Cloudian HyperStore
On-prem, S3-compatible, exabyte-scale object storage whose 8.2.6 release is NVIDIA-Certified and supports S3 over RDMA for direct GPU-to-storage data paths.
Summary
On-prem, S3-compatible, exabyte-scale object storage whose 8.2.6 release is NVIDIA-Certified and supports S3 over RDMA for direct GPU-to-storage data paths.
It sits beneath GPU clusters as a self-hosted S3 data plane, competing with public-cloud object storage and with software-defined stacks like MinIO and Ceph. Its differentiator is RDMA/GPUDirect throughput rather than just capacity economics, positioning it for AI-factory builds that want object storage to keep pace with NVMe and GPUs.
- The 35 GB/s and 210 GB/s figures require an RDMA/RoCE-capable network fabric; over plain TCP you get standard S3 throughput, not the headline numbers.
- "NVIDIA-Certified" here is the Foundation level (validated up to 128 GPUs), not an unlimited-scale guarantee.
acceleratesNVIDIA GPUDirect RDMA for S3 — moves objects into GPU memory bypassing CPU/HTTP.alternative_toMinIO — both are self-hosted S3, but HyperStore is appliance/enterprise-scale with RDMA.solvesEgress Cost — on-prem capacity model removes per-GB egress charges.
Definition
Cloudian HyperStore is an on-premises, exabyte-scalable S3-compatible object storage platform with native S3 API support. It runs on commodity hardware as a fully self-hosted alternative to public-cloud object storage, targeting AI training, fine-tuning, inference, and data-pipeline workloads. Version 8.2.6 is the current NVIDIA-certified release.
It gives teams a private, S3-API-native data plane that can sit directly under GPU clusters, so AI data never has to leave the building or traverse public-cloud egress. Its S3-over-RDMA path lets GPUs pull objects without the HTTP/TCP overhead that normally caps object-storage throughput, making it a credible local-first substrate for LLM and vector workloads.
AI training data lakes, vector-database storage backends, model checkpointing, GPU-fed inference pipelines, on-prem S3 for regulated/sovereign data, large-scale backup and archive.
Recent developments
- HyperStore 8.2.6 achieved the Foundation level of NVIDIA-Certified Storage, announced at NVIDIA GTC 2026. The designation validates the platform across the I/O patterns AI workloads demand for deployments involving up to 128 GPUs. Per Cloudian HyperStore Achieves NVIDIA-Certified Storage.
- S3-over-RDMA delivers 35 GB/s reads per node with linear scaling — a 6-node cluster sustains 210 GB/s. Cloudian implements S3 over RoCE (RDMA over Converged Ethernet), reporting a 7-8x throughput improvement over standard TCP-based S3 and over 40% lower CPU utilization on both compute and storage nodes. Per Supercharging Vector Database Indexing 8x Faster with Cloudian S3 RDMA and NVIDIA.
- GPUDirect integration moves data straight from object storage into GPU memory, bypassing CPU and system memory. Combined with GPU-accelerated indexing (cuVS), Cloudian cut Milvus index-build time on the MIRACL dataset 8x — from 2 hours to 16 minutes. Per Supercharging Vector Database Indexing 8x Faster with Cloudian S3 RDMA and NVIDIA.
- A single rack can scale to terabytes per second of sustained bandwidth. Per NVIDIA GTC 2026: Cloudian HyperStore Achieves NVIDIA-Certified Storage Designation.
Connections 8
Outbound 8
scoped_to1implements1accelerates1depends_on1alternative_to1competes_with1solves1Resources 4
Vendor announcement confirming 8.2.6, Foundation-level NVIDIA certification, and up-to-128-GPU validation.
Details the S3-over-RoCE numbers (35 GB/s/node, 7-8x over TCP, 40% CPU reduction) and the 8x Milvus indexing result.
Independent coverage tying the certification to NVIDIA GTC 2026.
Milvus's own writeup of the joint RDMA/GPUDirect vector-indexing benchmark.