Introduction
Notes on AI providers for UK and European organisations that want to know where their prompts and documents go, which country's law applies, and whether they could run a model themselves.
This page currently covers Mistral AI; more providers may be added.
How we describe providers:
We summarise what each provider is, where it is based and who owns it, what it says about sovereignty, and what you should check yourself. Claims are labelled [law], [government statement], [press report] or [vendor claim], and vendor claims are written as “X says…”.
These are notes, not endorsements or rankings.
On the US CLOUD Act: the law lets US authorities require providers subject to US jurisdiction to disclose data in their “possession, custody, or control”, wherever it is stored [law]. A non-US company with no US presence is generally outside its direct reach. Exposure can still come through US subsidiaries or offices, a US parent company, US sub-processors, and your own choices (for example, picking a US region or a US-owned host). So the jurisdiction of the company usually matters more than where the data sits. We don’t describe any provider as “immune”.
Facts were checked against providers’ own pages, filings or reliable reporting on the date below. Figures and ownership change. If you spot an error, please tell me.
Last checked: 28 September 2026.
Mistral AI - Making frontier AI open to all
Mistral AI: French AI models, open and commercial
Mistral AI is a Paris-based company that builds large language models and sells AI products, including its assistant Vibe (formerly Le Chat), an API platform and on-premises deployments.Many of its models are released as "open weights": you can download them and run them on your own hardware or with a host you choose. That includes its most capable model, Mistral Large 3, released under the Apache 2.0 licence in December 2025. Self-hosting gives you the most control, but a model this size needs substantial GPU hardware.If you use Mistral's own services, Mistral says your data is hosted in the EU by default. Data goes to the US if you choose its US API endpoint, and some features may temporarily transfer data outside the EU to its listed sub-processors [vendor claim]. Mistral also has a US subsidiary. So "European company" doesn't settle the jurisdiction question on its own. Check the plan, endpoint and features you actually use.
About Mistral: location and founders
Where it's based and who owns it: Mistral AI is a French company headquartered at 15 rue des Halles, Paris. It was founded in 2023 by Arthur Mensch (CEO), Guillaume Lample and Timothée Lacroix. It is privately held and backed by venture and strategic investors. In September 2025 the Dutch chip-equipment maker ASML led its €1.7 billion Series C with €1.3 billion, for a stake of about 11%, and was reported to be its largest shareholder [press report]. Mistral also has a US subsidiary, Mistral AI Inc., in Palo Alto, California [vendor claim: Mistral's own terms].
About Mistral: licensing
Licensing: Mistral uses a mixed model. It releases many models, including Mistral Large 3 and the Ministral 3 family, as open weights under the Apache 2.0 licence. Other models and products are available only through its paid services or commercial agreements. Check the licence of the specific model you plan to use.
About Mistral: editions and paid tiers
What it says about sovereignty [vendor claim]: Mistral promotes open weights, EU hosting by default and on-premises deployment as ways to keep control. Mistral says its flagship Mistral Large 3 is released under Apache 2.0, so you can download, fine-tune and self-host it without a subscription. You pay if you use Mistral's hosted API or Vibe plans, its packaged enterprise and on-premises software, or its support and custom training. Some tools and models are available only through those paid routes.What to check yourself:
- Which route you'll use: Vibe, the API (EU or US endpoint), a third-party cloud such as Microsoft Azure or Amazon Bedrock (both US-owned), or self-hosting. Each route has a different operator and jurisdiction.
- The current sub-processor list in Mistral's Trust Center, and whether you can switch off features that send data outside the EU (Mistral says Enterprise customers can switch off some [vendor claim]).
- Whether your prompts are used for training on your plan, and how long data is retained.
- If you self-host, whether you have the hardware and skills. Mistral says a single 8×H100 or 8×A100 node can run Large 3 in a compressed format [vendor claim].
About Mistral: features
What it is (features): multilingual text generation, code models (Codestral), image understanding, document understanding and OCR, an Agents API for multi-step workflows, the Vibe assistant (called Le Chat until May 2026), and on-premises deployment for organisations that want to run models on their own infrastructure.
