Nvidia GTC Berlin 2026: WWT and DDN Design Storage to Help AI Factories Fit More GPUs Within Their Power Limits

Every AI factory has a fixed power budget. Power the storage uses is power the GPUs can’t use. But storage still has to feed the GPUs data fast enough to keep them busy. A GPU waiting for data is wasted power.

World Wide Technology (WWT) and DDN announced a storage reference design for AI factory campuses of 50 megawatts and above, built to meet both limits at once. The design follows Nvidia’s reference architecture for the Nvidia 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 read throughput, which covers the 180GB /s a full rack needs.It’s designed to take less space and power, so operators can fit more GPUs per megawatt, with less idle time, which delivers faster time to first token and more token throughput per second.

For inference, long-context, reasoning and agentic workloads can create more key-value KV cache, the working memory a model builds for a prompt, than GPU memory can hold. Cache discarded without an offloaded copy must be recomputed when needed again, which costs GPU time and delays the first token.

Nvidia CMX adds a context memory tier built on Nvidia BlueField-4 STX, a layer between GPU memory and shared storage that holds KV cache. DDN Infinia runs on the STX infrastructure and shares the cache across inference nodes, so each node can reuse work already done instead of recomputing it. That cuts repeat GPU work and shortens time to first token.

DDN reports loading KV cache up to 55 times faster. In one 112K-token test, it took 2.1s to load vs. 57s to recompute. For the user, that would mean a two-second wait instead of nearly a minute. For the operator, it would provide GPU time back for new requests. Results depend on workload and configuration.

The design covers training and inference: data and checkpoints on DDN EXAScaler, KV cache on DDN Infinia, Nvidia Dynamo for kv-cache placement between tiers, and an Nvidia Spectrum-X Ethernet network of Nvidia Spectrum-6 SN6600 switches and ConnectX-9 SuperNICs with Nvidia BlueField-4 STX storage processors powering the context-memory tier. WWT integrates the components from each vendor and designs them to work as one system, so customers deal with one design instead of several vendors’ specifications.

“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 Felicien Mathieu, CTO, Middle East and Africa, WWT. “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.”

“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 Omer Asad, CPO, DDN.