives no timeline. By 2027, the government plans to replace Microsoft Teams and Zoom with the French-made, Outscale-hosted Visio. These projects reach the operating system, workplace communications and intelligence analysis—the layers where employees store work, permissions and routines.

The chatbot’s value depends on the workflow behind it

A citizen-facing assistant and an intelligence platform serve different users under radically different security conditions. Both sit between state information and a person trying to act on it. The assistant turns a request into an answer, while the intelligence platform turns data into analysis. A ministry makes either useful by connecting models to documents, tools, permissions, escalation paths and human decisions.

The June disclosure names Mistral but does not specify the chatbot’s retrieval sources, hosting arrangement, tool access or human override. A ministry that chooses a model still must make those four decisions. Together, they form a model procurement stack. The conversational window fits in a demonstration; the operating rules behind it determine how the institution acts.

Mistral has spent two years widening the deployment choices available behind that window. In 2024, the company released 3-billion- and 8-billion-parameter Les Ministraux models with 128,000-token context windows, which it designed for edge devices and offline assistants. In 2025, Mistral released a ten-model family under the Apache 2.0 license, including smaller Ministral variants. Customers can run those smaller models in controlled environments instead of relying solely on a remote frontier API.

Those releases give French ministries more options, but the June announcement does not identify which model, hosting arrangement or integration design the chatbot will use. A public assistant may prioritize broad access and service integration. An intelligence environment must prioritize controlled data, security and identifiable authority. Mistral supplies model choices; ministries still bear the expensive work of redesigning the processes around them.

AWS is learning to sell local control

In 2023, AWS announced a planned European sovereign cloud for governments and regulated industries, then launched it in January 2026 through a locally controlled EU parent company operated by EU citizens.

AWS is answering sovereignty requirements through its operating structure: where the cloud runs, who operates it and which organizational controls surround it. The launch does not establish whether French security agencies would accept that structure or whether it insulates workloads from every non-EU legal demand.

The comparison still matters. France can prefer ChapsVision for an intelligence workflow while allowing a foreign infrastructure provider to meet different requirements by changing how it operates. ChapsVision must fit DGSI’s workflow; AWS must satisfy the contracts and controls attached to a given workload.

If France required domestic ownership at every layer, it would narrow the supplier pool and assume French vendors could provide every component. Agencies can instead specify data residency, operator authority and substitution rights in contracts, making vendors compete over enforceable controls rather than flags alone.

Beijing, London and Paris intervene at different points

In April 2026, the UK government unveiled a £500 million Sovereign AI fund to invest in domestic startups and reduce dependence on technology from other countries. Wired reported on April 17 that the fund would begin with Callosum, whose software helps different chips work together. London is financing possible suppliers before ministries need them.

Beijing intervened later in the dependency cycle. The Wall Street Journal reported in March 2024 that Document 79, issued in September 2022, directed state-owned companies to replace foreign email, human-resources and other enterprise software. Reuters reported in October 2023 that government tenders showed spending on domestic replacements had increased since late 2022, beginning with computer equipment.

Beijing orders incumbents removed. London finances potential replacements. Paris combines investment in domestic AI with named foreign products it plans to displace, including Palantir, Windows, Teams and Zoom.

Agencies still have to carry data, permissions and operating knowledge into each replacement. Otherwise, governments may finance or mandate domestic products that their own departments cannot safely adopt.

A French vendor does not settle accountability

DGSI’s planned choice of ChapsVision would place a French supplier inside a sensitive system. The June disclosure does not specify who will own automated decisions, how analysts will challenge outputs, which privacy safeguards will apply or how the agency will handle security failures. Those omissions do not show that safeguards are absent; they mean the public documentation does not yet allow them to be evaluated.

The chatbot gives France a French voice; a completed DGSI migration would give it a door. France’s €655 million plan joins the visible interface to that exit, redrawing sovereignty as a handover document between Palantir and ChapsVision.