Nvidia CEO urges SK hynix to make more HBM chips - upi.com

At the 2026 Computex in Taipei, NVIDIA CEO Jensen Huang visited SK Hynix’s booth and met for a second day with SK Group Chairman Chey Tae-won, deepening their AI partnership. Huang wrote 'Please Make ... Read original →

Jensen Huang’s gesture in Taipei, China isn’t theater—it signals that AI’s bottleneck has decisively shifted from compute logic to memory bandwidth. HBM4E and SOCAMM are now existential for NVIDIA’s Blackwell Ultra and Rubin architectures, forcing SK hynix into aggressive capacity doubling. This triggers a cascade: TSMC’s CoWoS packaging, TSV yield curves, and hybrid LPDDR5X/HBM memory stacks must all be re-engineered. Geopolitical friction intensifies—U.S. subsidies for SK’s Arizona advanced packaging facility come with ‘friend-shoring’ strings, inflating capex by over 30%. Samsung will likely counter by fast-tracking HBM4 samples tied to AMD’s MI400 GPUs to claw back AI training share. Within 18 months, HBM transitions from optional to system-defining; any GPU vendor without secured HBM4 volume for 2027 effectively forfeits the high-end AI race.