Enterprises accelerating their AI initiatives increasingly face two bottlenecks: getting data close to compute and keeping high-value GPU investments continuously fed during training and inference. Azure NetApp Files addresses both challenges by serving as an enterprise file system integrated across all four layers of the Azure AI stack: Infrastructure, Managed Application Services, Data Fabric, and End User Integration.

For domain-specific model training, Azure NetApp Files provides HPC-class, ultra-high throughput and ultra-low-latency checkpointing, ensuring that NVIDIA GPU clusters running on Azure never stall. For AI inference, Azure NetApp Files delivers stateful, high-performance persistent storage for Azure Kubernetes Service (AKS), giving containerized workloads, including NVIDIA NIM microservices, consistent performance and automated provisioning.

With the rise of Microsoft Foundry, Model Context Protocol (MCP), and agentic AI architectures, Azure NetApp Files plays an even more strategic role. Through RESTful API integration, Azure NetApp Files serves as the unified data layer that feeds Microsoft Fabric, Microsoft Foundry, and NVIDIA’s agentic AI stack (NVIDIA NeMo, Nemotron open models, and the AI-Q Blueprint), creating a seamless, scalable foundation for next-generation AI workloads.

Co-authors:

Andy Chan, Azure NetApp Files, Principal Product Manager
Ranga Sankar, Azure NetApp Files, Product Manager
Vinay Maddi, GTM - Agentic AI | PAI, NVIDIA
Charlie Huang, Sr. Product Marketing Manager, Accelerated Computing, NVIDIA
Jitesh Mani, Sr. Product Manager Azure Storage, Microsoft

Data Mobility Before Compute and AI

AI initiatives often stall before training begins because the data and compute resources are not in the same location. GPU capacity is constrained and frequently available in regions different from where enterprise data resides. Azure NetApp Files provides multiple mechanisms to close that gap:

Azure NetApp Files cache volumes for low-latency access to large datasets
Migration assistant for seamless data transfer with preserved permissions
Cross-region replication for global AI deployments
High-performance access in Azure to on-premises and multi-cloud data powered by Azure NetApp Files cache volumes.

These capabilities help organizations to:

Improve data hygiene and consistency
Move data efficiently without compromising security
Position data as close as possible to NVIDIA GPU clusters and AI services
Accelerate distributed AI workflows with fast access to datasets across regions and environments
Simplify dataset management by updating data once at the source while remote locations access cached content

By reducing data movement overhead, Azure NetApp Files ensures that AI pipelines start fast and stay fast. Organizations can centrally manage training and inference datasets while giving geographically distributed AI teams, GPU clusters, and AI services high-performance local access, without sacrificing consistency or operational efficiency.

Azure NetApp Files: Your Data Gateway to Advanced AI Services in Azure

AI success begins and ends with data. Azure NetApp Files provides a unified, enterprise-grade file system that supports the full lifecycle of AI workloads:

Data mobility before compute/AI
Data performance during compute/AI
Data integration with NVIDIA’s AI software stack, Microsoft Fabric, Microsoft Foundry, and containerized AI services

Azure NetApp Files’ “one-stop-shop” model, combined with a comprehensive portfolio of services and dedicated cloud solution architects, ensures customers can onboard quickly, scale confidently, and operate efficiently across Azure’s AI platform.

Where Azure NetApp Files meets the stack

Stage	Azure NetApp Files	NVIDIA	Azure
Data mobility	Caching, migration assistant, cross-region/zone replication	GPU capacity, wherever it is available	Azure regions with GPU capacity
Training	Ultra-low latency checkpointing, high throughput NFS, QoS, short-term clones	Blackwell-generation GPUs; NeMo for data curation, fine-tuning, evaluation	ND GB200 v6 and ND GB300 v6 VMs
Inference	Trident CSI persistent volumes, NFS and SMB, QoS-controlled throughput	NIM microservices, Nemotron models	AKS
Agents	Object REST API – Integration with AI services with no copy and in-place data access.	NeMo Retriever, RAG Blueprint, AI-Q Blueprint, NeMo Agent Toolkit	Microsoft Foundry, Microsoft Fabric

AI Domain-Specific Model Training with HPC Infrastructure

“Azure NetApp Files scales dynamically to keep those high-value GPUs busy.”

AI training is fundamentally a throughput problem. High value GPU clusters must be continuously fed with data. Any storage-related stall directly impacts cost and time to result.

Azure NetApp Files solves this with:

Ultra-low-latency checkpointing
High-throughput NFS performance at scale
Dynamic QoS controls to match GPU demand

For training domain-specific LLMs and proprietary models, Azure NetApp Files provides HPC-class performance. Azure NetApp Files provides scalable storage performance designed to support fine-tuning and experimentation with large AI models.

Built to keep NVIDIA GPUs fed

Azure delivers NVIDIA Blackwell-generation systems, such as GB200 and GB300 NVL72, through ND GB200 v6 and ND GB300 v6 VMs, with NVLink and NVIDIA Quantum InfiniBand networking for large-scale training. NVIDIA NeMo adds the software to curate data, fine-tune Nemotron and other open models, and evaluate results. All of it consumes data at very high rates. Azure NetApp Files is the data layer that sustains those rates, so checkpoint writes and dataset reads never leave this GPU capacity idle.

Short-Term Clones for Rapid Experimentation

Azure NetApp Files short-term clones allow teams to instantly create writable dataset copies that are space-efficient, with no full copies. This is ideal for:

Training multiple model variants
Running parallel experiments
Creating isolated test environments

Clones ensure GPU clusters always have the data they need, without waiting for slow, full dataset copies.

Impact for Training Workloads

SaaS-like deployment with high throughput and predictable performance
HPC storage management abstracted away, so teams focus on models rather than file system tuning
Optimized cost efficiency by keeping NVIDIA GPU clusters fully utilized

Reference Use Case #1: Lightweight LLM Training on Azure NetApp Files

A global technology provider recently trained a lightweight local-language LLM with fewer than 20 billion parameters on Azure using Azure NetApp Files. The organization needed:

Low-latency storage for aggressive checkpointing
Regional availability aligned with GPU capacity
Scalable performance to keep pace with ever-increasing model parameters
Predictable throughput for both training and inference pipelines

Highlights

Low-latency, predictable throughput designed to keep GPUs productive
Scalable storage to support growing model sizes
Foundation for future domain-specific AI models

Business Challenge

The team faced GPU availability constraints and required a storage solution capable of sustaining high-performance checkpoint writes. Their objective was to optimize cost, performance, and scalability while ensuring seamless integration with existing AI workflows.

Solution

The organization deployed its lightweight LLM in an Azure region with available GPU capacity and selected Azure NetApp Files Ultra service level to support the training pipeline. The ultra service level provided low-latency checkpointing, high throughput for large single-file writes, and scalable capacity as model parameters increase.

During training, frameworks like TensorFlow and PyTorch often generate large, monolithic checkpoint writes, which can overwhelm software-defined storage. In this deployment, Azure NetApp Files demonstrated superior write latency, ensuring GPUs remained fully utilized.

Business Outcomes

Faster training cycles
Improved GPU utilization
Eliminated storage bottlenecks
A scalable foundation for future domain-specific LLM projects

Azure NetApp Files for Containers and AKS: High-Performance Storage for AI Inference and Agentic AI

As enterprises shift toward containerized, cloud-native architectures with Azure Kubernetes Service (AKS), Azure NetApp Files ensures consistent throughput, low latency, and dynamic provisioning for real-time AI applications. AI inference workloads increasingly run on containerized platforms such as AKS and OpenShift, and these environments demand storage that is stateful, high performance, dynamically provisioned, and easy to automate.

Azure NetApp Files meets these requirements with:

Trident CSI driver for automated persistent volume creation
NFS and SMB support for both Linux and Windows containers
QoS-controlled throughput to ensure inference pipelines remain responsive

Azure NetApp Files with NVIDIA’s agentic AI stack

NVIDIA NIM microservices package optimized inference engines as containers that run on NVIDIA GPUs in Azure and on AKS. NVIDIA NeMo, the Nemotron open models, and Blueprints such as the RAG Blueprint and AI-Q give teams a production path from model to agent. Every one of these components needs fast, shared, governed access to enterprise data, and this is where Azure NetApp Files fits:

Persistent storage for NIM and agent services. Trident-provisioned Azure NetApp Files volumes give containerized inference and agent workloads durable, shared storage that scales with them.
Fast retrieval for RAG. Low-latency shared storage can back the document corpora and indexes that NeMo Retriever and the NVIDIA RAG Blueprint depend on, so agents answer from your authoritative data.
A data flywheel. Usage data captured for NVIDIA’s Data Flywheel Blueprint, used to keep improving agent accuracy, latency, and cost, needs a durable, high-performance home. Azure NetApp Files provides it.
Hybrid reach. For data that lives on premises, NetApp’s AI Data Engine and AFX systems integrate with the NVIDIA AI Data Platform reference design, so the same data-to-agent pattern works wherever data resides.

Azure NetApp Files with Microsoft’s AI platform

The Azure NetApp Files object REST API lets Azure services such as Microsoft Fabric, Foundry, Azure OpenAI, Azure Databricks, and Azure AI Search reach NFS and SMB datasets directly, without copying them into a separate object store. This unified approach simplifies deployment, accelerates inference performance, and provides the reliability needed to scale intelligent services across Azure and edge environments.

Reference Use Case #2: AI Inference Modernization and Consolidation

A global consumer solutions provider recently modernized its AI inference platform by migrating to Azure. The organization selected Azure NetApp Files and AKS to build a unified, high-performance architecture capable of supporting next-generation personalization services.

Highlights

AI inference modernization using Azure NetApp Files as the high-performance NFS backbone
Performance uplift and complexity reduction compared to other solutions, especially for random and small-write data patterns
Unified data fabric via object/file duality, ingesting data from Azure and edge sources
AKS-based AI inference services delivering faster, more reliable recommendations at global scale

Business Challenge

The organization’s existing recommendation platform spanned multiple clouds and on-premises data centers with fragmented data sources, creating performance constraints, operational inefficiencies, and rising infrastructure costs. To scale personalization globally, they needed a unified data foundation, high-performance shared storage for inference, and a cloud-native architecture based on Microsoft Fabric capable of supporting advanced AI workloads.

Solution

Azure NetApp Files provides the ultra-low-latency shared file storage, scalability, and Kubernetes integration required for production AI inference. Other services were integrated as a single fabric to improve operational effectiveness and reduce complexity:

Azure Data Lake and ADLS Gen2
Azure Databricks integration
Microsoft Fabric for unified governance
Microsoft Foundry, enabling MCP pipelines and agentic AI workflows

Business Outcomes

Reduced infrastructure complexity
Improved inference performance
A unified, API-driven data fabric with seamless ingestion of multicloud and edge data
Enabled MCP and agentic AI workflows
A foundation for next-generation AI capabilities, including NVIDIA-accelerated agentic services

What Azure NetApp Files delivers, together with NVIDIA and Microsoft

For infrastructure leaders: fewer moving parts. One managed, enterprise-grade data layer that plugs into NVIDIA’s GPU and software stack and Azure’s AI platform, so less integration work and fewer places for performance to leak.
For executives: faster time to value and better return on GPU investment. Higher utilization during training, scalable serving in production, and agents grounded in trusted enterprise data.
For AI teams: a continuous path from data to trained model to deployed agent, with the flexibility to iterate quickly.

Conclusion

Azure NetApp Files is uniquely positioned to support the full lifecycle of enterprise AI: from data mobility, to HPC-class training on NVIDIA GPU infrastructure, to containerized inference with NVIDIA NIM on AKS, to modern agentic AI architectures built on NVIDIA’s agentic AI stack, Microsoft Foundry, and Microsoft Fabric. Its combination of high performance, low latency, dynamic scaling, and deep integration with the NVIDIA and Azure AI ecosystems makes it a versatile file system for organizations building domain-specific AI solutions.

Whether training proprietary models or deploying inference pipelines and agents across AKS, Azure NetApp Files ensures that data is always close, always fast, and always ready for the next stage of your AI journey.

Ready to get started

Explore how Azure NetApp Files can simplify data access across training, inference, and agentic AI workflows while delivering the performance and scalability required for enterprise-grade AI. Work with your Microsoft and NetApp teams to evaluate your data architecture, identify high-value AI use cases, and develop a deployment strategy that leverages Azure NetApp Files with NVIDIA AI, Microsoft Foundry, Microsoft Fabric, and AKS. Contact us at askanf@microsoft.com to get started.