AIStor is the Data Foundation for the NVIDIA AI Factory

The NVIDIA® AI Factory brings together the compute, networking, and software building blocks enterprises use to industrialize AI. AIStor is the object-native data store at the center of the AI factory.

Get a Demo

Get The Whitepaper

AIStor Fuels the AI Factory

NVIDIA STX defines how enterprise AI factories store, move, and access data at modern AI speeds. AIStor is the unified data foundation accelerating training, enterprise RAG, and real-time agentic inference throughout the AI lifecycle.

High-Throughput AI Training

AIStor saturates GPU interconnects during training by delivering data at the speed GPUs consume it, eliminating the storage bottleneck that limits cluster utilization.

GPUDirect RDMA for direct GPU-to-storage transfers

Parallel data loading across distributed training jobs

Eliminates CPU and network bottlenecks in the data path

Enterprise RAG at Scale

AIStor stores and serves the knowledge corpora that enterprise RAG pipelines depend on — with the performance and reliability production AI demands.

Millions of vectors and documents at sub-millisecond access

Native S3 compatibility for seamless pipeline integration

Scales with retrieval demand without reconfiguration

White share icon with three connected dots on a transparent background.

Real-Time Agentic Inference

AIStor offloads KV cache during inference, extending GPU memory capacity and enabling long-context, multi-agent workloads at production scale.

KV cache tiering to fast object storage

Sub-200μs latency for time-critical inference calls

Supports long-context and multi-model pipelines

Sovereign AI Infrastructure

Deploy AIStor in your data center, with no external dependencies. Your data stays where compliance demands it, while GPUs still run at full speed.

No egress to public cloud—full data residency and sovereignty on-premises

Deploys on commodity hardware within the NVIDIA STX architecture

12× GPU Throughput vs. Standalone NVMe

Sub-Millisecond Latency

5× More Training Data Processed per GPU Hour

Blog

NVIDIA® BlueField®

AIStor delivers Object Data at wire speed.

Whitepaper

NVIDIA AI Factory

Practical deployment patterns and customer outcomes for enterprises to scale AI without scaling complexity.

Blog

NVIDIA cuVS & Milvus

12X Faster Vector Index Creation with NVIDIA GPUDirect RDMA for S3-compatible storage.

Press Release

MinIO AIStor Brings Object Data Stores for the NVIDIA STX Reference Architecture

Blog

MinIO AIStor with NVIDIA GPUDirect® RDMA for S3-Compatible Storage: Unlocking Performance for AI Factory Workloads

Blog

Supercharging Inference for AI Factories: KV Cache Offload as a Memory-Hierarchy Problem

How AIStor + NVIDIA Works

AIStor deploys within the NVIDIA STX reference architecture as the unified storage layer, replacing fragmented multi-tier architectures with a single high-performance data store that feeds every stage of the AI pipeline.

AIStor Inside NVIDIA BlueField-4

AIStor runs natively on the NVIDIA BlueField®-4 DPU, putting storage compute directly on the SmartNIC and eliminating host CPU overhead entirely.

Storage processing offloaded to BlueField-4 ARM cores

Host CPU freed entirely from storage I/O

AIStor operates at wire speed on 800GbE fabric

Purpose-built for AI factory scale deployments

White lightning bolt icon on a transparent background.

Storage at Wire Speed

AIStor integrates with NVIDIA GPUDirect® RDMA for S3-Compatible Storage to enable direct memory transfers between storage and GPU, bypassing the CPU and system memory entirely.

GPUDirect RDMA for S3-Compatible Storage is available now as a tech preview.

Data moves from AIStor directly into GPU memory

Eliminates CPU bottlenecks in the training data path

Enables storage to keep pace with 800GbE GPU interconnects

Validated on NVIDIA BlueField-4 SmartNIC infrastructure

KV Cache Offload: Inference Memory Hierarchy

AIStor acts as a high-speed memory tier for KV cache during inference — extending effective GPU memory capacity without sacrificing latency.

Offloads KV cache to fast object storage under memory pressure

Sub-200μs retrieval latency preserves inference SLAs

Enables long-context and multi-agent inference at scale

Transparent to the inference stack via standard S3 APIs

Architected for NVIDIA STX

MinIO’s minimal footprint integrates natively with the NVIDIA BlueField-4 DPU that underpins the STX architecture.

Static binary under 200 MB

No external metadata databases or background services

Runs natively within BlueField-4 DPUs, eliminating dedicated x86 storage nodes

Proven Results

Performance validated across real AI factory workloads—training, inference, and retrieval at scale.

12× GPU Throughput

AIStor with GPUDirect RDMA delivers 12× the GPU throughput of a standalone NVMe configuration — saturating the interconnect so GPUs are never waiting on data.

Bar chart with four vertical bars of increasing height from left to right.

Sub-Millisecond Latency

KV cache offload to AIStor preserves inference SLAs - delivering sub-millisecond latency even under sustained production load.

White speedometer icon with needle pointing to the right on black background.

5× Training Data Per GPU Hour

Enterprises using AIStor in AI factory configurations process 5× more training data per GPU hour compared to traditional storage approaches.

106M Vectors in 120 Seconds

AIStor ingests and indexes 106 million embedding vectors in 120 seconds — enabling enterprise RAG pipelines to operate at a scale that matches production data volumes.

Built for Real-World AI Factory Workloads

AIStor and NVIDIA address the full spectrum of enterprise AI infrastructure, from data preparation to production inference.

AI Training

Large language model pre-training

Foundation model fine-tuning

Multi-modal training pipelines

Distributed training data loading

Inference & Serving

Real-time agentic inference

Long-context model serving

Multi-model pipeline orchestration

KV cache offload for throughput

Enterprise RAG

Vector embedding storage at scale

Knowledge corpus management

Research data governance across hybrid environments

Hybrid search pipelines

Data Preparation

Dataset curation and preprocessing

Feature engineering pipelines

Model checkpoint storage

Experiment tracking and artifacts

AIStor is the Data Foundation for the NVIDIA AI Factory

AIStor Fuels the AI Factory

High-Throughput AI Training

Enterprise RAG at Scale

Real-Time Agentic Inference

Sovereign AI Infrastructure

Read More

How AIStor + NVIDIA Works

AIStor Inside NVIDIA BlueField-4

Storage at Wire Speed

KV Cache Offload: Inference Memory Hierarchy

Architected for NVIDIA STX

Proven Results

12× GPU Throughput

Sub-Millisecond Latency

5× Training Data Per GPU Hour

106M Vectors in 120 Seconds

Built for Real-World AI Factory Workloads

Get AIStor running in your NVIDIA AI Factory environment in minutes.