Bedrock Cross-Region Inference and Vertex AI Global Endpoints: How to Keep Prompts in an Approved Region
To keep prompts in an approved region, use the narrowest routing option that meets your residency commitments: an in-region Bedrock profile or a regional Vertex AI endpoint. Allow a geographic or multi-region option only where the whole geography is approved. Treat global profiles and endpoints as non-approved by default. Restrict the options with IAM and organization-level policy, then enforce the same rule at runtime in the request path and keep per-request evidence of where each call was directed.
What to be able to show for each request
- Which workload and identity made the request.
- Which profile, endpoint type and region the request was directed to.
- Which policy was evaluated and whether the result was allow, block or escalate.
- Who approved any exception, and when.
- Which provider-side metadata corroborates the destination, and what it does not cover.
- How long the record is retained and where it is stored.
Layer one: provider-native controls
Why a routing choice is a residency decision
In both Amazon Bedrock and Google Vertex AI, the location where a prompt is processed depends on a provider-managed routing construct, not on where your application runs. In Bedrock that construct is the inference profile. In Vertex AI it is the endpoint type: regional, multi-region or global. Selecting a different profile or endpoint is often a one-line change to an identifier, and nothing else in the application looks different.
That is the core risk for regulated enterprises. A team can move prompts and completions across a border by swapping a model identifier or an endpoint hostname, with no change to the network path, the IAM role or the code review checklist. Data residency, privacy and contractual commitments are usually written in terms of geography, so the architecture has to express the same terms.
The practical approach is to classify every routing option by the geography in which it may process a request, decide which classes are approved for which data, and then make that decision enforceable rather than advisory.
Routing models side by side
The two providers use different names for similar ideas. The table maps them by the question that matters for residency: how wide is the set of locations that may process a request? Confirm the exact behavior, supported models and data-at-rest handling for each option in your own environment before approving it.
- Single region: Bedrock in-region profile, Vertex AI regional endpoint: Requests are directed to the one region you call. This is the narrowest option and the usual baseline for strict residency. Tradeoffs to check are model availability, quota and capacity in that region.
- Geography: Bedrock geographic cross-region profile, Vertex AI multi-region endpoint: Processing may occur in any region within a defined geography, and the provider chooses which one. This is acceptable only if the entire geography is within your approved boundary.
- Global: Bedrock global profile, Vertex AI global endpoint: Processing may occur outside any single geography. Treat as non-approved by default for residency-bound workloads, and approve only by explicit, recorded exception.
- Defaults and new models: A newly launched model may be reachable first through a cross-region or global option. Treat each new model identifier as a change that needs a routing review before use.
Where provider-native controls leave gaps
Cloud-native restrictions are static. They reflect the configuration at the time someone wrote them, and several gaps follow from that.
Drift is the first. Different teams, accounts and projects accumulate their own policies and exceptions, and the effective rule differs from the written one. New model launches are the second. If a model is offered through a cross-region or global option, teams are motivated to adopt it, and a permissive allowlist or a wildcard in an IAM policy lets it through without review.
The third gap is evidence. A one-time architecture review shows what was intended. An auditor asking about a specific prompt wants to know where that request was directed and under what rule. Provider logs and response metadata may show some of this and omit other parts, such as the policy decision that permitted the call. Do not assume they answer the per-request question until you have checked what they actually contain for each routing type.
These gaps argue for a second layer that evaluates each request at the moment it is made, instead of relying only on configuration that was correct when it was written.
Layer two: runtime enforcement in the execution path
A runtime enforcement layer sits in the path between the application and the model endpoint. For each request it compares the target profile, endpoint or region with the approved list for that workload and data class, and then acts. Permitted requests are allowed. Requests to non-approved endpoints are blocked. Ambiguous or exception cases, such as a request to a global option for a workload that has no approval, are escalated for human review.
This follows a general pattern in current public policy material. A draft GSA clause on safeguarding data in LLM systems, published for comment in June 2026 and applying only where Government Data is processed by an LLM, describes defense in depth through access controls, policy enforcement points and continuous monitoring. As of November 2026 it remains a draft, and it does not by itself set routing rules. It is useful as a reference for the layered structure, not as a residency requirement.
Trussed provides this layer. It enforces policy in the execution path, so decisions are enforced at runtime, not reviewed after the fact, and the scope is the individual interaction and tool call. It can be deployed self-managed in your own cloud, with configurable retention and data residency, so the enforcement and its evidence stay within your boundary. Each decision produces audit-ready evidence that records the destination a request was directed to and the policy outcome. Trussed enforces policy across 100+ models under one policy plane, which matters when Bedrock and Vertex AI sit side by side.
Three layers for region adherence
Choose
Select the narrowest routing option that fits your residency commitments.
Restrict
Limit callable profiles, endpoints and regions with IAM and organization policy.
Enforce and evidence
Allow, block or escalate each request at runtime and keep a record of the decision.
Make region adherence enforceable and provable
Talk through your Bedrock and Vertex AI routing options and where runtime enforcement and per-request evidence fit alongside your cloud-native controls.
Talk to an ExpertHow does your AI governance program compare?
See where your program has gaps in less than 2 minutes.