Nvidia Cuts Next-Gen AI Chip Memory Capacity by Up to 81%

Nvidia, the world's largest artificial intelligence (AI) semiconductor company, is considering reducing the high-bandwidth memory (HBM) capacity of its next-generation AI chips by up to 81%. The move comes as memory costs soar and supply constraints worsen. The 'Rubin Ultra' AI chip, scheduled for release in 2027, was previously announced to feature sixteen 12-high HBM4E stacks, providing a total of 1TB of memory.

However, Nvidia is now testing configurations that lower the HBM configuration to as little as 192GB, which is 81% less than the original plan and roughly 33% less than the 288GB found in the currently mass-produced predecessor, 'Rubin.'

The primary driver behind this design review is the astronomical rise in memory costs. According to an analysis by Morgan Stanley, memory costs accounted for $373,939 of the total bill of materials (BOM) for Nvidia's current-generation Grace Blackwell server product, representing a 9.4% share.