Do you need French AI to develop an application? Mistral, Dust, and your data

Mistral has signed with Microsoft, opened European regional access, and is targeting a gigawatt of computing capacity. What remains is to understand what that sovereignty covers when AI writes your software, and where the boundary lies between what stays with you and what leaves.

French AI for developing an application today means Mistral: it is the only French provider capable of running a coding assistant, at $14.99 excluding tax per month for Pro (mistral.ai/pricing, consulted September 2, 2026), and it opened European regional endpoints on August 11, 2026.

The misunderstanding around the name Vibe

A vocabulary clarification first: Mistral Vibe is the former Le Chat, renamed in May 2026 and repositioned as an agent for extended work, despite what the name suggests. The line that writes code is called Vibe for code and runs in the terminal, editor, or background, like other assistants on the market. Neither delivers a turnkey application to someone who cannot code. Maestro controls Mistral Vibe alongside Claude Code, Codex, and Cursor, with the option to switch mid-conversation.

What Mistral signed this summer

On July 29, 2026, Maddyness reported a multibillion-dollar agreement with Microsoft to provide computing capacity in Europe, a €1 billion investment announced by Samsung, a funding round of up to €3 billion, and an expected valuation of €20 billion, on top of more than €2 billion raised since 2023. The stated goal is one gigawatt of capacity by 2030, with a data centre in Bruyères-le-Châtel, Essonne. On August 11, 2026, Mistral opened regional endpoints in Europe and the United States and presented itself as the only European lab offering both a choice of processing region and a contractual service level.

Dust builds agents, not applications

The other French name mentioned in these conversations is Dust, a platform where teams create agents connected to internal tools: Notion, Slack, or a document workspace. It is useful for connecting AI to what your business already knows, and Maestro does not replace it. Dust delivers agents that live within its platform; the software still needs to be written elsewhere: the specification, code, and application that runs without it come from another tool. We do not list its prices or server locations: the pages consulted on September 2, 2026 did not provide them.

Where your data goes when an assistant writes your software

Two flows run in parallel and are often confused. The first is your project: with Maestro, code, documents, and version history stay on your Mac, keys are in macOS Keychain, and no server on our side sees your files; online backup is optional, to your own account. The second flow is what the assistant sends to the model provider to reason, and that does leave. With a Mistral regional endpoint, it stays in Europe. With a personal Claude account, Anthropic documentation consulted September 2, 2026 indicates retention of up to five years depending on a setting, versus thirty days otherwise, and exclusion from training under commercial terms. The details of what stays with you depend on that boundary.

What sovereignty does not cover

A European model does not review your code for you. A review of the ten leading vibe coding tools reports that 40–62% of generated code contains vulnerabilities (Blog du Modérateur, August 14, 2026), and that proportion does not depend on the provider's flag. In August 2026, free-mode Mistral Vibe delivered code that did not compile; we tightened the method instead of changing models, and published the account here. Mistral also retains an advantage Maestro does not have: deployment on a bank's or local authority's infrastructure, with audit logs and enterprise authentication that can be defended before a security committee.

Ask the question in the right place

Restate your requirement before choosing a provider. If it concerns processing location, ask for the region and service level in writing; a European regional endpoint answers that. If it concerns ownership, look at where the code lives once written and what happens if the tool loses its engine. If it concerns responsibility, read what the CNIL's note says about agents, and who answers for what. These three requirements call for three distinct answers that a single flag cannot cover.