World Wide Technology (WWT) and DDN have published a storage reference design for AI factory campuses of 50 megawatts and above, announced at NVIDIA GTC Berlin. The pitch rests on a simple budget: every watt the storage draws is a watt the GPUs cannot use, yet the storage still has to keep them fed.
The design follows NVIDIA's reference architecture for the Vera Rubin NVL72, a 72-GPU rack, and targets more than 2.5GB/s of read throughput per GPU. A single 2U DDN AI400X3M appliance is rated at 190GB/s of reads, which covers the 180GB/s a full rack needs.
Inference is the second half of the design. Long-context, reasoning and agentic workloads can build more key-value (KV) cache, the working memory a model creates for a prompt, than GPU memory can hold. Cache thrown away without an offloaded copy has to be recomputed, which costs GPU time and delays the first token.
To avoid that, NVIDIA CMX adds a context memory tier built on BlueField-4 STX processors, sitting between GPU memory and shared storage. DDN Infinia runs on that infrastructure and shares the cache across inference nodes, so each node can reuse work another has already done. DDN reports loading KV cache up to 55 times faster: in one 112K-token test it took 2.1 seconds to load against 57 seconds to recompute. The companies note that results depend on workload and configuration.
Training data and checkpoints sit on DDN EXAScaler, KV cache on DDN Infinia, with NVIDIA Dynamo placing cache between tiers and a Spectrum-X Ethernet network of Spectrum-6 SN6600 switches and ConnectX-9 SuperNICs tying it together. WWT integrates the components and validates them in its AI Proving Ground.
Felicien Mathieu, CTO Middle East and Africa at WWT, said AI factory builders across EMEA are making GPU, networking and storage decisions at the same time as they secure power and land, and that those decisions are hard to reverse. "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."
Omer Asad, CPO at DDN, said: "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."
The two companies were showing the design at DDN's booth, number 3012, at GTC Berlin on 21 and 22 October 2026.
To stay across the latest in cloud, AI and enterprise tech analysis from Compare the Cloud, subscribe to our weekly newsletter at https://www.comparethecloud.net/newsletter
