Dell Tackles AI Infrastructure With Disaggregated Servers And Storage

A report by Arm last month outlined enterprises’ enthusiasm for AI – 82 percent of the 655 business leaders surveyed said their organizations are using AI applications – and lack of the required infrastructure: 80 percent have budgets put aside for AI, but only 29 percent said they have the infrastructure that can automatically scale to meet AI demands.

What the infrastructure will look like is still being worked out. As AI training and inference jobs expand, the infrastructure will have to include distributed computing, energy efficiency (there’s a lot of talk about liquid cooling), datacenters with modular designs, a mix of powerful GPUs and CPUs, and large memory capacities, according to Arm.

The OEMs Chase GenAI, Hope For Profits

System makers are pushing hard in that direction. That includes Cisco – and its Universal Computing System (UCS) converged server-switching systems – and Hewlett Packard Enterprise. Cisco at its Partner Summit last fall unveiled AI hardware offerings that leaned on its UCS and validated designs, and in February announced a deal with AI giant – and datacenter competitor – Nvidia to give organizations a broader range of AI networking options that include technology from both vendors.

Also, in February, HPE rolled out the latest editions to its Gen12 ProLiant servers that include Intel’s “Granite Rapids” Xeon 6 CPUs, following up the launch of the first systems in the family aimed at AI workloads and jammed with Nvidia’s newest GPUs.

And speaking of Nvidia, at GTC the company introduced what executives said was an exaflop-capable rack system of servers based on the company’s GB200 NVL72 system and targeting AI and similarly compute- and power-hungry workloads.

Dell Is Up Next

Dell has been working on enhancements to its PowerEdge servers and various storage and data protection offerings to make them better for AI, which are being announced now, and promises more coming next month at its Dell Technologies World 2025 conference.

The new products address what Varun Chhabra, senior vice president of product marketing for Dell’s ISG unit, said the top challenges enterprises are trying to manage. They’re looking for infrastructure to help them balance the need to run traditional workloads like virtual machines, databases, systems of record – think ERP and CRM – but also AI, edge applications, and containerized jobs.

“Looking across this estate requires organizations to think very differently about their infrastructure strategy,” Chhabra explains. “What we hear from customers most often is that they are trying to architect their technology to avoid vendor lock-in. It is no longer just about supporting virtual machine-based workloads. It’s about virtual machines, containers, and bare metal. They’re also looking at making sure that they get the most out of their technology investments so that there’s efficient resource utilization, not just from a cost perspective, but also from a power and real estate space perspective.”

Distributed Datacenters

As AI takes hold in the enterprise, businesses are seeing their datacenter become increasingly distributed and they need to ensure their AI workloads have access to all of their data in a consistent format regardless of where it’s located. There needs to be a shared data layer across the entire infrastructure, which needs to be able to dynamically scale to meet the myriad changing demands of the workloads, he said.

It’s what Dell calls a disaggregated architecture, where compute, storage, and networking are pulled together to create shared and adaptable resource pools that can be applied to any workload organizations need to run and that deliver better scalability and availability. Chhabra said disaggregation is an evolution of a three-tiered strategy – independent scaling of compute, storage, and network, which is complex and often multi-vendor – and hyperconverged (HCI), which is simpler but locks organizations into a single vendor or operating system, limiting needed flexibility for scaling or dynamic workloads.