Dell AI Factory 2.0: New servers with Nvidia B300 and proprietary file system

Contents

At its in-house trade fair, Dell presented new and expanded integrations with Nvidia, Intel, Google, Cohere, Hugging Face, Palantir, and ServiceNow. In addition, there were extensions to the Dell AI Data Platform, the Dell Data Lakehouse, and the new Project Lightning parallel file system. The goal is to bring AI closer to enterprise data. According to Michael Dell, 85 percent of companies plan to run generative AI workloads on-premises within the next 24 months.

More on-prem models and Partner Stacks

The Dell AI Factory with Nvidia remains a central building block. Dell has announced version 2.0 for this. It combines new AI servers, networking, storage, cooling, managed services, and software components for training, inference, and agentic AI. These include, among others, the PowerEdge systems XE9780L and XE9785L, each with eight Nvidia B300 accelerators, as well as support for Nvidia RTX Pro 6000. Dell is also expanding the AI Factory with Intel Gaudi 3 accelerators and the PowerEdge XE9680. Dell positions the platform as a validated complete solution with PowerEdge servers, PowerScale storage, PowerSwitch networking, services, and an open-source software stack, such as PyTorch, Kubernetes, Grafana, and Prometheus.

Dell is also expanding its offerings for models and agent platforms. Together with Google Cloud, Gemini models will be provided on-premises. The solution will be exclusively available to Dell customers and is intended to enable more control over data and infrastructure. With Cohere, Dell is also bringing the Cohere North agent platform to its infrastructure. North is designed to connect with Dell storage and all data sources.

Dell lists Hugging Face, Palantir, Reflection, ServiceNow, and SpaceXAI among others as validated on-premises solutions. In addition, there are validated AI solutions with Mistral, Fogsphere, Ipsotek, UneeQ Digital Humans, and Poolside, which can be provided through the Dell Automation Platform catalog. This covers a broad spectrum for Dell: models, agent platforms, development tools, workflow automation, computer vision, and various forms of data access.

However, on-premises models alone do not solve the central problem: agents, RAG applications, and domain-specific models require access to clean, discoverable, and quickly available data. This data is often distributed across file and object stores, databases, ERP and CRM systems, or at edge locations. This is where Dell's expanded AI Data Platform comes in. It is designed to unlock data sources, accelerate data flows, and make enterprise data more usable for AI applications. In the keynote, Dell COO Jeff Clarke described the platform as a new element of the AI Factory. It is intended to help prepare, organize, process, and protect data for AI workflows.

Lakehouse, PowerScale, and ObjectScale

One building block of the AI Data Platform is the Dell Data Lakehouse. It combines the scalability and cost-efficiency of a data lake with the structure and reliability of a data warehouse. A federated query engine is intended to make data discoverable, analyzeable, and processable across different environments. It is also intended to help automate orchestration for agents.

As storage foundations, Dell names PowerScale and ObjectScale. PowerScale is responsible for file and object storage for unstructured data and AI analysis, and ObjectScale for S3-compatible object workloads. Dell had already announced shortly before the event that it would integrate PowerFlex into Dell Exascale Storage. This means the architecture covers block access via PowerFlex, file workloads via PowerScale and Lightning, and object storage via ObjectScale.

This separation is relevant for AI applications because training, inference, and RAG workloads often require different data paths than traditional enterprise applications. While ERP and database data are often processed on a block basis, many AI workloads are dominated by unstructured data, file and object access, vector indexes, metadata, and context information.

Lightning File System for GPU Workloads

With Project Lightning, Dell announced a new parallel file system for AI and HPC workloads. It is designed for high throughput, low latency, and parallel access from many GPU nodes. Dell names training, inference, checkpointing, key-value caching, and metadata analysis as application areas, among others. Per node, Dell promises 97 GByte/s random read throughput and 97 percent bandwidth efficiency. The architecture uses RDMA networks and supports Nvidia GPUDirect Storage. This allows Lightning to achieve 67 percent faster data access and twice the throughput compared to the competition.

In addition, Dell is working with Nvidia on a highly scalable caching solution for accelerated inference. It integrates with Nvidia's NIXL library and targets agentic workloads and reasoning models with long thinking and context phases. Dell cites up to 100 times more tokens per second and 80 percent lower latency for this.

An On-Prem Platform for All of AI

Dell is evolving its AI Factory from a pure GPU and infrastructure platform to a comprehensive environment for enterprise AI. The partner stacks provide models, agents, and integrations; the AI Data Platform organizes and accelerates access to enterprise data. This aligns with the on-premises trend: AI should run where data, latency requirements, access rights, and governance specifications are located.

However, whether this works in heterogeneous environments depends less on the model catalog and more on data quality, governance, access rights, and actual data paths. While Dell is providing more building blocks for this, the actual work remains the same: companies need to know what data they have, how relevant it is in individual cases, and whether it is usable for AI at all.

(mki)