Amazon SageMaker HyperPod introduced new capabilities for enterprise inference, including data capture that records inputs and outputs at multiple points along the inference path—from the endpoint, to the load balancer, to the model pod—providing deep observability and auditability through declarative custom resource definition configuration. Users can deploy models directly from popular community hubs without pre-staging weights, with built-in support for gated access, revision pinning, and token isolation across runtimes such as vLLM, TGI, and SGLang. Loading weights from node-local NVMe storage reduces cold-start latency, with automatic fallback to cloud storage. HyperPod automatically manages custom domain DNS records, and granular pod-level AWS Identity and Access Management permissions provide fine-grained security control. For data capture, prerequisites include an Amazon Simple Storage Service bucket and the right IAM permissions. Each tier has required settings: for the SageMaker Endpoint level, turn it on, set a destination S3 URI, pick at least one capture option, and choose a sampling percentage; for the Load Balancer level, enable it with an s3:// URI; for the Model Pod level, enable it, defaulting to capturing both input and output at 100 percent sampling.