Overview
- Samsung is reported to begin large‑scale HBM4 production in February 2026, with most output earmarked for NVIDIA’s next‑gen Vera Rubin system and some for Google’s seventh‑generation TPU.
- Industry reports say Samsung’s HBM4 scored highest in NVIDIA’s SiP testing on speed and power efficiency, bolstering expectations it will pass final qualification.
- SK Hynix plans to deliver 12‑high HBM4 final samples to NVIDIA in early January 2026 and, pending quick quality checks, aims to ramp production between February and March.
- SK Hynix, NVIDIA, and TSMC have adopted a three‑way collaboration for SiP integration and reliability fixes, with data sharing and TSMC providing the substrate die.
- HBM3E contract prices have risen by roughly 20% on tight supply as hyperscalers seek long‑term deals; one media report claims Google dismissed a procurement executive over HBM shortages, while ASUS formally denied any plan to build a DRAM wafer fab and China’s state security authority issued AI safety guidance.