European AI Providers Beyond ChatGPT

European AI comes in two kinds: vendors with their own models, mainly Mistral in Paris and Aleph Alpha in Heidelberg, and platforms that run other vendors' models in German or European data centers, such as IONOS, STACKIT, T-Systems and OVHcloud. Two more routes are often forgotten: a US model like Claude or GPT on EU servers from AWS, Google or Microsoft, and a model on your own hardware. What decides is not the country of origin but six criteria: where the provider is based, where processing happens, the data processing agreement, training on your data, the models on offer and the pricing model.

To be clear up front: I work with Claude myself, a US model. The information on the providers comes from their own pages (as of 31 October 2026), not from a test of my own, except for Langdock (see below). I am a developer, not a lawyer, and this is not legal advice, just the operator's perspective. More on AI in everyday work is on my AI page.

Six criteria instead of a ranking

"German AI" and "AI made in Germany" sound like one property but hide six questions:

1. Seat: Where is the contracting party based, and who owns it?
2. Place of processing: Which data center, and with what exceptions?
3. DPA: Is there a data processing agreement under Article 28 GDPR, and how is it concluded?
4. Training: Are your inputs used for training, and can you switch that off?
5. Models: Own models, or open models from other vendors?
6. Pricing model: Per user, per token, or on request?

A provider can have an exemplary data center and still train on your chats by default. No single question is a verdict.

European AI providers with their own models

Mistral AI is based in Paris. Its chat is now called Vibe, formerly Le Chat, plus the API through Mistral Studio. According to the help center, data is hosted in the EU by default, in the US via the US API endpoint, and depending on the feature it can go temporarily to subprocessors outside the EU. On training: in Vibe, inputs and outputs are used for training by default unless you opt out; only Enterprise customers are opted out by default. The Data Processing Addendum is part of the agreement for commercial customers. You pay by subscription (Free, Pro, Team per user, Enterprise) or per token for the API. Some Mistral models have open weights under Apache 2.0; which ones is covered in Open Weight vs Open Source: Licenses.

Aleph Alpha from Heidelberg presented its model Kolibri on 5 October 2026: 78 billion parameters, about 3 billion of them active per token, German and English, trained by Aleph Alpha's teams in Germany, according to the company, and designed to run on infrastructure that customers control themselves. The weights are on Hugging Face under Apache 2.0; the stated minimum is, for example, two 80 GB A100s or one H200, which is data center equipment. Kolibri is also available through an API after sign-up; where it runs, what happens to the data and what it costs is not stated on the website. Important: Aleph Alpha and the Canadian company Cohere announced on 16 September 2026 that they had signed a binding business combination agreement. The combined company is to operate as Cohere, with headquarters in Berlin and Toronto; regulatory approvals are still pending.

German and European platforms for open models

The platforms run other vendors' open models in their data centers behind an OpenAI-compatible API, for companies building their own applications.

• IONOS AI Model Hub (IONOS Cloud GmbH, Montabaur): According to the docs, all inference runs in IONOS data centers in Germany; content is not logged, stored or used for training. The data processing agreement applies as part of the cloud contract without a separate signature. Models include Qwen, Mistral Small, gpt-oss-120b and Llama 3.3 70B. Billing is per token, for gpt-oss-120b 0.15 euros per million input tokens and 0.65 euros per million output tokens, plus VAT, business customers only.
• STACKIT AI Model Serving (Schwarz Digits Cloud GmbH & Co. KG, Bad Friedrichshall): STACKIT states that its companies, data centers and servers are located exclusively in Europe. According to the docs, requests are not stored and not used for training. There is a data processing agreement for the STACKIT Cloud Services. Models include gpt-oss-120b, Llama 3.3 70B, Gemma 4 and Qwen; billing is per token.
• T-Systems LLM Serving (T-Systems International GmbH, Frankfurt am Main): Open models run at Telekom in Germany, and according to T-Systems prompts are not stored persistently and never used for training. T-Systems also passes through frontier models (GPT-5 on Azure, Claude on Google Cloud or Azure, Mistral Medium in the EU, Gemini on Google Cloud worldwide) under each vendor's own terms. Data processing agreements are available, with a German contracting party. Billing is per token, self-service from 1,000 euros a month.
• OVHcloud AI Endpoints (group: OVH Groupe SA, Roubaix, France; contracting party for customers in Germany per the terms: OVH GmbH, Cologne): According to OVHcloud, hosted and run in the EU; your data is never used for training, and only what billing requires is kept. More than 20 open-weight models, billing per token. The data processing agreement is part of the contract with OVHcloud; under the AI Endpoints terms, OVHcloud has no knowledge of inputs and outputs and does not reuse them, except for the Batch API.
• Langdock (Langdock GmbH, Berlin) is not a model API but a workspace with chat, agents and workflows across several vendors' models, billed mainly per user and month, workflows additionally per workspace. According to its security page, the multi-tenant version runs on Microsoft Azure in the EU, as do most models (configurable), with no training on customer data. More in Langdock vs Claude: What I Use and Why.

What "sovereign AI" means here, and what it doesn't

First, the German platforms mostly run models from the US and China: gpt-oss from OpenAI, Llama from Meta, Gemma from Google, Qwen from Alibaba. An open model in a German data center sends nothing to its vendor, but it is something different from "AI made in Germany". Second, seat and infrastructure differ: Langdock is based in Berlin and runs on Microsoft in the EU, T-Systems passes frontier models through to their providers, most of them on Azure and Google Cloud. Third, owners can change, as the planned combination of Aleph Alpha and Cohere shows. STACKIT explicitly advertises that its European seat and data centers protect against access by US authorities under the CLOUD Act. How much that weighs for your data is a question for your data protection officer.

The third route: a US model on EU servers

If you want Claude or GPT with processing in the EU, you go through AWS Bedrock, Google Vertex or Microsoft Foundry; for Claude, Bedrock and Vertex offer an EU option, while Foundry, according to Anthropic, offers only "global" or a US data zone. Anthropic's own docs only offer "global" and "us" as processing locations, and the contracting party is a US company in every case. Which Claude models have an EU profile on Bedrock is covered in Claude on AWS Bedrock: EU, GDPR and Cost. How this route compares with a business plan and your own server, I cover in Private LLM: Vendor, EU Cloud or Own Server?.

The fourth route: local

An open-weight model on your own hardware needs no provider and no DPA with one, but local does not automatically mean compliant; more in Local AI and Privacy: Cloud or Server?, and the hardware question in Your Own AI Server: Hardware and Cost.

How I decide for myself

My team of fourteen AI employees runs on Claude with a DPA day to day. I tried Langdock and stayed with the folder for my business. How I choose between Claude, EU cloud and a local model depending on the data class is in Private LLM: Vendor, EU Cloud or Own Server?.

That is not a verdict against European providers, just my weighting of the same six criteria. For your own application on open models with a German contracting party, look at IONOS, STACKIT or T-Systems. For a team chat, look at Mistral and check the training setting. How to keep your team off private accounts meanwhile is in Shadow AI: When Your Team Uses AI in Secret.

Frequently asked questions

Is there a European alternative to ChatGPT?

As a ready-made chat, Mistral Vibe from France comes closest in this overview, with a free plan for individuals. For teams, Langdock from Berlin offers a workspace across several vendors' models, running on Microsoft Azure in the EU. IONOS, STACKIT and T-Systems aim their model APIs at companies building their own applications.

Is European AI automatically GDPR compliant?

No. An EU data center is one building block; it replaces neither the legal basis for entering personal data, nor the DPA, nor the check whether the provider trains on your data. This is not legal advice; for your specific case, take the question to your data protection officer.

What does sovereign AI mean?

The term is not protected. Providers usually mean a seat and data centers in Europe, sometimes their own models or operation on your infrastructure. Ask about the six criteria, not the label.

Does Mistral train on my data?

In Vibe, yes by default, until you opt out in the settings; Enterprise customers are opted out by default. Feedback given with thumbs up or down may be used by Mistral on every plan. For the API (Mistral Studio), the free mode lets Mistral use inputs and outputs for training, with an opt-out at any time; on pay-as-you-go, Mistral says you keep control and can also opt out at any time. Labs and preview models are used for training regardless of your setting, per the contract.

Where to go from here

Take your most sensitive task and run the six criteria for the two or three candidate providers. Usually one point decides. How I set what data an AI gets to see at all is in AI and Privacy: What the AI Gets to See. Which obligations apply regardless of provider is in EU AI Act: Overview and Deadlines 2026.

How my AI employees work within these limits I show in my community.