Nvidia is evaluating significant memory spec reductions for its upcoming Rubin Ultra GPU due to HBM supply shortages from SK Hynix, Samsung, and Micron. The chip was originally planned with up to 1 TB of HBM4E memory but is now being tested with configurations as low as 192 GB using 8-Hi stacks, representing a 33% reduction from the current Vera Rubin’s 288 GB. TrendForce reported in early August 2026 that Nvidia is testing four distinct HBM configurations as DRAM and HBM shortages are expected to deepen into 2027. This reduction could force enterprises running large AI models to purchase more GPUs to handle the same workloads, increasing power consumption, rack space, and infrastructure costs.
Source: Read the original article
