Doll... Dell is adding agentic AI capabilities to its AI Data Platform, promising GPU-based data processing that runs 20 times faster in batch mode and four times faster overall than CPU-based processing. The updates, announced this week, target what executives call the real bottleneck in enterprise AI: data readiness, not compute access.

The release builds on findings from a Dell survey of 3,800 enterprise IT decision-makers and AI experts worldwide, conducted earlier this year. Respondents named data — its quality, availability, management, and security — as the top challenge in adopting and scaling AI.

Why is data the bottleneck?

Dell executives have pushed this message since launching the AI Data Platform two years ago. Varun Chhabra, senior vice president of Dell's Infrastructure Solutions Group, told journalists at a recent media briefing: "The infrastructure is ready, but the data isn't. The real bottleneck is not access to compute, it's not often access to models. It's actually access to data."

Chhabra said enterprise data sits fragmented across clouds, datacenters, file systems, databases, applications, and edge locations. "Much of it is trapped in silos based on workloads, and there is often not a unified way to reach it," he said. "It's often dark, which means it's effectively invisible to AI. It is often ungoverned, so the teams are stuck between AI that needs access and policies that require control."

The platform, a foundational element of Dell's AI Factory, organizes these capabilities in three layers:

• Data Orchestration Engine — takes in, prepares, labels, and enriches data, with unified pipelines, distributed control that decouples compute from storage, and native access to Nvidia NIM microservices, AI Blueprints, and templates.
• Data Engines — analytics, processing, and search capabilities that Chhabra said help organizations find the right data in hours instead of weeks.
• Storage Engines — PowerScale NAS for unstructured data and high-throughput AI workloads; ObjectScale with S3-over-RDMA, Nvidia CUDA libraries, and periodic AI model snapshots; and the Lightning File System, a software-defined parallel system announced at Nvidia's GTC 2026 show in March and released a month later for high-scale training and inferencing.

The platform runs on Nvidia technology, including Nemotron Retriever models for document parsing, embedding, and reranking, plus the open source cuVS library of GPU-accelerated algorithms for vector indexing and search.

What do the new features change?

The updates focus on giving AI agents shared context. Chhabra explained that agents accessing the same data currently must reconstruct their understanding and regenerate tokens on every query. "This costs extra tokens," he said. "Often that means the answers that you're getting out of this are not fully trustworthy. What all of [the new features drive] is fewer steps and lower costs because agents don't have to keep rebuilding context in every request."

Three components address this:

• Unified Semantic Layer — provides rules and definitions so terms mean the same thing wherever they appear, backed by a searchable glossary. Chhabra cited the word "defect" carrying different definitions across manufacturing plants. Dell also uses Nvidia's Auto-Ontology open source library to build knowledge graphs from enterprise data.
• Enterprise Knowledge Graph — maps relationships between structured and unstructured data using metadata, lineage, and query history. Chhabra said it can connect, for example, defect frequency to specific suppliers or distributors without forcing the model to rerun and generate tokens each time.
• Knowledge Agents — trusted per-topic agents built on the semantic layer and knowledge graph, with configurable rules for guidance, token spending limits, and data visibility. They use Nemotron Retriever models for reasoning and visual data understanding.

Knowledge Agents also let enterprises choose model sizes per use case, including general-purpose models in the cloud or open source models on-premises.

What about performance tooling?

Dell is adding Nvidia's cuDF GPU-accelerated library to the Data Processing Engine, alongside Apache Arrow for moving data efficiently between storage and processing so data can be queried where it sits. The result: GPU-based processing delivers 20 times faster batch processing and four times the data processing speed compared with CPUs. The cuDF-and-Arrow combination complements existing cuVS-augmented search and cuDF analytics.

Dell also released the open source Storage Performance Tool for ObjectScale and PowerScale. "Organizations can define the workload they want to benchmark our storage platforms with — checkpoint-style writes, high-concurrency reads, mixed read-write environments, even high-score queries," Chhabra said. Users run the same tool Dell Engineering uses internally.

The additions aim to close what Chhabra calls the "pilot reproduction gap" — use cases that succeed in demos but stall at scale due to data governance problems and dirty data. "That's really where the real enterprise scale is held back," he said.