AWS adds live inference data capture to SageMaker HyperPod

AWS says SageMaker HyperPod now provides Data Capture, which records live model inputs and outputs to Amazon S3 for analysis, auditing and training. Teams can enable capture at the endpoint, load balancer or model container, with separate controls for each. For capture at the model container, a proxy copies requests and responses into a collector that batches them for storage. AWS demonstrates using those records to identify a missing support category and fine-tune a model while another GPU continues serving requests. The walkthrough uses synthetic data; complete captured payloads can contain personal information and require protection before reuse for training.