As NASA and other Earth science agencies migrate petabyte-scale archives to cloud platforms like Amazon Web Services (AWS), they confront new challenges and opportunities in long-term data stewardship. With NASA’s Earthdata Cloud hosting upwards of one billion files ("granules"), traditional data access and management techniques designed for on-premises systems often fall short. Many granules are stored in formats that assume fast, local, and random file access and require specialized API libraries to extract values. To enable performant data access in cloud-native object stores like AWS S3, we (OPeNDAP, NASA, and partners) have developed access methods that bypass traditional libraries. These methods rely on auxiliary "sidecar" metadata files—one per data file—which describe how to locate and reconstruct data chunks using HTTP Range-GET operations. This doubles the number of managed files and introduces new preservation concerns. Sidecar files support services like NASA’s OPeNDAP server and client applications like VirtualiZarr, both perform equivalent low-level operations to reconstruct structured data. However, this approach introduces new stewardship obligations: if a data file location or content changes, the associated sidecar must be updated. Likewise, inconsistencies in metadata stored in NASA’s Common Metadata Repository (CMR) can impair data discovery and access.While these problems are familiar in computer science, they are novel in scientific data stewardship. Cloud platforms enable decentralized collaboration across organizations, yet our data systems require centralized consistency for archival integrity and access. This tension highlights emerging risks and the need for updated strategies in cloud-based data stewardship.