Scaleway: our review

Scaleway hosts APIs for open language models — Mistral, Qwen, Llama, Gemma and about twenty others — billed by the token consumed. It's a French cloud provider, a subsidiary of the Iliad group, which builds its AI offering on its own network of data centres rather than on infrastructure rented from an American giant.

What you can plug into it

An application that needs a language model — a conversational agent, a summarising tool, an internal search engine — calls the Scaleway API the way it would call another provider's: the same request logic, a choice of model. The bill follows actual usage, per million tokens read and written, rather than a fixed subscription. A batch of one million tokens is offered on sign-up, enough to compare several models before committing a budget. Beyond text, the catalogue covers vision (reading an image), embeddings (indexing documents) and audio transcription.

Sovereignty: what's proven, what's promised

This is Scaleway's central argument, and it's worth splitting in two. What's established: the models run in a data centre located in Paris — the publisher notes this "may change in the future" — and Scaleway states it does not read, collect or reuse the content of prompts and responses generated by the APIs. What remains in progress: the SecNumCloud qualification, the one the State and operators of vital importance require, has not been granted to date. Scaleway launched its application in late 2025 and presented it as an end-of-year goal; by mid-2026, the publisher still describes it as "under evaluation." The ISO 27001 and HDS (health data hosting) certifications, on the other hand, are secured. The nuance matters: hosting in France is not the same as a SecNumCloud qualification, and confusing the two exposes an SME to recommending a service it cannot yet use for certain public contracts.

Where it falls short

Three reservations before committing. First, there's no interface: Scaleway sells an API, not a finished product. Without a developer to integrate it, the tool is useless. Second, the SecNumCloud certification is still missing, which rules Scaleway out of specifications that formally require it today. Finally, per-token billing shows no documented cap: a badly written loop or an overly long prompt can run up the bill with no known automatic alert — a point to watch from the first tests, as with any API of this kind.

Our recommendation

If you have a technical team and want an open model hosted in France, Scaleway is a sound choice, notably for building an in-house AI agent. If your constraint is an immediate SecNumCloud certification, wait for the qualification to complete or turn to a host already certified. And if you're looking for a ready-to-use product rather than a technical building block, this isn't the right tool: our automation services suit that need better.