Pedro Olivares
EU Hosting: Choose Where Your AI Runs

When an agent sends a prompt, a document or a tool result to a model, that request is processed somewhere. For many European teams, where is a real question. It comes up in security reviews, in contracts and in conversations with a data protection officer, and it is a different question from where your data is stored.
EU hosting is how Timbal answers it. An admin of an Enterprise organization turns it on in Settings, and from then on every model call the organization makes through Timbal is served by a deployment in the EU, or it is refused. There is no quiet fallback to a global endpoint. This post explains what that means, what changes when you turn it on, and the places it does not reach yet.
How a request is handled
With EU hosting on, Timbal sends your model calls only to EU deployments of the providers. OpenAI models are served from OpenAI’s EU endpoint. Claude is served through Amazon Bedrock, using EU inference profiles that route only to EU regions. Gemini is served through Google Vertex AI in the EU.
If a model has no EU deployment, calls to it are refused. Timbal does not retry on a global endpoint, with another key, or on another provider outside the boundary. A refused call is the intended behaviour: the alternative is data leaving the place you asked it to stay.

What happens when you turn it on
Turning it on changes which models people in the organization can use, so Timbal shows the effect first. A preview replays the last 30 days of your model usage against EU hosting and lists the models and projects that would be blocked, with call counts, and the ones that would keep working. You decide with the numbers in front of you.
Models without an EU deployment are blocked for everyone in the organization, admins included. The rule is locked: a rule you write cannot turn a blocked model back on.
The list of allowed models is explicit. A new model or provider stays blocked until it has been checked for the EU, so a launch elsewhere never opens a gap.
Provider-side state does not move. On OpenAI, responses, files and vector stores created before the switch belong to a different project from the EU one, and nothing is migrated. A conversation that relies on one of them has to start again.
Knowledge bases keep working. Embeddings from the EU endpoint are identical to the ones from the global endpoint, so nothing needs to be reindexed.
Turning EU hosting off again needs an explicit confirmation that data may leave the EU, and both changes are written to the audit log.
Timbal’s own model calls follow the rule
Your agents are not the only thing that calls models. Timbal uses them to title sessions, split documents into chunks for knowledge bases, embed text and route requests in ACE. With EU hosting on, those calls use EU deployments too. Where no EU model fits, the feature gives way instead of going global: a session keeps a plain title, a document is split by rules instead of by a model.
What it costs
Some providers charge 10% more for processing in the EU, on some of their models. Timbal bills each call at the rate of the deployment that served it, so those models cost 10% more on an organization with EU hosting on and the rest stay at their usual price. The confirmation shows the difference before you commit.
What it does not cover
EU hosting covers model calls. Other parts of the platform run on other services, and those are not blocked on an EU organization and not covered by EU hosting yet. The confirmation lists them before you turn it on, and so do we:
Knowledge base parsing. PDF and Office files go through a document-parsing service, and audio and video through a speech-to-text service, both run on Timbal’s accounts.
Speech-to-text. Transcription, including realtime transcription, runs on a speech provider.
Platform tools. Web scraping, image and video generation and speech run on provider accounts that Timbal owns.
Voice agents. Speech-to-text and text-to-speech for voice agents use speech providers.
Composer. The app builder runs on an external coding agent and the model providers behind it.
App code with its own provider key. A call your app makes straight to a provider skips Timbal, so that provider’s own settings apply.
Some provider features also differ on EU deployments. Claude’s built-in web search is not available on Amazon Bedrock, so on an EU organization Claude calls Timbal’s search tool instead, and web fetch and code execution are refused.

Where models run is not where data lives
EU hosting is about where models process your requests. It is not a commitment about where records are stored, and by itself it does not describe where every other part of the platform runs. A requirement about where prompts are processed needs an inference answer, and a requirement about where data is stored needs a storage answer. Many reviews need both, so we keep them apart and say which one each feature gives you.
Turning it on
EU hosting is available on the Enterprise plan. An admin can find it in Settings, then Models, under Hosted in the EU, and read the preview before turning it on. If EU processing is a requirement for you and you are on another plan, talk to us.


