What changed
Amazon SageMaker AI inference routers are generally available for real-time endpoints in supported commercial Regions. A routing target can span Regions and accounts, prefer local capacity, shift traffic after endpoint errors and expose the destination and routing reason on each response.
Why it matters
Managed failover can replace application-specific health checks and routing, but it also creates a data-residency, retry and version-consistency boundary across every registered endpoint.
What to do
Register identical canary endpoints in two Regions, pin residency-sensitive traffic locally, saturate the primary, then verify one-retry behavior, destination telemetry, failback timing, model parity and remote-capacity limits.
Caveat
AWS and Salesforce report the implementation and one joint test; the router does not copy artifacts, create capacity or move traffic for latency alone. Results are vendor and customer evidence, not an independent reliability guarantee.