Sovereignty & compliance

LLM gateway in France: the French options

Published 2026-11-18 · 9 min read

An LLM gateway in France can mean three things: a gateway run by a French company, models served from data centres in France, or a French qualification such as SecNumCloud. Three gateways are run by French companies — Frontière AI, Eden AI and LowRouter — and they differ on where the models run. You can also call French providers directly: Mistral AI, Scaleway in Paris, OVHcloud in Gravelines, or Cloud Temple, whose LLM service is hosted and operated in France. Or run a gateway yourself on a server in France, with Etalab's OpenGateLLM or LiteLLM. On Frontière AI, 7 models run entirely in France today, and every response names the provider that served it.

What "an LLM gateway in France" can mean

An LLM gateway is a single API in front of several large language model providers: one key, one format — usually OpenAI-compatible — and one bill, while the gateway picks the provider, retries elsewhere when one fails and tells you which one answered. Our guide to the European AI gateway covers the category across Europe. Here the question is narrower: what does "in France" add, and which options actually meet it?

"In France" can mean three different things, and vendors rarely say which one they mean:

  • A French company. The operator is incorporated in France and answers to French and EU law. This decides who can compel it to hand over data. The US CLOUD Act reaches companies under US control wherever their servers are; a French company sits outside its direct reach — not outside every foreign law. Our GDPR vs. CLOUD Act guide explains the mechanism.
  • Models served from data centres in France. The provider that runs the model does so on servers located in France. This is a fact about place, not control: the Paris region of a US cloud is in France but under US control, and a French provider's service can run outside France.
  • A French qualification. SecNumCloud, the cloud qualification of ANSSI, France's national cybersecurity agency, adds criteria of protection against non-EU law. It is what French public bodies and operators of vital importance are often asked for. It qualifies a specific service, not a company, so "our company is SecNumCloud" means nothing until you know which service is on ANSSI's list.

The three combine differently from one option to the next. A gateway can be French while routing to models in Germany or the United States. A provider can run its models in Paris without any of its AI services being qualified. The comparison below says which of the three each option meets, from its own pages.

The options in France, compared

Eight services come up when you look for an LLM gateway or an LLM API in France. They fall into three groups: gateways run by French companies, French providers you can call directly, and gateways you host yourself. The facts come from each company's own pages, read on 18 November 2026 unless a line says otherwise.

OptionTypeCompanyWhere the models runHow it is priced
Frontière AIGatewayFLEECE AI SASU, FrancePer model: 7 in France, the other sovereign models elsewhere in Europe, fast-access models at US providers — each labeled before you call itMargin in the per-token price, VAT included, no top-up fee
Eden AIGatewayEden AI SAS, FranceAbout 1,000 models from 80+ providers worldwide; a separate EU endpoint refuses non-EU modelsPer token, "a small markup around 5%" (its own comparison)
LowRouterGatewayCarbonifer SAS, Nantes, FranceEU and US providers; a region pinned in the model ID fails rather than leaving itUpstream price, no per-token markup, a fee on each top-up (checked 12 October 2026)
Cloud Temple LLMaaSProviderCloud Temple, French companyHosted and operated in FrancePer token
Mistral AIProviderMistral AI, ParisEU endpoint covering EU and EFTA data centres, no single country namedPer token
Scaleway Generative APIsProviderScaleway, FranceParis onlyPer token
OVHcloud AI EndpointsProviderOVHcloud, FranceGravelinesPer token
OpenGateLLMSelf-hosted gatewayEtalab project, MIT licenceWherever you run itFree software; you pay your servers and providers

Frontière AI

Frontière AI is published by FLEECE AI SASU, a French company. It runs no GPUs: it routes to European inference providers through one OpenAI-compatible API, with the Anthropic Messages format and embeddings as well, and it labels every model by jurisdiction before you call it. "EU sovereign" models stay on European providers, failover included — the call fails rather than leaving Europe. "Fast access" models run at US companies and say so. Each response names the provider that served it in its headers. The section below lists which models run in France. Prepaid from €10, VAT included, open to individuals as well as companies; per-model prices are on our pricing page.

Eden AI

Eden AI is the other French gateway with a large catalogue. Its site describes one OpenAI-compatible API to around 1,000 models from more than 80 providers, with routing and fallbacks, "zero data retention by default", SOC 2 and ISO 27001 certification and a data processing agreement (edenai.co). Its dedicated EU endpoint answers only with EU-hosted models: "Non-EU models are refused, never billed." Its own comparison of European gateways presents Eden AI SAS as "a French company under EU law" and its pricing as the provider's rate plus "a small markup around 5%" (Eden AI comparison). The difference with us is the catalogue and the default: Eden AI reaches US closed models as easily as European ones, and Europe is something you choose by using the EU endpoint; on Frontière AI the jurisdiction is decided per model.

LowRouter

LowRouter, published by Carbonifer SAS in Nantes, passes the upstream provider's price through with no per-token markup and charges a fee on each top-up instead. It routes to European and US providers alike, and a request without a region preference goes where its router picks; pinning a region in the model ID makes the request fail rather than leave that region. On the same model at the same provider — Qwen3 235B at Scaleway in Paris — it is cheaper than we are, fee included. Our LowRouter comparison has the figures and sources.

Cloud Temple LLMaaS

Cloud Temple is a French cloud provider that also sells language models as a service. Its documentation says the LLMaaS infrastructure "is hosted in France and operated in France by Cloud Temple, a company governed by French law", that the API is OpenAI-compatible, and that the infrastructure is qualified SecNumCloud 3.2 (Cloud Temple LLMaaS FAQ). It is a provider with its own catalogue rather than a multi-provider gateway. If a SecNumCloud qualification is a hard requirement, it is the option to examine first — check the exact scope against ANSSI's list of qualified services before relying on it.

Mistral AI, Scaleway and OVHcloud: going direct

If the models you need come from one provider, call that provider directly: it is cheaper than any gateway and one contract is simpler than two. Mistral AI, a Paris company, serves mainly its own models; its EU endpoint covers EU and EFTA data centres without naming one country. Scaleway runs its Generative APIs in Paris only and does not store prompts, except requests that cause errors or abuse, which it may keep for up to two weeks. OVHcloud runs its AI Endpoints in Gravelines. Both Scaleway and OVHcloud are French companies serving open-weight models — Llama, Qwen, Mistral, GPT-OSS and others — and neither of these AI services is SecNumCloud-qualified. Our list of European inference providers compares their catalogues and terms.

OpenGateLLM and LiteLLM: running your own gateway in France

You can also run the gateway yourself, on a server in France. OpenGateLLM is an open-source project published by Etalab under the MIT licence; its repository describes it as a "production-ready API gateway for Generative AI" and an alternative to LiteLLM, TensorZero and OpenRouter, built for self-hosted infrastructure. LiteLLM is the most widely used proxy of this kind. Self-hosting gives you full control over routing and logs; you then sign every provider's contract, hold their keys and operate the service. The two approaches combine: a self-hosted LiteLLM can use Frontière AI as one upstream, as our LiteLLM guide shows, tested end to end. Our comparison of open-source AI gateways covers the alternatives.

Which models run in France on Frontière AI

Today, 7 of our 16 EU-sovereign models run entirely in France, at Scaleway in Paris, with OVHcloud in Gravelines as fallback where one is listed:

Mistral Large 4 and Mistral Small 4 119B are served by Mistral AI, a French company, on its EU endpoint — in Europe, but without a guarantee of France.

The other 7 sovereign models run at European providers outside France, in Switzerland and Germany; our list of EU-hosted AI models gives the provider and country of each, and every model has its own page in the catalogue.

Two more facts complete the picture. The gateway itself runs in the Paris region of Vercel, a US company: requests transit through it, and Vercel keeps request logs but neither prompts nor responses. And Frontière AI does not write API prompts or completions to its database, only token counts. Each response carries the provider that served it in the x-routed-provider header and whether it is sovereign in x-sovereign, so where a request ran ends up in your own logs. Our privacy policy lists every sub-processor.

How to choose

  • You need SecNumCloud. None of the three gateways claims it. Start with Cloud Temple and check the qualified scope on ANSSI's list.
  • One model family is enough. Go direct: Mistral AI for Mistral's models, Scaleway or OVHcloud for open-weight models served in France. It is the cheapest route and the simplest contract.
  • You need several models, one key, and to know where each request ran. Use a gateway. Frontière AI if you want the jurisdiction decided per model and named on every response; Eden AI if you also need US closed models behind the same API; LowRouter if upstream prices matter most and you will pin regions yourself.
  • You want full control of routing and logs. Run OpenGateLLM or LiteLLM on a server in France, and connect the providers yourself.

Whatever you pick, ask the same three questions: which company operates it, where the model runs, and what each party keeps. The answers are in each vendor's legal notice, documentation and sub-processor list, and the ten questions in our European AI gateway guide go further.

For the wider view, read what sovereign AI means in practice and what EU-hosted AI models really means.

FAQ

Is there an LLM gateway based in France?

Yes. Frontière AI (FLEECE AI SASU), Eden AI (Eden AI SAS) and LowRouter (Carbonifer SAS, Nantes) are French companies that put several model providers behind one OpenAI-compatible API. They differ on where the models run: Frontière AI decides the jurisdiction per model and names the provider on every response, Eden AI reaches about 1,000 models worldwide with a separate EU endpoint, and LowRouter routes to EU and US providers at upstream prices.

Which LLMs run in France?

Scaleway serves open-weight models from Paris, OVHcloud from Gravelines, and Cloud Temple says its LLM service is hosted and operated in France. On Frontière AI, 7 models run entirely in France today: Qwen3 235B, Qwen3.5 397B, DeepSeek V4 Flash 0731, Llama 3.3 70B, Qwen3.6 35B A3B, GPT-OSS 120B and Mistral Small 3.2 24B. Mistral AI is a French company, but its EU endpoint covers EU and EFTA data centres without naming one country.

Does a French LLM gateway keep my data in France?

Not automatically. The company that runs the gateway, where the gateway is hosted and where each model runs are three separate questions. Frontière AI is French, its gateway runs in the Paris region of Vercel, a US company, in transit only, and each model runs at the provider named on its label, in France for some and elsewhere in Europe for others.

Which LLM option in France is SecNumCloud-qualified?

None of the three gateways claims it, and neither Scaleway's Generative APIs nor OVHcloud's AI Endpoints is SecNumCloud-qualified. Cloud Temple says its LLMaaS infrastructure is qualified SecNumCloud 3.2; check the exact scope on ANSSI's list of qualified services before relying on it.

Which is the cheapest way to call an LLM in France?

Going directly to the provider that serves your model is usually the cheapest. Among gateways, LowRouter passes the upstream price through and charges a fee on each top-up, and Eden AI describes a markup of around 5%. Frontière AI builds a fixed margin into its per-token price, VAT included, with no top-up fee; on the same model and provider, LowRouter is cheaper.

Ready to try it?

Create an account and call any model in the catalog in minutes.

Create an account