NVIDIA is reducing the default SOCAMM configuration in its VR200 NVL72 racks from 192GB to 96GB modules, effectively halving the CPU-side LPDDR5X memory per rack from ~54-55TB to ~28TB. This adjustment aims to address LPDDR5X supply constraints and optimize costs, preventing memory and storage expenses from escalating to $2.1 million (29% of total BOM). The specification cuts prioritize faster system deployment and higher rack availability, with general server DDR5 specs also expected to be reduced by roughly 50% per CPU.
NVIDIA is actively de-contenting AI server memory to manage costs and accelerate deployments, indicating LPDDR5X supply is a binding constraint.