An online delivery company has one XGBoost model per city stored in S3 and currently hosts each model on its own EC2 instance (5% utilization). They want the web app to select the fastest courier in real time for both existing and new users while minimizing operational overhead. Which solution achieves this with the least management effort?

The correct answer leverages Amazon SageMaker's multi-model endpoint capability, which allows hosting multiple models on a single endpoint. This significantly reduces operational overhead compared to managing individual EC2 instances or separate SageMaker endpoints per city. The multi-model server dynamically loads models from S3 based on the TargetModel parameter in the inference request, making it efficient for a large number of models (cities) and minimizing resource utilization. Incorrect options: Batch transform is suitable for offline processing, not real-time inference, and reading from S3 files would introduce latency and complexity for real-time lookups. Consolidating to a single EC2 instance still requires managing the instance and scaling it manually, which doesn't fully minimize management effort or address the 5% utilization issue efficiently. Creating separate SageMaker endpoints per city would lead to high operational overhead and cost, as each endpoint would incur its own infrastructure costs, similar to the original problem of having many underutilized EC2 instances.