Nvidia Weighs Lower HBM Specs for Rubin Ultra Amid DRAM Squeeze

High-bandwidth memory is a critical component in AI accelerators, providing the rapid data access needed for increasingly complex workloads. Its capacity and stack configuration have a direct bearing on chip performance, power consumption and cost. For Nvidia’s next-generation Rubin Ultra platform, tight DRAM supply and uncertainty surrounding HBM4e validation and manufacturing have emerged as potential constraints on production scale and system economics.

Nvidia is evaluating lower HBM specifications, capacity and stack counts for Rubin Ultra as DRAM supplies are expected to remain tight through 2027 and the timing of HBM4e mass production remains uncertain. The changes could ease procurement costs and supply bottlenecks. Some cloud service providers are also considering reducing HBM capacity in their internally designed ASICs. No final configuration, procurement value or implementation date has been disclosed.