#Hbm4

SK hynix has surpassed 70% yield for HBM4 memory, outperforming its own forecasts. Competition with Samsung (approx. 80%) and Micron is heating up, shifting the focus to scaling speed and cost efficiency.

Samsung has approved the architecture for its Fab 2 plant in Taylor, Texas: 250k m² dedicated to 2nm HBM4 base logic for AMD and Nvidia. Construction starts in November 2026, with mass production in 2030.

Nvidia has lowered HBM4 and NVLink 6 bandwidth in the Vera Rubin platform by 12–17%. Adjustments affected NVL72 and HGX Rubin NVL8 systems, while some pages still display old values.

Most NVIDIA Rubin accelerators will retain 12-layer HBM4 memory for peak AI performance, despite rising costs and component shortages.

Samsung is discussing pricing for next-generation HBM4 memory at $4-5 per gigabyte, which is three times higher than the cost of HBM3E.