French AI for enterprises: what sits behind the word “sovereign”

We notice (and we welcome it) that more and more companies care about the sovereignty of their tools, especially in AI. And France has a particularly rich ecosystem of services firms and software vendors.

Since Snowden, we have known that any data within their reach is watched by the US administration. Since Trump's second term, the executive has chosen to cut access in a targeted way to US services. That trend is likely to intensify, and from then on it puts companies at significant risk.

Choosing a vendor founded by French people inside a US structure does not protect you. Likewise, if your company's data is hosted on a US cloud or sent to a US AI provider, the problem remains the same.

So we wanted to check which French companies were in a position to guarantee the sovereignty of their products and protect their customers.

The three questions that matter, and what they cost you

Who you contract with, and who owns the capital

Two checks, in that order, because the first one is eliminatory.

First, which company signs the contract. A French team, French founders and an office in Paris say nothing about the entity that sells you the service. A group can perfectly well have a French subsidiary that employs the staff, and a US company that commercialises the product and declares itself the data controller. In that case, you contract with a company under US law, and the rest of the analysis becomes irrelevant: it does not matter where the servers sit, the company that holds your data falls under a foreign jurisdiction. That takes three minutes to check on the legal notice or the /legal page.

Then, who owns the capital. Here, a company can be perfectly French, incorporated, with no foreign parent, and still have majority extra-European capital after a few funding rounds.

What that changes for you: nothing today, on your data. But whoever controls the capital controls what comes next. A sale, a change of headquarters, dropping the sovereign line if it costs too much, shifting the infrastructure to a provider the new shareholders know better: these are decisions taken at the board, not in the product team. The risk therefore sits on a three-to-five-year horizon, and it is worth knowing without turning it into a drama.

Who hosts your data

This one is the most important of the three, and by far the most damaging.

A host under US law remains subject to US law, including when its servers are in Paris or Frankfurt. The CLOUD Act, adopted in 2018, lets US authorities demand from a US company the data it holds, including when that data is stored abroad. Server location is not a defence.

The other risk took longer to become credible: the service being cut off. Since 2025, several unilateral US decisions have suspended access for political reasons, with no useful notice for the European customer.

What makes this criterion frustrating is that it is entirely avoidable and it brings the user nothing. OVHcloud, Scaleway and 3DS Outscale host databases and application servers as well as AWS or Google Cloud. Nobody in your company will ever see the difference in use. A French vendor that hosts with a US hyperscaler has made an engineering-comfort choice, not a product choice.

Who runs inference

Inference is the moment when the AI is executed, and turns an input (a query, documents, and so on) into an output (a reply to a message, generated documents, generated content, and so on).

Concretely: company data that travels with your query goes to whoever runs the model.

Here the vendor has a mitigating circumstance. The best models on the market are American, and the one that plugs in GPT or Claude often delivers a better service to its users. The choice is entirely defensible. It is simply rarely written in black and white on the product pages.

And one question remains, the one almost nobody asks.

The table

Fourteen vendors, in alphabetical order.

Legend. ✅ the criterion is met by default. 🟠 it is reachable, but it takes a move: a written request, a contractual clause, a manual opt-out provider by provider. ❌ it is not met. ? the information is not public.

The vendors, one by one

We sorted them into five groups.

A useful clarification before we get into the details: these three criteria say nothing about product quality or the seriousness of the team. A vendor can be excellent at its job and still end up in the last group.

We also found no lies in the commercial claims: vendors that did not make sovereign choices do not pretend they did, and they own their positioning.

Group 1: European hosting and European inference

Ask This Guy. ATGU SAS, 100% of the shareholding held by French individuals or personal holding companies. Data on dedicated OVHcloud servers in Gravelines, redundancy in Frankfurt, plus our own GPUs at Verda, a Finnish company. Two organisation-level inference policies, “EU only” or “Worldwide”, changeable from the admin console. This is our product; we come back to it below.

Sources: our approach to sovereign AI, managing AI providers.

IA-Souveraine.fr. Published by DRI (Digital Rural Informatique), a SAS created in 2001 in Le Mans, controlled by a French holding company. This is the most vertically integrated file in the panel: DRI is not a software vendor that rents cloud, it is a twenty-five-year-old host that added an AI layer. The independence sits on the infrastructure, not on the models, which remain open weights published by others, Google, Meta or Mistral depending on the case. Its legal notice says it in so many words: it is “publisher and host” of the platform. Own datacenter in Le Mans, own network and autonomous system number, ISO 27001 and ISO 14001 certifications, HDS compliance available, hosting exclusively in France for the sovereign offer.

On inference, the sovereign offer is clear: open-weight models are run on DRI's machines, in France, therefore by a French company. And unlike the multi-model platforms in this list, there is nothing to lock down, because there is nothing to exclude. Sovereignty is the product, not an option to switch on. One reservation, though: the platform advertises more than 80 models in total, including “reference commercial models” that are never named. We read those as a distinct offer from the sovereign one, which is consistent with how DRI presents the two, but the question is worth asking if you subscribe.

Sources: legal notice, privacy policy, sovereignty page, DRI launch press release.

LightOn. A public limited company listed on Euronext Growth Paris since November 2024. Models developed in-house, deployment possible on-premise and as far as air-gap for classified environments. Information we could not find: LightOn nowhere names the host of its SaaS offer, presented as a “sovereign European cloud” with no further detail.

Sources: developer privacy policy, deployment options, Euronext Growth IPO sheet.

Mistral AI. The only vendor in the panel that produces its own models and operates its own datacenters: Bruyères-le-Châtel and Les Ulis in Essonne, Borlänge in Sweden, with an Île-de-France campus announced at 1.4 GW at full capacity. Capital is described by the company as “over 60% European”, with ASML as the largest shareholder at around 11%, alongside Bpifrance, US funds and the Emirati sovereign wealth fund MGX.

Two nuances, which take nothing away from the industrial depth of the file. And let's say up front that we only know them because Mistral publishes them: its Trust Center is one of the most complete in the panel, with a tab that names its subprocessors. Many vendors on this list expose nothing comparable, and that is precisely why their write-ups are shorter. That Trust Center therefore cites two US companies in the chain: Microsoft as the cloud infrastructure provider for European Studio customers, and CoreWeave as an inference compute provider. And the European lock goes through a written request: Enterprise customers have the features that involve a transfer outside the EU disabled at organisation level, by asking Mistral.

Sources: where is my data stored?, regional inference documentation, Trust Center, Subprocessors tab, infrastructure and open models announcement.

Wikit. Wikit SAS, Lyon, specialist in the internal support assistant. Hosting on OVHcloud in France. Wikit names seven third parties in total, which is already rare, and above all it publishes the best documentation in the panel on the model layer: a table that states, provider by provider, whether data stays in the EU, with a row that frankly answers “no” for OpenAI. Nobody else does that. Provider choice is set at the project level; no customer-facing admin console is documented.

Source: privacy and security, Wikit documentation.

Group 2: The infrastructure is there, the door stays open

Craft AI. Craft.ai SAS, chaired by a French holding company. The file illustrates the “listed is not mandated” trap: the annex to its DPA authorises five clouds, including 3DS Outscale, OVHcloud and Scaleway, while production runs on AWS and Scaleway. In other words, the French building block exists, contractually, but it is never presented as a catalogue option. On the model side, of the eight authorised providers, five are American (AWS Bedrock, Google Vertex AI, Microsoft AI Foundry, OpenAI, Voyage AI), one is Canadian (Cohere), and only two are European: Mistral AI and Scaleway. A fully European inference setup is therefore possible on paper, but it means sticking to those two lines.

Sources: DPA annexes, annex B for actual hosting and annex C for authorised subprocessors, additional terms on foundation models.

Delos. The most frustrating file in the panel. Delos Intelligence hosts all data with Scaleway in the Paris region, with triple replication, ISO 27001 certification and a trust center. On hosting, nothing to criticise, and this is the standard offer, not a paid option.

Then it gets more complicated. The security page states that models are “hosted exclusively in the European Union”. But the catalogue on display is GPT 5.2, Claude Opus 4.5, Gemini 3 Pro and Mistral Large 3: three closed models out of four, therefore necessarily operated by their US publishers, wherever the server runs. And the interface lets the user pick their model, with nothing documenting the ability for an administrator to restrict the list. To its credit, and this is the only serious mitigation when models are closed: the security page mentions processing on pseudonymised data before transmission to the model. What remains is to get a precise answer on what is pseudonymised, and at which point in the chain.

If that administrator setting existed, Delos would become one of the most sovereign solutions on the market without touching its infrastructure.

Sources: security page, privacy policy, Assistant product page for the model list, fundraising, FrenchWeb.

Group 3: Backed by a US hyperscaler

Ask for the Moon. Franco-German capital, €2.5m raised in March 2024 from Matterwave Ventures and UI Investissement, with no extra-European investor identified: that is the strength of the file. Services are hosted on Microsoft Azure in Paris, and inference is not documented publicly. One point to have confirmed in writing: their own data-protection page states that data at rest is not encrypted, which is unusual.

Sources: protecting your data, profile and fundraising, Maddyness.

Dust. Permutation Labs SAS, a French company. The capital, however, is majority American. Hosting rests exclusively on Google Cloud, only the region varies, and no documented alternative exists: that is the one point in the file that cannot be fixed with a setting.

Two things to put to their credit. Dust is very transparent: legal notice, group entities page, security measures, dated subprocessor list, specific terms for foundation models, everything is public. And it is the only one to have publicly contractualised a per-provider opt-out right, self-serve for inference. The mechanism is nevertheless a blacklist, not a whitelist: compliance is obtained by disabling five providers to keep one, and in a strictly European mode, only Mistral remains.

Sources: security measures, subprocessor list, foundation model terms, entities and affiliates.

Outmind. Outmind SAS, Palaiseau. Processing by default in the EEA, host AWS EMEA in Paris, dedicated instance per customer with its own encryption key. Inference is not documented. Another point of attention: user authentication goes through Clerk, a San Francisco company, and nothing indicates that it is optional.

Sources: security page, privacy policy.

Group 4: Too little public information to conclude

For these three, we simply did not find enough public material. A thirty-minute conversation would move all three write-ups, one way or the other.

Cominty. Cominty SAS. The product is presented as “agnostic”: Docker, GCP, AWS, OVH. That is likely a choice left to the customer, but default hosting is not documented, and no model provider is named. On security, by contrast, Cominty says more than the group average: SOC 2 Type II, GDPR compliance, AES-256 encryption at rest and TLS 1.2+ in transit. Be careful not to conclude too fast: agnosticism can just as well cover an organisation setting as a free user choice.

Sources: cominty.ai website, trust center.

Limova. Brand operated by Kainotom.ia SAS, in Nice. The privacy policy claims hosting “exclusively on servers located within the European Union”, but the only subprocessor named in the entire documentation is Google Firebase, a Google LLC service. The documentation describes neither a model provider nor a processing chain, for a product whose agents read emails and take phone calls. Also worth noting, on form: the legal notice publishes neither the SIREN, nor the RCS number, nor the host, even though the law requires them.

Sources: legal notice, privacy policy, the Limova model, Maddyness.

Prisme.ai. Prisme.ai SAS. The distinctive point changes the whole reading: the platform is deployed in self-hosting or the customer's private cloud, via Kubernetes, on OVHcloud as well as on AWS or Azure. Sovereignty is therefore entirely delegated to the customer: whoever deploys on OVHcloud gets a French stack, whoever deploys on AWS does not, and the vendor guarantees neither. No model provider is named, no certification is mentioned, and the capital is not published.

Sources: Prisme.ai documentation, product pages.

Group 5: The false friend

Slite. Slite is regularly listed among the French alternatives: French founders, French team, Paris office, a knowledge-management product with AI search. Its own legal page is nevertheless unambiguous.

The contracting entity and data controller is Slite, Inc., a Delaware company, based in San Francisco. Slite France SAS “does not market the Service and does not act as the contracting entity for customers: it provides employment and support services within the group”.

Production data is hosted on Google Cloud, in Saint-Ghislain, Belgium. A European customer therefore contracts with a US company, on US infrastructure, with servers in Belgium. That is perfectly legal, the company does not hide it, and its security is solid, with SOC 2 Type II. Slite does not promise sovereignty.

It is also, among fourteen vendors, the only case of a company controlled from abroad.

Sources: legal information, privacy policy, security page.

A small extra reality check: where their own website is hosted

This one has no operational bearing. A marketing site does not hold your data, and handing it to a US provider commits nothing about the product architecture. But it is the only test you can run yourself, in thirty seconds, before the first meeting. And it says a little about the house mindset.

Here is what we were able to determine:

The three questions to ask in a demo

If you take only one thing from this article, it should be these three questions. They take ten minutes to ask and they call for verifiable answers.

1. “Who hosts our data, by name, and is there a European alternative to your standard offer?” A serious vendor names its host. Refusing to name it is an answer in itself.
2. “Can we choose, lastingly and simply, never to use non-European inference providers?”
3. “If we switch to a fully European setup, how many providers do we have left?” Compliance that rests on a single provider is a single point of failure.

What we did with this, at Ask This Guy

We built our product by trying to answer these questions in advance, and that is also why we wanted to write this article. Our offer is laid out in detail on our sovereign AI page.

Concretely: infrastructure and databases sit on dedicated OVHcloud servers in Gravelines, with redundancy in Frankfurt, and we operate our own GPUs at Verda, in Finland and Iceland, to run open-weight models ourselves. On inference, an administrator switches the whole organisation between two policies from a console menu: “EU only”, which allows only European providers and our self-hosted models, or “Worldwide”, which opens the model layer without ever moving the hosting. The setting can be changed at any time.

We do not claim this is the only good architecture. Several vendors on this list made different, perfectly defensible choices. We simply think a buyer should be able to observe, lock and prove, without having to negotiate.

The detail of our processing chain is documented publicly in our approach to sovereign AI and in our AI provider management.

Frequently asked questions about French and sovereign AI

What is sovereign AI for a company?

Sovereign AI is a solution whose vendor company, data host and inference operator all fall under European law, so that no foreign authority can demand access to the data or interrupt the service. The term has no legal definition and no official label: you therefore have to check the three links separately, in the vendor's public documents.

Is AI hosted in France necessarily sovereign?

No. Hosting in France with AWS, Microsoft Azure or Google Cloud is still operated by a company under US law, subject to the CLOUD Act regardless of where the servers sit. Data residency satisfies a contractual localisation clause, but it does not change the jurisdiction that applies to the provider. That is the distinction most often skipped in commercial pitches.

Is Mistral AI a French company?

Yes. Mistral AI is a simplified joint-stock company under French law, registered under SIREN 952 418 325, with no parent company outside the European Union identified. Its capital is described by the company as over 60% European, with Dutch ASML as the largest shareholder at around 11%, alongside Bpifrance, US funds and the Emirati sovereign wealth fund MGX.

Is Dust a French company?

Yes on the legal plane. The contracting entity is Permutation Labs SAS, a French company registered under SIREN 949 205 314, with no foreign parent, which its legal notice and group entities page confirm. Its capital, however, is majority American after two rounds led by Sequoia, and its infrastructure rests exclusively on Google Cloud.

Can you use ChatGPT or Claude and stay sovereign?

No, in the strict sense. GPT and Claude are closed models whose weights only OpenAI and Anthropic hold: they alone can run them, and the processing therefore falls under US law even on a European server. Genuinely European setups rely on models (open-weight or not) operated by a European actor, or on a vendor that produces its own models.

How do you check for yourself whether an AI vendor is actually French?

Open three pages, it takes about ten minutes. The legal notice or the /legal page gives the contracting entity and lets you spot a foreign parent. The privacy policy or the DPA annex names the host and the subprocessors. The SIREN number can be cross-checked for free on annuaire-entreprises.data.gouv.fr. If any of those three pieces of information is missing, that is already a result.

---

In short

Of fourteen vendors, thirteen are indeed European companies. Good news, but it is also the least discriminating point on the market. Everything turns on two engineering decisions.

Hosting, where only four vendors rely by default on a provider under European law, even though the choice costs the end user nothing. And control of inference, where almost nobody lets an administrator impose a European setup on the whole organisation, without a written request or a manual opt-out.

If you are preparing an enterprise AI project and you want to put these criteria against your own context: book a demo.