Wally on Google Vertex AI EU — data residency for GDPR

Vertex AI offers Anthropic Claude in several EU regions (europe-west1, europe-west4, europe-west9, etc.). Pick this path when your contracts, customers, or applicable regulators require AI inference on data of EU subjects to stay within the EU.

What you need:
• A GCP project with Vertex AI enabled and Anthropic Claude models enabled on the publisher catalog
• A service account or workload identity with aiplatform.endpoints.predict permission on the Claude publisher endpoint
• The corresponding API key, OAuth token, or ID token that your gateway can present as an Authorization or x-api-key header

Endpoint format:
A Vertex Anthropic endpoint looks like:
https://europe-west4-aiplatform.googleapis.com/v1/projects/YOUR_PROJECT/locations/europe-west4/publishers/anthropic/models/claude-haiku-4-5:streamRawPredict

Most teams put a small proxy in front (Cloud Run, Cloud Functions, or API Gateway) that accepts Anthropic-format messages and forwards to Vertex with the right auth header. The Settings page accepts either the raw Vertex URL or your proxy URL — anything that speaks the Anthropic Messages API.

What this gets you:
• Inference runs in the EU region you selected. Prompts containing EU subject data never leave that region for the AI hop.
• Your Vertex bill comes from Google directly; the Turtini +15% chargeback does not apply to the inference path
• Your DPO can name a specific GCP region in your records of processing activities

Region selection:
The region field on the Settings page is informational — what actually pins the region is the endpoint URL you paste. Choose europe-west4 (Netherlands), europe-west1 (Belgium), or europe-west9 (Paris) depending on your residency needs.

What still runs outside the EU:
The Turtini application itself (Firestore, Cloud Functions, Cloud Run) currently runs in Google Cloud's us-central1 region. Wally orchestration metadata — which user asked, what tool was called, the audit log entry — lives there. If your contractual posture requires the entire application stack to be in the EU, that's a separate Turtini deployment conversation.