Pure Storage Reference Architecture for NVIDIA Enterprise AI Factory Accelerates Intelligence at Scale

The Pure Storage AI Factory with NVIDIA is a validated AI infrastructure platform designed to accelerate enterprise AI projects. As AI models and reasoning systems grow more complex, this solution aims to speed up insights while simplifying operations.

Purpose-built for Modern Enterprise AI

Traditional IT struggles to meet the demands of modern AI workloads, continuously ensuring computing  and AI teams are productive. On the other hand, high-performance AI storage needs to incorporate enterprise capabilities such as guaranteed uptime, agility, and resilience. The Pure Storage solution for the NVIDIA Enterprise AI Factory validated design  addresses these issues with a simple-to-operate, integrated, high-performance, and resilient architecture. It supports advanced workloads such as  generative AI and computer vision by combining NVIDIA’s latest accelerated computing and software platforms with Pure Storage® FlashBlade//S™ storage, delivering the performance and scalability needed for large-scale model training and inference. The Pure Storage Flashblade//S is an NVIDIA-Certified™ storage system that can scale to support up to 1,024NVIDIA GPUs and multiple petabytes of data, with high-speed networking tightly integrating compute and storage. 

Key Benefits

• Non-disruptive upgrades: FlashBlade allows for data-in-place upgrades, enabling expansion or refreshes without downtime—unlike traditional storage that often requires planned outages.
• Ease of use: A unified platform simplifies the AI data pipeline from ingestion to inference, designed for enterprise IT and AI developers to operate with minimal hassle.
• Consistent performance: The architecture delivers reliable performance at any scale, allowing organizations to grow deployments without downtime or performance loss.
• Agility and future-proofing: The open, modular design supports evolving AI technologies and avoids vendor lock-in, maximizing flexibility and interoperability.
• Efficiency: The high density and lower power consumption of FlashBlade reduce data center footprint and operational costs, while the unified stack simplifies maintenance.

Validated Configurations 

Together, Pure Storage, NVIDIA, and other system partners are building products, software, and services to accelerate the enterprise IT shift to an AI Factory. Following the NVIDIA Enterprise AI Factory validated design, Pure is leading the charge to create a new class of on-premises infrastructure — featuring NVIDIA-Certified RTX™ PRO and HGX Blackwell servers, NVIDIA Spectrum™-X Ethernet networking, NVIDIA BlueField® DPUs, and NVIDIA AI Enterprise software paired with Pure’s NVIDIA-Certified storage systems. 

NVIDIA Enterprise AI Factory validated design solutions are based on recommended hardware configurations from NVIDIA Enterprise Reference Architectures (Enterprise RA), tailored for enterprise-class deployments ranging from 4 to 32 nodes with 16 to 256 GPUs. Each Enterprise RA follows a prescriptive design pattern, called a Reference Configuration, built around an NVIDIA-Certified server to ensure optimal performance in a cluster. These reference configurations standardize the description of compute nodes based on CPU, GPU, networking, and bandwidth specifications. The C-G-N-B nomenclature simplifies system selection by clearly indicating compute power, networking capabilities, and bandwidth performance. Each digit (e.g., 2-8-5-200) refers to the number of CPU sockets, number of GPUs, number of network adapters (NICs, SuperNICs, or DPUs), and average East-West bandwidth per GPU (GbE), respectively.

To simplify adoption, Pure Storage has aligned its systems with the reference configurations used in the NVIDIA Enterprise AI Factory validated design, enabling effortless deployment and scaling of balanced systems that avoid bottlenecks and suboptimal performance. Below are the repeatable reference configurations that Pure has designed and tested for. 

These represent validated building blocks that customers and partner OEMs can use to deploy the solution at different scales. Here’s what those mean:

• 2-4-3-200 (PCIe optimized): A compute node with 2 CPUs, 4 GPUs, 3 high-speed NIC/DPUs, and 200 Gbps networking bandwidth per GPU.It might correspond to an NVIDIA-certified server with four PCIe GPUs (NVIDIA L40S or the NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs , or NVIDIA H100 NVL for large models) and three 200Gb Ethernet interfaces (often leveraging NVIDIA BlueField-3 DPUs for offloads). The 2-4-3-200 building block offers a balanced entry point for AI factories—sufficient for tasks like small model training and fine-tuning or serving medium-sized models in production. Enterprises can start with a cluster of these 4-GPU servers knowing that the architecture will scale predictably.
• 2-8-5-200 (PCIe optimized): A larger PCIe-based node featuring 2 CPUs, 8 GPUs, 5 NIC/DPUs, and 200 Gbps network bandwidth per GPU. With eight GPUs in a single node, this configuration packs more AI compute per server, suitable for heavier training jobs or consolidated inference farms. This configuration aligns to the recently announced NVIDIA RTX PRO Server, with up to eight RTX PRO 6000 Blackwell Server Edition GPUs. The “2-8-5” design pattern ensures even more network bandwidth and adapter count to handle the increased GPU count, maintaining a full 200 Gb/s pipeline to each GPU. Use cases include medium-large model training (such as moderate NLP models or vision models) and high-throughput inference for many concurrent users. This configuration might be delivered via servers like NVIDIA-certified HGX A100/H100 PCIe systems or custom OEM designs that meet the spec. It provides an excellent scale-up option while still using standard Ethernet fabric.
• 2-8-9-400 (HGX optimized): A high-end configuration with 2 CPUs, 8 GPUs (in an NVIDIA HGX form factor), 9 network adapters, and 400 Gbps networking per GPU. This corresponds to an NVIDIA HGX platform (e.g., HGX H100 or upcoming HGX H200) where 8 SXM GPUs are connected via NVIDIA NVLink with ultra-high GPU-to-GPU bandwidth (on the order of 900 GB/s internally). The “9” network interfaces (such as BlueField-3 DPUs or NVIDIA ConnectX-7 NICs) provide massive aggregate throughput—8x 400Gb for the GPUs plus extras for storage or inter-cluster communication—effectively delivering 400Gb/s to each GPU for the most I/O-intensive workloads. This 2-8-9-400 pattern is aimed at the most demanding scenarios: large-model training (multibillion-parameter models), heavy inference on large context inputs, or any AI workload where maximum throughput and low latency are paramount. Given its power, it naturally suits use cases like advanced generative AI, complex simulation, or massive-scale inference serving.

These reference configurations can be scaled out cluster-wide. For instance, an AI factory might start with a (4) 2-8-5-200 nodes and later expand up to 32nodes per cluster—the FlashBlade storage and NVIDIA Spectrum-X Ethernet networking will scale accordingly (guidance is provided for how many FlashBlade chassis and uplinks are needed as one grows, ensuring linear scalability). This gives enterprises a clear, modular growth path: Start with a proven building block and add more as AI demand grows.

To ensure broad accessibility, NVIDIA and Pure Storage are validating solutions with major server OEMs to bring these AI Factory building blocks to market. Dell Technologies, HPE, and Supermicro servers will be tested with NVIDIA Blackwell platforms combined with Pure Storage FlashBlade as complete AI Factory solutions. This will accelerate NVIDIA AI Factory design and deployments with the customer’s vendor of choice and provide confidence that solutions are fully validated to NVIDIA-Pure Storage specifications.

The AI Factory is not a single black-box appliance, but rather an open reference architecture that multiple vendors can build upon. Enterprises can adopt the solution through the hardware vendor they prefer or even integrate components themselves following the reference design. This openness contrasts with some competing approaches that tie customers to one vendor’s ecosystem. In all cases, the customer benefits from the jointly validated design and tight integration work done by the engineering teams.

Unified Software Stack: NVIDIA Base Command Manager, Run:ai and Portworx by Pure Storage

A standout feature of the NVIDIA-Pure Storage AI Factory is its unified software stack for data and compute orchestration. Three key software components play a role here: NVIDIA Base Command Manager, Portworx® by Pure Storage, and NVIDIA Run:AI. Together, they form a cohesive layer that simplifies running AI at scale:

• NVIDIA Base Command Manager is a cluster management solution offering fast deployment and automates the provisioning and administration for clusters of any size. In the AI Factory, Base Command Manager provides provisioning, monitoring, and management capabilities in a single tool that spans the entire lifecycle of the cluster. By integrating Base Command Manager, the AI Factory ensures that organizations optimize the use of their cluster resources. The software includes dashboards for displaying cluster utilization and health data to help tune performance or catch issues (thus, “infrastructure optimization tools” are built). In short, NVIDIA Base Command Manager ties the hardware together into a single AI supercomputer from the user’s point of view, abstracting away the complexity of the underlying cluster.
• Portworx by Pure Storage is a Kubernetes-native data platform that excels at managing persistent storage for containerized applications. Within the AI Factory, Portworx can be leveraged to orchestrate data on FlashBlade for containerized AI workloads. As AI development increasingly uses container- and Kubernetes-based pipelines, Portworx ensures that data is available to containers wherever they run, with features like dynamic volume provisioning, storage snapshots, and replication across clusters. This is especially important in hybrid cloud scenarios—for instance, if some AI microservices run in the cloud while the main training happens on-prem, Portworx can help seamlessly move or sync data between environments. It effectively provides a cloud-ready layer on top of FlashBlade fast storage. The Portworx integration means developers don’t have to manually handle data logistics; they can request data volumes through Kubernetes and trust that the platform will deliver the data with the needed performance. In combination with the raw speed of FlashBlade, this ensures that containerized AI tasks (say, model inference services or data prep pipelines) get low-latency, high-throughput data access on demand.
• NVIDIA Run:ai is a specialized AI workload orchestration platform that maximizes  GPU utilization and simplifies job scheduling in the AI Factory.. Run:ai extends Kubernetes with intelligent, policy-driven scheduling capabilities that dynamically allocate GPU fractions, queue jobs, and enforce quotas enabling efficient sharing of GPU infrastructure across multiple users and teams. For example, if one experiment is only using a portion of a GPU’s memory, Run:ai can schedule additional workloads on the same GPU, increasing utilization in ways traditional schedulers cannot. The result is less idle time and higher workload throughput  within the AI Factory. Together, Run:ai and Portworx essentially create a single unified data and resource management plane across GPU and storage, providing lower latency data access and better resource utilization for AI jobs. This unification is a game-changer for productivity: Data and compute resources are orchestrated in tandem, automatically.

By leveraging Base Command Manager, NVIDIA Run:ai, and Portworx, the NVIDIA-Pure Storage AI Factory solution delivers a full-stack approach. It’s not just raw hardware; it’s hardware plus intelligent software. This means enterprises get a turnkey platform that not only has world-class performance but also the smart automation and integration needed to use that performance effectively. Developers can launch AI experiments without worrying about mounting file systems or finding where their data resides—the platform handles it. IT operators can ensure governance and efficiency with fine-grained control policies. The unified stack effectively turns the complex matrix of GPUs, network, and storage into a cohesive experience for end users inside the company.

Boosting Developer Productivity and AI Pipeline Efficiency

One of the most exciting aspects of the NVIDIA-Pure Storage AI Factory is how it can accelerate the work of AI developers and data scientists. By removing infrastructure bottlenecks and complexity, it allows teams to focus on building models and solutions rather than wrangling with hardware and data plumbing. 

Here’s how the AI Factory improves productivity across the AI pipeline:

• Simplified data access: In many enterprises, data is siloed across different storage systems, and getting the right data to feed into AI experiments is a slow, manual process. With the AI Factory’s unified FlashBlade storage, all stages of AI data—from raw ingest to refined features—reside in one high-speed repository. This means training clusters and inference servers can access data sets directly, without tedious copy or transfer steps. Moreover, the ability of FlashBlade to serve both file and object workloads means compatibility with a wide range of AI frameworks and data formats. Developers can use standard protocols (like NFS or S3) to read/write data at GPU speeds. The integration of RoCE networking further ensures that GPUs can pull data from storage with minimal latency and CPU overhead. In practice, this translates to shorter data loading times and the ability to iterate faster. A training job that might have stalled waiting for I/O now proceeds unhindered, keeping expensive GPU resources fully utilized.
• Minimal infrastructure tuning: Tuning infrastructure for AI can consume countless hours of IT effort. The Pure Storage with NVIDIA Enterprise AI Factory solution greatly reduces this burden. This reference architecture is pre-validated and balanced for AI workloads, meaning out-of-the-box performance is already optimized. NVIDIA Base Command software (included in the stack) provides tools to optimize GPU utilization and handle scheduling, so admins don’t have to script their own solutions for managing multi-user clusters. Overall, the platform’s design choices—like using standard Ethernet and RDMA, versus obscure proprietary interconnects—also mean it plugs into existing environments easily. IT teams spend less time tinkering and firefighting, and AI developers spend less time waiting. As a result, organizations can achieve faster iteration cycles—more experiments, model tweaks, and tests in the same amount of time—accelerating the path from idea to insight.
• End-to-end pipeline integration: The NVIDIA AI Factory with Pure Storage isn’t just about training; it’s a holistic approach covering data prep, training, validation, and deployment. For instance, consider a typical enterprise AI workflow: Data engineers ingest and curate large data sets, data scientists train models on GPU clusters, and ML engineers deploy those models for inference in production. Traditionally, each stage might occur on separate systems, requiring data to be moved and environments to be managed separately. In this solution, the entire pipeline can be unified. A single FlashBlade can host the raw data, intermediate pre-processed data, model checkpoints, and final model files, plus RAG embeddings and KV cache context for inference. The same GPU infrastructure can be partitioned (with appropriate scheduling) to handle both training jobs and inference services. This cohesion means that as soon as a model is trained, it can be tested and served to users on the same platform, drastically shortening deployment time. Furthermore, the consistent environment reduces errors when transitioning models from development to production. All of this leads to an agile AI development process where ideas swiftly progress to production deployments, without the traditional friction between siloed teams and systems.

Ultimately, by delivering performance with simplicity, the AI Factory lets enterprises achieve what was previously very difficult: an AI platform that is both blazingly fast and operationally efficient. This combination yields tangible business benefits—quicker insights from data, the ability to iterate on AI models more frequently, and a more productive development team that isn’t bogged down by infrastructure concerns.

FlashBlade: A Storage Backbone Optimized for AI Workloads

At the heart of the NVIDIA Enterprise AI Factory with Pure Storage is FlashBlade//S, an all-flash, scale-out storage platform engineered specifically for modern analytics and AI workloads. In this solution, FlashBlade functions as the high-performance data platform that underpins both AI training (which requires reading massive data sets efficiently) and AI inference (which benefits from fast, low-latency access to reference data and model files). The architecture of FlashBlade is uniquely suited to these challenges. It provides multi-dimensional performance and scalability; low-latency, parallel data access; non-disruptive and resilient operations; and the ability to unify file and object storage, enabling data scientists and engineers to work with data in whatever format their tools expect, without resorting to using silos of different storage.

Pure Storage FlashBlade provides an NVIDIA Enterprise AI Factory with a high-performance, highly efficient data foundation—delivering multi-petabyte capacity, tens of terabytes per second of throughput, and consistent low latency at scale. Just as importantly, it delivers this performance with the enterprise features (snapshots, encryption, replication) and ease of use that enterprise IT expects, unlike niche HPC file systems. 

Conclusion: Accelerating AI Innovation with Confidence

The launch of the NVIDIA Enterprise AI Factory with Pure Storage marks a significant milestone in the evolution of AI infrastructure. It combines the best of two worlds: NVIDIA’s cutting-edge AI computation and orchestration capabilities and innovative flash storage technology from Pure Storage. For enterprise AI leaders, this means they can finally deploy an AI platform that delivers supercomputer-level performance and enterprise-grade reliability/efficiency in one package. The platform’s ability to handle dynamic, multi-modal AI workloads with ease while simplifying operations and scaling increases organizational agility in a domain where speed matters.

By leveraging this jointly engineered solution, enterprises can transform their data centers into true “AI factories”—facilities that continually ingest data, train AI models, and produce insights or intelligent services that drive the business. The Pure Storage with NVIDIA Enterprise AI Factory reference architecture helps organizations remove the historical barriers that slowed these initiatives: No longer will data handling be a throughput roadblock, no longer will adding capacity require downtime, and no longer will developers be stymied by infrastructure quirks. Instead, they get a streamlined, high-octane platform to rapidly experiment and deliver AI results.

As the AI landscape continues to evolve (with ever-larger models and emerging applications in generative AI, real-time analytics, and beyond), having a flexible and powerful foundation is key. The AI Factory provides that foundation—performance at scale, adaptability for the future, and operational elegance built in. For any organization aiming to lead in the AI era, the NVIDIA AI Factory with Pure Storage offers a compelling path forward: Accelerate your AI projects today, with confidence that your infrastructure will meet the needs of tomorrow. This is AI infrastructure reimagined for agility, performance, and success.

Learn more:

• Download the Pure Storage with NVIDIA Enterprise AI Factory Reference Architecture
• Pure Storage AI solutions
• FlashBlade//S

*For 1,024+ GPU environments, explore FlashBlade//EXA™.

Try FlashBlade

No hardware, no setup, no cost—no problem. Experience the self-service capabilities of FlashBlade.

Take a Test Drive