WWT and DDN Design Storage to Help AI Factories Fit More GPUs Within Their Power Limits
WWT, Wednesday, October 21st, 2026
WWT and DDN publish a storage reference design for 50MW+ AI factories that saves power for GPUs and speeds KV cache loading.
World Wide Technology and DDN announced a storage reference design for AI factory campuses of 50 megawatts and above, aimed at feeding GPUs fast enough while leaving more of the fixed power budget for compute.
Built on NVIDIA's Vera Rubin NVL72 reference architecture, it targets more than 2.5 GB/s of read throughput per GPU, with one 2U DDN AI400X3M appliance rated at 190 GB/s for a full rack.
For inference, DDN Infinia on NVIDIA BlueField-4 STX shares KV cache across nodes; DDN reports loading a 112K-token cache in 2.1 seconds versus 57 seconds to recompute.
WWT integrates the components as one system.