Provider / Model Trust Matrix
A side-by-side of who sees what, where, under whose DPA.
A comparative table covering AWS Bedrock, Google Vertex, and Azure Foundry — data residency, training opt-out, inference isolation, audit log availability, and contractual chain.
Comparative columns
- AWS Bedrock, Google Vertex AI, Azure OpenAI / Foundry
- Data residency guarantees per region
- Training opt-out (default vs. configured)
- Inference isolation and tenant boundary
- Audit log availability and retention
- Contractual chain (DPA, sub-processors)
Citations
- Every cell traces to a vendor documentation URL or master agreement clause
- Last-verified date stamped per row
Read the first pages before you take it.
Row: Training on customer data (default behaviour). AWS Bedrock — never. Google Vertex AI — never for first-party models; configurable for tuned models. Azure OpenAI — never; explicit opt-out documented in the Service Specific Terms.
Row: EU data residency. AWS Bedrock — eu-west-1, eu-central-1, eu-west-3 with regional inference profiles. Google Vertex AI — europe-west and EU multi-region with data residency commitment. Azure OpenAI — Sweden Central, France Central with EU Data Boundary.
Row: Per-invocation audit log. AWS Bedrock — CloudTrail + CloudWatch, full invocation metadata. Google Vertex AI — Cloud Audit Logs, request/response logging configurable. Azure OpenAI — Diagnostic settings to Log Analytics, content logging opt-in.
Row: Sub-processor visibility. All three publish a sub-processor list with notification of additions; AWS and Azure provide 30-day prior notice, GCP publishes ongoing.
One email. No marketing list.