Wedoany.com ReportWedoany.com Report on Mar 17th, , on 16 local time MinIO announced that its AIStor product will support object data storage for the NVIDIA STX reference architecture. AIStor, designed based on the NVIDIA STX rack-scale reference architecture, aims to provide unified, high-performance data storage covering the complete AI lifecycle, from large-scale model training to enterprise RAG and real-time agent inference. As part of the collaboration, MinIO will join the NVIDIA STX ecosystem as an AI data platform partner.
NVIDIA STX defines how an enterprise AI factory stores, moves, and accesses data at modern AI speeds. Relying on NVIDIA Vera Rubin, NVIDIA BlueField-4 processors, and NVIDIA Spectrum-X Ethernet networking, STX provides a modular, rack-scale storage architecture for production AI infrastructure. AIStor operates on the STX reference architecture as an object-native data foundation, positioned at the core of the AI factory.
As AI systems evolve from isolated inference to distributed multi-agent workflows, infrastructure demands expand in terms of training throughput, RAG indexing performance, and low-latency context access. AIStor meets these demands on a single, consistent platform that works in concert with the NVIDIA STX architecture.
AIStor provides a single data store suitable for the complete AI lifecycle. For high-performance data storage oriented towards training and analysis, it leverages the NVIDIA STX reference architecture to deliver the optimal throughput and power density required to keep GPU clusters fully utilized and energy-efficient. Its fully disaggregated design aligns with STX's rack-scale model, separating storage from compute for independent scaling and fault isolation without compromising low-latency performance.
For enterprise AI data, AIStor provides the data foundation for RAG and multimodal indexing pipelines. As part of the STX design, running natively on NVIDIA BlueField-4 processors, AIStor accelerates embedding, indexing, and retrieval at line speed while maintaining S3 API compatibility, reducing latency and operational complexity compared to traditional storage tiers.
For contextual memory in agent AI, AIStor extends the memory hierarchy beyond GPU constraints, collectively forming the rack-scale infrastructure required for these workloads. Within STX, AIStor provides scalable, high-performance object storage to preserve and retrieve inference state across distributed agents with near-DRAM predictability.
"AI is being reshaped by data as much as by models," said Garima Kapoor, Co-founder and Co-CEO of MinIO. "Infrastructure that succeeds in this era isn't retrofitted for AI; it's designed for AI. AIStor's integration with the NVIDIA STX reference architecture reflects years of careful engineering around performance, scale, and architectural simplicity. We are delivering the unified, object-native data foundation that production AI factories require."
The fit between AIStor and NVIDIA STX reflects deliberate architectural decisions made before rack-scale AI factories became mainstream. AIStor is delivered as a single static binary, under 200 MB in size, requiring no external metadata databases or background services. Its lightweight design runs natively within BlueField-4, which underpins the STX storage layer, eliminating dedicated x86 storage nodes and reinforcing STX's disaggregated architecture.
Within the NVIDIA STX architecture, AIStor unlocks advanced capabilities accelerated by NVIDIA DOCA and NVIDIA BlueField-4, including NVIDIA GPUDirect RDMA for S3-compatible storage, hardware-accelerated erasure coding, zero-copy data transfer, and hardware-accelerated encryption, checksums, and compression. These capabilities collectively ensure storage operates as an integrated component of STX, not as an external bottleneck.
MinIO AIStor based on the NVIDIA STX reference architecture on NVIDIA BlueField-4 is expected to be generally available in the second half of 2026, with eligible enterprise customers able to access early experience. AIStor with NVIDIA GPUDirect RDMA support for S3-compatible storage is now available as a technology preview.
