Data residency
Placement is specific to a resource or model route. An organization's location does not establish residency for every product.
Compute placement
Tarka chooses the current default pool for new Agent Hosts, sandboxes and jobs. Read ListLocations for each product's default and whether it accepts new resources; the creation quote records the exact location before you accept. Existing resources retain their location and volumes. Requesting a location that is closed to new resources returns an error instead of silently moving work.
| Location | Provider and country | Scope |
|---|---|---|
us-1 | Amazon Web Services, United States, Northern Virginia (us-east-1) | Elastic compute workers. The control plane is separate. |
eu-1 | Netcup, Europe. The exact country is not yet verified in the deployment inventory. | Existing compute and the European control and data services. |
US compute can pull images and send control-plane metadata to European services. It is not a promise that every associated byte stays in the United States. Storage, Hosted Git, API authentication and monitoring have their own placement.
Model processing and OCR
Model routing is independent of compute placement. The -eu aliases select their configured European inference endpoint; they do not change the location of other model calls, OCR, stored files or control-plane records.
The utility GPU worker serving glm-ocr is rented through Vast.ai. Its physical country is not yet verified in the deployment inventory. Do not rely on that OCR route for a country-specific residency requirement until Tarka confirms its location. Other OCR aliases require their own verified route inventory; they do not inherit a country from their model name.
Country-specific requirements
Check the accepted quote and resource location, and obtain confirmation for any required processing, storage, logs and backup countries before sending restricted data. An unverified country is an unresolved placement detail, not a residency guarantee.