AI infrastructure changes the game when it comes to the role of storage. In the traditional enterprise world, storage was often separated by protocol and workload. The file storage for applications and users, object storage for cloud-native data lakes, block storage for databases, and sometimes a separate high-performance file system for HPC clusters.
As enterprises transition from traditional data centers into high-performance "AI factories,” data silos are inefficient and increasingly difficult to manage. The same dataset must cross different phases of an AI pipeline, starting from ingestion, preprocessing, training, fine-tuning, retrieval, inference, governance, and finally, long-term retention.
In this new architecture one of the most significant bottlenecks is the segmentation between file and object storage. Bridging this gap can simplify data pipelines, reduce unnecessary data movement, and accelerate AI initiatives.
One Dataset with Many Access Patterns
Modern AI pipelines rarely rely on one standard data-access pattern. The same dataset may be consumed by GPUs through high-performance file protocols such as NFS, or through S3 object APIs, depending on how the training, fine-tuning, inference, or data-processing workflow is designed.
Training and Shared Environments: NFS provides familiar file semantics for datasets, checkpoints, and shared development environments.
Large-Scale Ingestion: S3 provides scalable object access for distributed ingestion, preprocessing, and model artifacts.
RAG: Data can arrive through S3, be transformed and indexed, and then be accessed by downstream applications through file or object interfaces.
Data Science and Inference: Spark, Trino, Python, RAPIDS, and Kubernetes workloads may require different interfaces to the same underlying data.
A unified file and object platform lets each workload use the interface it needs while operating on the same dataset without requiring copies, conversions, or separate storage silos.
The Combined Benefit of File and Object Storage
In traditional setups, data often lives in separate storage silos. This increases operational complexity and slows down AI pipelines because data must be copied and updated between file systems, object stores and application specific repositories.
The VAST AI Operating System addresses this challenge by offering a truly unified multi-protocol platform. The VAST DataStore enables simultaneous access to the exact same dataset across both file (NFS over RDMA, SMB) and object (S3) interfaces. This convergence offers several benefits:
Zero Data Movement: Applications can write data through NFS and make that same data available through S3 without creating duplicate copies or building complex data-movement workflows.
Simplified AI Pipelines: Eliminates the need to stitch together multiple storage tiers, reducing operational complexity and TCO.
Native Multi-Protocol Performance: Unlike legacy gateway-based approaches, a native file and object architecture allows data to be consumed through the protocol that best fits the application without compromising performance or manageability.
Consistent governance and protection: When file and object data live in the same platform, enterprises can apply consistent policies for security, snapshots, replication, auditing, and long-term retention.
The Foundation: VAST DASE and NVIDIA ERA
VAST's unified file and object architecture, built on our Disaggregated, Shared Everything (DASE) architecture, complements the NVIDIA Enterprise Reference Architecture (ERA), providing a validated blueprint for deploying high-performance AI infrastructure. VAST enables storage resources to scale independently as AI workloads and GPU clusters grow, while maintaining a common data layer across file and object access.
The power of this architecture is its ability to support varying scales and configurations of NVIDIA-Certified Systems. The integration of VAST's object storage is validated to meet the demanding requirements of different AI use cases:
The 2-4-3-200 Architecture: Each node is equipped with 2 CPUs, 4 PCIe GPUs (such as L40S, RTX PRO 6000, H100 NVL, or H200 NVL), and 3 high-performance network adapters each delivering 200 Gbps of bandwidth.
The 2-8-5-200 Architecture: Engineered for multi-node AI or hybrid applications, this configuration increases density and computational power per node. It features 2 CPUs, 8 PCIe GPUs, and 5 network adapters per node.
Empowering Neoclouds and AI-Forward Enterprises
Neocloud providers must plan for diverse customer requirements and storage access patterns. One customer may require S3 for a cloud-native training pipeline; another may rely on NFS for PyTorch or shared research environments; another may need Kubernetes-native provisioning and management. Enterprises add existing file shares, S3 buckets, databases, data warehouses, analytics tools, and backup systems to the mix.
Supporting those requirements with separate storage systems increases infrastructure complexity, cost and makes data harder to share across workloads. A unified file and object platform with secure multitenancy allows different applications and teams to use the interfaces they need while maintaining a common data layer, helping organizations modernize their infrastructure without forcing every application to change at once.
Different customers and workloads need different interfaces. The infrastructure shouldn't require different copies of the data.
A Unified Storage Foundation for the AI Factory
AI factories require more than just a new storage model, they require a unified data platform. File storage remains essential for training, shared workspaces, checkpoints, analytics, and enterprise applications. Object storage is essential for cloud-native applications, data lakes, model artifacts, RAG pipelines, tenant services, and long-term retention. By unifying file and object storage on a single AI data platform, VAST Data and NVIDIA ERA provide a practical blueprint for next-generation AI data centers.
Download the white paper: NVIDIA ERA for Object Storage with VAST Data
