Dell AI Data Platform explained

The Dell AI Data Platform is Dell's storage and data-software layer for AI: the products that hold an organisation's files and objects, plus the engines that prepare, index and query that data for models and agents. Dell calls it the data foundation of the Dell AI Factory with NVIDIA, and it is Dell's version of the NVIDIA AI Data Platform reference design that NVIDIA announced in March 2025 with DDN, Dell, HPE, Hitachi Vantara, IBM, NetApp, Nutanix, Pure Storage, VAST Data and WEKA.

This article describes what is in it as of October 2026, what has shipped, what is still a roadmap, and what to check before it goes into a design.

Two kinds of engine

Dell organises the platform into storage engines, where the data lives, and data engines, which do something with it.

Storage engines

Product	What it is	Status
PowerScale	Scale-out file storage (the former Isilon line) with NFS, SMB and S3 access. The F710 holds NVIDIA Cloud Partner certification.	Shipping. Support for up to 500 tenants per cluster and mTLS over NFS announced for November 2026.
ObjectScale	S3-compatible object storage, as an appliance or as software on PowerEdge servers.	Shipping.
Lightning File System	A new parallel file system for training and inference, with NVIDIA ConnectX and BlueField connectivity.	Announced March 2026; Dell's materials give conflicting timing.
Exascale Storage	A software-defined design that runs file, object, parallel-file and block storage on shared PowerEdge hardware.	Announced for the second half of 2026.

Data engines

Engine	Built on	What it does
Data Analytics Engine	Starburst (Trino)	SQL across data lakes, databases and object stores without copying the data, with a natural-language assistant.
Data Processing Engine	Apache Spark	Batch and streaming data preparation. GPU acceleration with NVIDIA cuDF on RTX PRO Blackwell Server GPUs is due in December 2026; Dell claims about 4× faster on average and up to 20× on batch tasks.
Data Search Engine	Elastic	Semantic and hybrid search for RAG over PowerScale and ObjectScale, with MetadataIQ discovery and NVIDIA cuVS indexing.
Data Orchestration Engine	Dataloop (acquired December 2025)	Discovery and ingestion, labelling and transformation pipelines, human review, a marketplace of models and NIM microservices, and support for NVIDIA's AI-Q blueprint. It keeps embeddings and indexes current as files change.

What was announced in October 2026

On 9 October 2026 Dell added a layer above the engines, due in the first half of 2027:

• A Unified Semantic Layer: shared business definitions and a glossary across structured and unstructured data.
• An Enterprise Knowledge Graph: the relationships between tables, data products, files and vector indexes, refined from metadata, lineage and query history, filtered by the user's permissions.
• Knowledge Agents: each limited to a part of the graph, with customer-set rules for data access, quality thresholds and spending limits.

Available now from the same announcement: the Dell Storage Performance Tool for sizing and comparing storage, and expanded implementation services. Apache Arrow becomes the data-movement format between storage and processing.

How it connects to OpenShift and RAG

Two integrations matter for the platforms we build:

• Dell Container Storage Modules. Dell's CSI drivers, including the PowerScale driver, are certified in the Red Hat Ecosystem Catalog and installed through an operator, with modules for replication, observability, authorisation and resiliency. This is how OpenShift workloads get PowerScale volumes.
• Search for RAG. The Data Search Engine is Elastic, and Elasticsearch is the default vector database of the NVIDIA RAG Blueprint. Dell also ships a PowerScale RAG Connector. A RAG pipeline on OpenShift AI can therefore index directly against Dell's search engine rather than running its own.

Questions to ask before it goes in a proposal

1. Which parts ship today? The storage engines and three of the four data engines are shipping; Lightning, Exascale, the GPU-accelerated Spark and the 2027 semantic layer are announcements. Tender responses should separate the two.
2. Whose numbers are these? The performance claims (6 TB/s per rack for Lightning, 4× for cuDF) are Dell's, from internal testing, with no third-party benchmarks published at the time of writing.
3. How much NVIDIA is required? The platform assumes NVIDIA GPUs, networking and software at every layer; that is a dependency on one supplier's pricing and delivery.
4. Does it replace what the customer already runs? Dell must show Dataloop and Starburst working with existing Databricks, Snowflake or Oracle estates rather than replacing them. Ask for the integration, not the demo.
5. What are the alternatives? HPE, NetApp, VAST Data and IBM offer similar stacks on the same NVIDIA design; the comparison belongs in the assessment.

Where it fits for us

Dell is already in our designs: the sovereign AI cloud proposals use PowerEdge XE9680 GPU servers, and PowerScale is the file tier in the Army data centre design. The open-source engines inside the platform, Spark, Trino and Elasticsearch, are the ones our labs teach, so the skills carry over whether a customer buys the Dell engines or runs the open-source versions on OpenShift. The AI data platform service covers both routes.