This article presents a reference architecture for implementing a hybrid AI/ML data access strategy using Amazon SageMaker, addressing enterprises with both on-premises and cloud infrastructure.

• Define hybrid data strategy: map ML workloads, connectivity, single source of truth, and storage performance requirements
• Use AWS Direct Connect for high-speed on-premises to cloud connectivity
• Employ AWS DataSync and Storage Gateway to migrate on-premises data to Amazon S3
• Designate Amazon S3 as the primary source of truth for ML datasets
• SageMaker Studio users download S3 data to EFS-backed home directories for experimentation
• SageMaker training jobs access data via S3 (File, Fast File, Pipe modes), FSx for Lustre, or EFS
• Use Amazon File Cache for high-speed caching and FSx for NetApp ONTAP for cloud bursting
• Archive data with S3 Glacier to reduce storage costs

This architecture enables enterprises to leverage cloud-native ML capabilities while maintaining on-premises infrastructure and data management practices.