To unlock accelerated AI storage performance, organizations are turning to RDMA for S3-compatible storage, a technology that leverages remote direct memory access to dramatically improve data transfer speeds for AI workloads. As AI demands grow—projected to generate nearly 400 zettabytes of data annually by 2028, with 90% being unstructured—traditional storage solutions struggle to keep pace with the need for high throughput, low latency, and scalability across hybrid environments. Object storage has traditionally served as a cost-effective solution for archives, backups, data lakes, and logs. While increasingly used for AI training, it has historically lacked the performance needed for large-scale, concurrent GPU workloads. RDMA for S3-compatible storage addresses this gap by optimizing the S3 API-based protocol using RDMA, enabling faster, more efficient data transfers directly between memory spaces without involving the CPU. Powered by NVIDIA networking, this solution delivers higher throughput per terabyte, improved throughput per watt, lower cost per terabyte, and significantly reduced latencies compared to TCP, the standard transport protocol. By integrating RDMA client libraries into AI GPU compute nodes, data can be accessed from object storage much faster, boosting GPU utilization and accelerating AI training cycles. The architecture is open and extensible. While initially optimized for NVIDIA GPUs and networking, the libraries are designed to allow other vendors and developers to contribute, customize, and integrate RDMA support into their own software ecosystems. NVIDIA is collaborating with leading storage providers to standardize and accelerate adoption. Cloudian, Dell Technologies, and HPE have already incorporated RDMA support into their high-performance object storage platforms—Cloudian HyperStore, Dell ObjectScale, and HPE Alletra Storage MP X10000—delivering unmatched scalability and performance for AI environments. Jon Toor, Cloudian’s chief marketing officer, emphasized that RDMA for S3-compatible storage enables faster, more efficient data handling while maintaining compatibility with existing S3-based tools and applications across on-premises and cloud deployments. Rajesh Rajaraman of Dell Technologies highlighted that AI factories—whether on-premises or in the cloud—require seamless data portability and high-performance storage under massive concurrent access. Dell ObjectScale’s integration with RDMA delivers dramatically lower latency and enhanced scalability. Jim O’Dorisio from HPE noted that RDMA for S3-compatible storage is transforming how data moves at scale. With RDMA capabilities built into HPE Alletra Storage MP X10000, organizations gain higher throughput, reduced latency, and lower total cost of ownership—key advantages for AI-driven workloads. NVIDIA’s RDMA for S3-compatible storage libraries are currently available to select partners and will be generally released via the NVIDIA CUDA Toolkit in January. Additionally, NVIDIA has launched a new Object Storage Certification as part of the NVIDIA-Certified Storage program, helping ensure interoperability and performance for AI workloads.