Where your AI runs: three paths and their terms
Companies want AI that keeps their data in Europe and under their own control. We show three ways to run a language model and the conditions each one brings.
Companies in Germany want AI from their own country. 96% would prefer AI from German providers.1 Behind that is a clear wish: their data should stay in Europe, under European law and under their own control.
How well that can be met depends on where a model runs. There are three ways. Each comes with its own conditions for your data, for model quality and for operations.
Level 1: international providers
Many start with ChatGPT, Claude or Gemini, on a team plan or through the provider's API. You get the strongest models at once and have nothing to operate. Processing happens in the provider's data centres, not necessarily in Europe.
For many companies this rules the route out for customer data. Many clients require their documents to be processed in European data centres, under the GDPR and similar rules. For public information and non-critical internal texts it still works well.
Level 2: leading models in European data centres
The large cloud providers offer many leading models from European data centres. Examples are Amazon Bedrock in Frankfurt, Azure under Microsoft's EU Data Boundary and Google Vertex AI with EU endpoints.2,3 Your data is then processed on European soil or in your region and is not routed to international data centres.
The terms state that your data is not used for training.4 The provider runs and updates the models.
Level 3: open models in a sealed-off environment
Open models such as Llama, Mistral, Qwen or gpt-oss can run on your own servers, fully offline if need be. Alternatively you work with confidential computing in the cloud, which also relies on open models. There, data stays in protected hardware even while it is processed.5 Both are real options today.
This level comes with the highest effort and the highest cost. Results at the level of modern AI tools need plenty of computing capacity. They also need people who run the models and keep them current. Since January 2026, the strongest open models have trailed the best closed ones by about four months on average.6 This level is meant for companies in security-critical industries whose data, as a rule, may not leave their own computers.
International providers
ChatGPT, Claude, Gemini as plan or API
- strongest models, available now
- nothing to operate
- processing not necessarily in Europe
- often ruled out for customer data
For public and non-critical internal content
Leading models in Europe
e.g. Bedrock Frankfurt, Azure EU, Vertex AI EU
- many leading models
- processing in Europe, no training on your data
- nothing to operate
- still in the cloud, not for extreme security requirements
For most project and customer data
Open models, sealed off
own servers or confidential computing in the cloud
- full control over the data
- a route for security-critical industries
- highest effort, highest cost
- models a few months behind the frontier
For security-critical and sealed-off areas
Our own summary based on provider documentation, as of 10 Nov 2026.
Which data belongs on which way
If you are unsure, a thought experiment helps. Ask yourselves whether you would send these documents by email today, through Microsoft 365 or similar providers. Everything you would send that way fits level 2. In the normal case that covers most of the work in an industrial setting.
What lives only in separated, offline or locally sealed environments may call for level 3. Some projects contain both. Then it is about separating which data may be used and which may not, and introducing clear rules for it. That way you find a solution instead of rejecting the whole project over a few parts.
The choice can differ per project. A new office building might fit level 2. The documentation of a defence project belongs in your own data centre or a confidential cloud environment. What matters is that you can switch without starting over.
Engera runs on level 2 by default. That ensures your data is not used for training and stays in Europe, as requirements and security demand. Level 3 works too: either you run the open models yourselves, or we use our partner for confidential computing in the cloud. We set that up by agreement.
As of 10 November 2026. Provider terms change often; the contract and current documentation are what count. This text is orientation, not legal advice.
What to do now
- 01
Sort your documents
Run your documents through the email question and decide what belongs on which level.
- 02
Set the rules
Write down which data may be used and which may not.
- 03
Keep the switch open
Set up data and permissions so you can change the way per project.
- 04
Start small
Begin with one concrete, bounded use case on the level its data actually needs.
Sources
- Bitkom: Digital-Gipfel: Unternehmen wollen KI aus Deutschland (companies want AI from Germany), 02.11.2026
- Microsoft: What is the EU Data Boundary?, updated 21.07.2026
- Google Cloud: Vertex AI: Data residency, accessed 16.11.2026
- Google Cloud: Generative AI and data governance (Vertex AI), accessed 16.11.2026
- Microsoft Learn: Azure confidential computing products, updated 09.11.2026
- Epoch AI: Open-weight models lag the closed frontier (ECI gap), 29.05.2026
Read on
Which challenge is keeping
your team busy right now?
Tell us, and in the demo we show you how Engera makes it easier.

