WWT & DDN launch AI storage design for NVIDIA Vera Rubin

World Wide Technology and DDN have introduced a storage reference design for AI factory campuses of 50 megawatts and above, built around NVIDIA's Vera Rubin NVL72 architecture.

The system is intended to help operators manage a basic constraint in large AI facilities: power used by storage is power that cannot be used by GPUs, yet storage must still deliver data fast enough to keep those GPUs busy.

Under the design, storage is sized for NVIDIA's 72-GPU rack format and targets more than 2.5 GB/s of read throughput per GPU. A single 2U DDN AI400X3M appliance is rated at 190 GB/s of read throughput, versus the 180 GB/s required by a full rack.

That allows operators planning large campuses to map storage and infrastructure needs alongside GPU deployment and site power limits, rather than treating storage as a separate layer later in the design process.

Power trade-off

The announcement reflects a wider issue in AI infrastructure, as developers try to fit more compute into fixed power envelopes. As GPU deployments grow, supporting systems including storage and networking are drawing closer scrutiny because they also consume power and space in data centres.

The reference design covers both training and inference workloads. Data and checkpoints sit on DDN EXAScaler, while KV cache is handled on DDN Infinia. The setup also uses NVIDIA Dynamo for KV-cache placement between tiers, and an NVIDIA Spectrum-X Ethernet network based on Spectrum-6 SN6600 switches, ConnectX-9 SuperNICs and BlueField-4 STX storage processors for the context-memory tier.

A central part of the design is KV cache, the working memory a model builds while processing a prompt. In long-context, reasoning and agentic inference workloads, that cache can grow beyond available GPU memory.

If the cache is discarded and no offloaded copy is kept, the system must recompute it when the information is needed again. That adds to GPU workload and can delay responses.

The companies said NVIDIA CMX creates a context-memory tier between GPU memory and shared storage using BlueField-4 STX. DDN Infinia runs on that infrastructure, allowing cache to be shared across inference nodes so completed work can be reused instead of recalculated.

DDN said it has recorded KV-cache loading speeds of up to 55 times faster than recomputation. In one 112K-token test cited by the company, loading took 2.1 seconds, compared with 57 seconds to recompute, though results depend on workload and configuration.

Integration role

WWT's role is to integrate components from different suppliers into a single design. The aim is to reduce the burden on customers that would otherwise need to reconcile separate vendor specifications across storage, networking and infrastructure planning.

Felicien Mathieu, Chief Technology Officer, Middle East and Africa, at WWT, said customers in the region are making multiple infrastructure decisions at the same time as they secure sites and electricity supply.

"AI factory builders across EMEA are making investment decisions across GPU, networking and storage at the same time they're securing power and land, and those decisions are hard to reverse so choosing technology partners that offer both performance and power efficient design is crucial," said Mathieu.

"This reference design gives them a tested starting point: storage sized to the GPUs, from checkpoints through KV cache, with the network and facility requirements already worked out. Because we validate it in our AI Proving Ground first, customers can move from design to deployment with fewer surprises."

Omer Asad, Chief Product Officer at DDN, said the partnership combines DDN's storage with WWT's testing and integration work.

"WWT brings the integration and testing scale that large AI factories need. Together we give customers an NVIDIA-certified storage design for Vera Rubin, from training to inference," said Asad.