Amazon SageMaker AI in 2025, a year in review part 2: Improved observability and enhanced features for SageMaker AI model customization and hosting

This article reviews Amazon SageMaker AI's 2025 enhancements focused on observability, model customization, and hosting capabilities for generative AI workloads.

• Enhanced metrics provide instance and container-level tracking of CPU, memory, GPU utilization, and invocation performance
• Rolling updates enable zero-downtime deployments with configurable batches and automatic rollback on CloudWatch alarms
• Serverless model customization automatically provisions compute resources for fine-tuning with advanced techniques like RLVR and RLAIF
• Bidirectional streaming enables real-time multi-modal applications with persistent connections for voice agents and transcription
• PrivateLink support and IPv6 compatibility enhance enterprise security and compliance requirements
• Integrated MLflow experiment tracking for automatic metric logging during model customization
• Pay-per-token pricing model eliminates upfront infrastructure investments for model customization

SageMaker AI's 2025 improvements significantly enhance observability, deployment safety, and accessibility for enterprise AI workloads at scale.