AWS AI Updates: April 24, 2026

1. SageMaker HyperPod Gets Automatic Slurm Topology Management

AWS. HyperPod now automatically selects and continuously maintains the optimal network topology for Slurm-based distributed training clusters. The feature removes one of the more error-prone manual steps in standing up multi-node training — picking the right topology for the hardware and then keeping it tuned as nodes fail in and out. For teams running long pretraining or fine-tuning jobs on HyperPod, it means less babysitting of the scheduler and fewer wasted GPU hours on suboptimal layouts. Source

2. SageMaker Notebooks Get a Natural-Language Data Agent for IAM Identity Center Domains

AWS. SageMaker Unified Studio notebooks now ship with a built-in AI data agent for IAM Identity Center domains. The agent generates Python code and SQL from natural-language prompts inside the notebook, targeting the repetitive “glue” work of exploring new tables, writing joins, and producing boilerplate pandas snippets. It complements rather than replaces manual editing — generated cells land in the notebook and can be edited before running. Source

3. Elastic Beanstalk’s AI Environment Analysis Adds Windows Support

AWS. Elastic Beanstalk’s AI-powered environment analysis, which sends logs and configuration to Amazon Bedrock to surface root causes for failing deployments, now works on Windows environments in addition to Linux. It’s a narrow but practical win for the large fleet of .NET and IIS workloads still on Beanstalk — those teams can now get the same Bedrock-assisted diagnostics that Linux customers have had, without shipping logs to a separate analysis tool. Source

4. SageMaker Unified Studio Notebook Kernels Can Now Run Inside a VPC

AWS. Unified Studio notebook kernels can now be attached to a customer VPC, giving data scientists network isolation for interactive ML workloads and direct, private access to VPC-only data stores. The feature closes a common gap for regulated customers who had standardized on Unified Studio but needed to run exploratory notebooks against databases and feature stores that do not expose public endpoints. Source