A media company needs low-latency article recommendations for readers in one city. Traffic peaks during predictable times (morning, lunch, after work) and is light otherwise. Inference responses are under 4 MB. Which deployment option minimizes latency and is most cost-effective?

Serverless inference with provisioned concurrency is the most cost-effective and low-latency option. Serverless inference (e.g., AWS Lambda or SageMaker Serverless Inference) scales automatically to handle traffic peaks and scales down to zero during low traffic, minimizing costs. Provisioned concurrency eliminates cold starts, ensuring low latency even for infrequent requests. Real-time inference endpoints with auto scaling would incur costs even when idle, as instances remain provisioned. Asynchronous inference and batch transform jobs are unsuitable for low-latency, real-time recommendations, as they introduce significant delays by processing requests in batches rather than individually and immediately.