Dell VP says discrete beats disaggregated storage for AI

With VAST Data leading the disaggregated storage charge into the AI market, Dell exec David Noy says this is the wrong approach because "discrete architecture wins at AI scale." This is despite Dell espousing disaggregation for AI just nine months ago. 

Noy is a product management VP at Dell, with all the obvious biases that position includes. He's now making the argument that his company's PowerScale storage has been right all along.

A disaggregated storage architecture separates the controllers and compute from storage (the drive enclosures) so that each can be scaled up or down separately. This means you don't get wasted storage if you need more controller processing capacity, or wasted and idle compute if you simply need more storage capacity. 

When you have a discrete, integrated controller-and-drive chassis design, then you have to add both if you want more of each one. This is emphastically true if you have a monolithic, scale-up design. 

But Noy argues a clustered, scale-out design lets you add smaller discrete units when you run out of capacity or compute.

Noy writes that "for years, disaggregated architectures were positioned as 'flexible': compute in one layer, storage in another, connected by additional networks, switches, and orchestration logic." He asserts that PowerScale's discrete platform features "less rack space, fewer switches, and lower power consumption" than an equivalent disaggregated storage system.

He adds: "It is not surprising that vendors who initially championed disaggregation are now introducing discrete systems of their own..."

VAST Data's basic design uses stateless C-Nodes (controllers), each of which can connect to any of its D-Nodes (drive enclosures). The EBox concept was introduced in 2024 and stuffs a C-Node and D-Node in one server-style chassis, which does not allow the independent scaling of compute and storage.

VAST says this configuration is a good fit with the needs of hyperscalers that deploy thousands of a very specific server configuration, cloud providers that only offer virtual machine instances, and server vendors like Supermicro and Cisco who can deliver the VAST Data Platform to customers using their own servers.

Noy comments: "Although this shift is not controversial, it's simply an acknowledgement of the architectural limitation of disaggregated architectures for AI at scale." For example, with VAST, "performance effectively tops out once two CBoxes are paired with a DBox – meaning additional compute heads don't increase bandwidth."

There's more in his blog about this topic.

Ironically, the announcement of Dell's AI Data Platform with Nvidia in March last year, boasted: "Its disaggregated framework enables independent scaling of compute, storage, and data processing, giving organizations the flexibility to adapt swiftly to evolving demands."

Moreover, in May 2025, Varun Chhabra, SVP of infrastructure and telecom marketing at Dell, told Silicon Angle: "Disaggregated infrastructure is where the puck is going for our customers because it combines the flexibility of three-tier with the simplicity of hyperconverged to enable dynamic resource allocation from a shared resource." He was talking at the launch event of an expanded Dell AI Factory portfolio with infrastructure, software, and services updates at the Dell Technologies World event.

Now it's back to the future, Noy writes: "Now, as AI exposes the limits of layered, controller-heavy architectures, we're watching a familiar cycle repeat. Vendors are returning to discrete systems and framing them as something 'new.'"

He adds: "For those of us who have worked in enterprise infrastructure long enough, we recognize a pattern: what is framed as innovation often just turns out to be rediscovery."