Data and Memory Foundation Optimized for the NVIDIA AI Factory

Proven at memory speed and exabyte scale for both training and inference, AIStor is the object and S3 native data store optimized for NVIDIA AI Factory.

Get a Demo

Read the AI Inference White Paper

MinIO Doubles the Sessions Your GPUs Hold

$16M

In annual compute savings on 1024-GPU cluster at 128k context

90%+

GPU utilization on your NVIDIA GPU clusters

40-60%

Lower cost per token, same cluster

77%

Of the Fortune 500 run on MinIO

4.6x

Faster GETs with NVIDIA GPUDirect for S3-Compatible Storage

2.4x

Tokens per GPU once sessions outgrow HBM

800 GbE

Full line rate delivered on NVIDIA BlueField-4

The concurrent sessions from the same GPUs

Learn More

Purpose Built for AI, Not Bolted On

AIStor is the unified data foundation accelerating training, enterprise RAG, and real-time agentic inference designed to run natively on the NVIDIA BlueField-4 STX Storage Processor.

Feed Every GPU at Full Speed

Help keep GPUs supplied with data and fully utilized. AIStor streams training datasets straight into GPU memory over NVIDIA GPUDirect® for S3-compatible storage, sustaining training throughput across the storage fabric. Checkpoints save and restore on the same path, so the job keeps moving.

Direct-to-GPU data delivery with device pointer support for PyTorch and CuPy

Parallel dataset loading across distributed training jobs

Checkpoint saves and restores that don't stall the run

Enterprise RAG at Scale

Ground every answer in your own data. AIStor stores the documents, embeddings, and vector segments behind production retrieval-augmented generation (RAG) pipelines, with S3-native access on the same interface Milvus and the rest of the stack already speak.

106 million vectors indexed 12x faster with NVIDIA cuVS and RDMA, in a published end-to-end benchmark with Milvus

Pipeline-proven with NVIDIA NIM microservices, from embedding generation to segment storage

Scales with retrieval demand, no pipeline reconfiguration

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

Real-Time Agentic Inference

Reuse context instead of recomputing it. MemKV recalls shared KV cache in microseconds, so returning sessions can reuse cached context instead of recomputing the corresponding portion of the prefill. AIStor persists context, logs, and models durably beneath. Run either tier alone or both together.

Shared context memory across the GPU cluster, microsecond recall

Durable KV cache offload to object storage over RDMA

Long-context, multi-agent sessions at production scale

See the Benchmarks

Sovereign AI Infrastructure

Keep your data where it needs to stay. AIStor deploys entirely inside your own data center, including air-gapped environments, so the data foundation of your AI factory stays under your jurisdiction. Data residency holds while supporting high-performance GPU workloads.

On-premises and air-gapped deployment for strict data residency

Software-defined on industry-standard servers, edge to core

Encryption at rest and in transit, with keys that never leave your control

Proven Results

Measured performance across the workloads that define an NVIDIA AI Factory. Training, inference, and retrieval at scale.

Faster Checkpointing

Checkpoint saves and restores move directly between GPU servers and object storage with NVIDIA GPUDirect for S3-Compatible Storage, no host staging to slow access, at 33.5 GB/s per storage node with linear scaling. Checkpoint frequency becomes a schedule decision, not a storage constraint.

Shared Context Memory

HBM is per-GPU and finite; the fastest memory in the data center is also the scarcest. MemKV pools context on flash at petabyte scale, written once and recalled by any node serving the model, with microsecond context-retrieval latency, so returning sessions can reuse cached context instead of recomputing the corresponding portion of the prefill.

Direct-to-GPU Data Path

MemKV moves KV cache as raw blocks over NVIDIA NIXL, used by NVIDIA Dynamo for efficient KV transfer, with no filesystem or object API in the path. Blocks sizes match the engine's attention page, so recalled context lands in GPU memory ready to use.

GPU-Accelerated Vector Indexing

NVIDIA cuVS collapsed index-build compute from hours to minutes. Storage set the ceiling from there, and moving the data path to RDMA improved every remaining phase of the published pipeline,speeding up ingestion alone by 22.8%. Fast index builds are a pipeline property, not a GPU property.

AIStor in Production

From global vehicle fleets to factory floors, AIStor is the data layer powering GPU-intensive AI workloads at scale.

North American Automaker Builds Exabyte-Scale AI Training

A leading North American automaker ingests more than 6 petabytes of vehicle sensor and video data every day from over 100,000 vehicles, and GPU clusters training autonomous driving models are never waiting on data. AIStor handles it all on commodity hardware, at exabyte scale.

3+ exabytes of usable capacity

6PB ingested daily from 100,000+ vehicles

600,000+ S3 requests per second per cluster

300+ node clusters stood up in a single weekend

Read The Full Customer Story

Life Sciences Leader Industrializes Drug Discovery with AI

A pioneering TechBio organization replaced legacy parallel file systems with a unified, S3-compatible data lakehouse spanning on-premises, colocation, and public cloud environments. By repatriating primary storage, the company regained control over its proprietary research IP while significantly reducing cloud egress and operational costs.

20+ PB of AI research data managed across hybrid cloud environments

2.2M automated biological experiments conducted per week

3 PB of research data repatriated from public cloud

Top 500 ranked supercomputing cluster

Read The Full Customer Story

Our work requires high scalability, throughput, and efficiency in handling AI workloads. Based on preliminary testing, we believe that MinIO AIStor will increase efficiency in the CPU's on our AI compute infrastructure, ultimately enhancing the performance and economics in our data environment.

- Director, High Performance Compute Engineering and Operations

Life Sciences Leader

Learn More

We are excited to see MinIO bring AIStor to the NVIDIA AI ecosystem and to explore how AIStor and GPUDirect Storage together perform under the specific demands of our workloads.

- Director, High Performance Compute Engineering and Operations

Life Sciences Leader

Learn More

We chose MinIO over alternatives because of its scalability and performance.

- Data Architect

Leading North American Auto Manufacturer

Learn More

Built for NVIDIA AI Factory Workloads

AIStor supports the full spectrum of enterprise AI infrastructure, from data preparation to production inference to long-running agents.

These workloads are designed for NVIDIA AI Factory deployments using NVIDIA AI infrastructure.

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

KV cache offload for higher throughput

Shared context memory through MemKV

Enterprise RAG

Vector embedding storage at scale