NVIDIA Trims Vera Rubin Memory as HBM4 Prices Threaten to Eat 29% of Every Rack’s Cost – Report

NVIDIA is reportedly trimming memory in its Vera Rubin NVL72 AI racks to curb rising system costs.

GF Securities says SOCAMM memory per rack will be cut from 192GB to 96GB, and CPU-rack memory will also be reduced.

HBM4 allocations for GPUs will stay unchanged, preserving core accelerator performance.

The changes are meant to offset surging memory prices, which could otherwise account for nearly 29% of total system cost.