Dell PowerScale for Azure Goes GA as AI Data Platform Adds cuDF, mTLS and Knowledge Graphs

Dell announced the update on October 6, 2026. It has five parts:

• a semantic layer
• a knowledge graph
• topic-specific agents
• GPU-accelerated data prep
• storage and security upgrades

Most of it is not shipping yet. Here is what is available now, what is scheduled, and where the claims need caution.

The pitch: data without context is just noise​

Dell's argument is that enterprises have spent years making data accessible without making it usable. Blocks & Files quotes Arthur Lewis, president of Dell's Infrastructure Solutions Group, as saying an agent that can find a customer record but doesn't know what it means or whether it can be trusted is "just fast."

Varun Chhabra, a Dell senior vice president, told The Next Platform that "the infrastructure is ready, but the data isn't." He also described a "pilot reproduction gap." In that gap, a demo works, but scaling it up runs into governance problems and dirty data.

Chhabra also made a cost argument. Agents that rebuild context on every query burn extra tokens and trigger back-and-forth with humans. Dell says shared context should mean fewer steps and lower compute costs. That is Dell's claim, and the reporting does not give measured token savings.

The three new context layers (1H 2027)​

All three are due in the first half of 2027.

Unified Semantic Layer. It applies shared business definitions, rules and a searchable glossary to structured and unstructured data.

• Dell's example is that one system's "client" and another's "account" can be recognized as the same entity.
• Chhabra's example is that "defect" may mean different things in two plants.
• Customers can import existing ontologies and taxonomies.
• Dell is also enabling Nvidia's open-source Auto-Ontology library, which builds knowledge graphs from enterprise data.

Enterprise Knowledge Graph. It maps how data is related, using metadata, lineage and query history, and keeps tuning itself as activity changes.

• When an agent asks a question, the platform pulls in related tables, data products, multimodal data and vector indexes that the agent is permitted to see.
• Dell's example is tracing an odd sensor reading to a machine, its repair history, the supplier batch and the orders at risk.

Knowledge Agents. Each agent covers one topic, grounded in a defined slice of the graph.

• Customers set what guidance it follows, what data it can see, what quality bar it must meet and how much it may spend.
• Nvidia Nemotron Retriever models handle reasoning and visual understanding.
• SiliconANGLE reports that all three components run inside the customer's own data center, because they hold sensitive information.

Reality check. Permission scoping and a spending cap are guardrails. They do not make an agent's answers correct. A knowledge graph is only as good as the metadata and lineage behind it, and a messy estate will produce a messy graph. IT teams should treat these as promising 2027 features to pilot, not as finished products.

GPU data prep: cuDF and Arrow (December 2026, more in 1H 2027)​

Dell is adding Nvidia's cuDF to its Data Processing Engine. cuDF is an open-source, GPU-accelerated library for tabular data, built on Apache Arrow's columnar memory format. Filters, joins, group-bys, sorts and aggregations can run across many GPU cores with less CPU-to-GPU data movement. Apache Arrow is meant to move data between Dell storage and the engine so jobs can query it in place.

Dell's numbers need careful handling:

• Dell says the engine, running cuDF on Nvidia RTX PRO 4500 Blackwell Server Edition GPUs, is nearly 4 times faster on average than CPUs alone.
• It also claims up to 20 times faster on batch workloads.
• SiliconANGLE adds the test detail. Dell's September tests used a PowerEdge R770, with Spark jobs averaging 3.9 times faster and a batch data-mining job reaching 20.4 times.
• Those tests used default settings with no tuning.

These are vendor tests. Your speedup will depend on your workload mix, data shapes and configuration, and the 20x figure is the best case, not the typical one. The Nvidia acceleration stack is due in December 2026. Further Arrow acceleration is expected in the first half of 2027.

PowerScale: 500 tenants and mTLS (November 2026)​

PowerScale will support up to 500 tenants per cluster. Each tenant gets more granular role-based access control. PowerScale will also be able to encrypt and authenticate NFS file traffic with mutual TLS. SiliconANGLE says Dell is pitching this at AI service providers and enterprises running shared platforms. The enhancements are scheduled for November 2026.

For admins, mTLS over NFS matters because plain NFS has long been a weak spot in zero-trust designs. Tenant-level RBAC also matters for anyone consolidating teams onto one cluster.

The Azure angle: managed PowerScale​

Dell's cloud-native PowerScale service for Azure is available now, starting with a US East region, according to Blocks & Files. Microsoft's own posts add detail:

• Microsoft announced "Dell PowerScale for Azure – Dell managed" as generally available on September 29, 2026.
• It is an Azure Native Integration that brings PowerScale and the OneFS operating system into Azure.
• Dell handles deployment, monitoring, maintenance and upgrades.
• Admins can provision and manage it from the Azure portal, with Azure Resource Manager integration and billing on the Azure invoice.
• It integrates into an Azure Virtual Network via VNet Injection, with data encrypted at rest using Microsoft-managed keys.
• Dell's solution brief says it supports NFS, SMB and S3 with unified permissions.
• Dell's brief also lists up to 8.4PB usable capacity in a single namespace, three performance tiers and 1-year or 3-year terms.

Dell's performance comparison is a marketing claim. It says the service delivers 4 times the performance, a 4 times larger namespace and 2 times the cluster resiliency of its "closest competitor." Blocks & Files understands that competitor to be Qumulo's Azure Native Qumulo. I found no published methodology or independent benchmark for those comparisons, and the competitor identification is the reporter's understanding, not something Dell confirmed. Do your own proof of concept before basing a purchase on them.

Also note the Dell service description in the research notes says it excludes backup, disaster recovery, migration and identity and access management. Those stay on the customer.

Storage Performance Tool (available now)​

Dell's open-source Storage Performance Tool is on GitHub. It benchmarks S3-compatible object storage, and Dell says it covers ObjectScale and PowerScale. Chhabra says users can define workloads such as:

• checkpoint-style writes
• high-concurrency reads
• mixed read/write patterns
• Iceberg queries

He says it measures throughput and latency live, verifies persisted data end to end, and records versions and provenance so runs are repeatable. He also says it is the same tool Dell Engineering uses internally.

Per the research notes, the project documentation points to a dedicated benchmark bucket. Write and mixed workloads modify the target, and mixed runs include DELETE operations. Do not point it at production data.

A benchmark tool helps with sizing and with comparing vendors on equal terms. It does not prove a system will sustain your production workload.

Services​

Dell is expanding Professional Services with AI Data Platform implementation across its data and storage engines. The aim is a production-ready foundation, with analytics, processing, search and orchestration activated and tuned. Blocks & Files lists them under the announcement. SiliconANGLE says customers can get the new services today.

Timeline​

What it means for IT teams​

• Azure shops now have a Dell-managed OneFS option billed through Azure. Check region availability first, since the Blocks & Files report says only US East at launch.
• Teams planning on-prem GPU analytics get a December target for cuDF in Dell's engine. Benchmark on your own data before counting on the headline numbers.
• Multi-tenant file environments should look at the November mTLS and RBAC changes. They address a real security gap.
• Data and AI leads should treat the semantic layer and graph as 2027 pilots, and start now on the metadata and lineage they depend on.

Dell's thesis is sound: AI projects stall on data readiness. But most of the headline features are announced, not shipped, and the speed and competitor claims are Dell's own. The benchmark tool is the most practical item here, because it lets you check the claims yourself.