Dell Expands Its AI Data Platform With Data Context, Processing, and Storage Features

Data remains a barrier to enterprise AI adoption

Dell recently published survey results from 3,800 enterprise IT decision-makers and AI experts worldwide. Respondents identified data quality, availability, management, and security as the leading challenges to adopting and scaling AI.

Dell executives have emphasized this issue since the company launched its AI Data Platform two years ago. According to Varun Chhabra, senior vice president of Dell’s Infrastructure Solutions Group, organizations generally have access to infrastructure and models, but their data is often not prepared for enterprise-scale AI workloads.

“Enterprise data is just not in a place where it's ready to support scaling of AI workloads,” Chhabra said at a media briefing.

AI systems often need data distributed across clouds, datacenters, file systems, databases, applications, and edge locations. Much of this information remains isolated in workload-specific silos. It may also be unstructured, difficult for AI systems to discover, or subject to governance policies that limit access.

The three layers of Dell’s AI Data Platform

Dell’s AI Data Platform is a foundational component of the company’s AI Factory. It is organized into three primary layers.

The Data Orchestration Engine ingests, prepares, labels, and enriches data. It provides unified data pipelines, distributed control that separates compute from storage, and native access to Nvidia NIM microservices, AI Blueprints, and templates.

The Data Engines support analytics, processing, and search. Dell says they help organizations identify and use relevant data in hours rather than weeks.

The Storage Engines include several systems:

• PowerScale: Network-attached storage for unstructured data and high-throughput AI workloads.
• ObjectScale: Support for S3-over-RDMA and Nvidia CUDA libraries, along with large object repositories and periodic snapshots of AI models.
• Lightning File System: A fast, software-defined parallel file system announced at Nvidia’s GTC 2026 event in March and released the following month. It is designed for high-scale training and inference.

The platform also incorporates Nvidia technology, including CUDA libraries, NIM microservices, Nemotron Retriever models for document parsing, embedding, and reranking, and cuVS, an open source library of GPU-accelerated algorithms for vector indexing and search.

Dell’s objective is to provide unified storage layers with exabyte-scale capabilities so GPUs do not remain idle while waiting for data. The approach also addresses what Chhabra calls the “pilot reproduction gap,” in which an AI use case works in a demonstration or small pilot but encounters data quality and governance problems during broader deployment.

New features for agentic AI

Dell is adding capabilities focused on agentic AI. The features are intended to provide agents with shared context and consistent definitions for terms found across structured and unstructured data. They also make it easier to find and use that context.

Agents can access the same data, documents, and history, but they may need to reconstruct their understanding and regenerate tokens each time a query runs or additional data is analyzed. Dell says this can increase token usage, create repeated exchanges with human users, slow responses, and reduce confidence in the results.

The new capabilities are designed to reduce repeated context-building, compute requirements, and token generation.

Unified Semantic Layer

The Unified Semantic Layer provides common rules and definitions for terms wherever they appear. For example, the term “defect” might have different meanings at two manufacturing plants. The layer includes a searchable glossary that allows agents to interpret the term consistently across both locations.

Dell is also using Nvidia’s Auto-Ontology open source library to create knowledge graphs from enterprise data.

Enterprise Knowledge Graph

The Enterprise Knowledge Graph maps relationships between structured and unstructured data. It uses metadata, data lineage, and query history to continuously refine the graph.

The semantic layer defines the meaning of a term, while the knowledge graph connects that term and its associated data to information in other systems. For example, it can help relate the frequency of a manufacturing defect to a particular supplier or distributor without requiring a model to rebuild those connections and generate new tokens for every request.

Knowledge Agents

Organizations can use the semantic layer and knowledge graph to create Knowledge Agents focused on specific topics. Rules can define the guidance an agent follows, the number of tokens it can use, and the data it can access. The agents also use Nvidia Nemotron Retriever models for reasoning and visual understanding of data.

Knowledge Agents can support different choices of AI models for different use cases. Enterprises can select general-purpose models in the cloud or open source models deployed on-premises.

GPU-accelerated data processing

Dell is adding Nvidia’s cuDF GPU-accelerated library to its Data Processing Engine. It is also integrating Apache Arrow to move data efficiently between storage and processing systems, allowing data to be queried where it resides.

These additions enable organizations to use GPUs instead of CPUs for data processing. Dell reports that GPU-based processing can deliver batch processing up to 20 times faster than CPU-based processing and provide four times the data processing speed.

The cuDF and Apache Arrow integration complements the cuVS-enhanced search and cuDF analytics already available in Dell’s Processing Engine.

Storage performance benchmarking

Dell has introduced the open source Dell Storage Performance Tool for ObjectScale and PowerScale. The tool is intended to help organizations measure storage requirements for AI infrastructure more accurately and maintain data flow to GPUs.

Users can define the workloads used to benchmark Dell storage platforms. Supported scenarios include checkpoint-style writes, high-concurrency reads, mixed read-write environments, and high-score queries. The benchmarks can run on an organization’s own infrastructure using the same tool Dell Engineering uses internally to evaluate its products.