Samsung Electronics and SK Hynix plan to begin mass production of 8 layer HBM4 memory stacks in the second quarter of 2027 to supply Nvidia. The move comes after Nvidia requested 8Hi configurations for its upcoming Rubin architectures to stabilize production yields and reduce costs compared to taller 12 layer modules. Initial volume shipments are scheduled for mid 2027, supporting adjusted memory targets for next generation enterprise computing platforms.
Industry reporting from ZDNet Korea reveals that both memory makers initially focused production lines on 12 layer HBM4 designs. Building 12 layer stacks requires bonding more DRAM dies together, which increases cost per stack and reduces overall factory yields due to the physical height of the silicon. Discussions between Nvidia and both manufacturers in late 2026 led to a strategic shift toward 8 layer HBM4E memory to optimize supply chain output. Producing 8Hi modules allows foundries to deliver higher volumes of functional silicon while avoiding the defect rates common to 12 layer manufacturing.
Nvidia plans to integrate these 8 layer stacks across both the standard Rubin and Rubin Ultra platforms. The company adjusted specifications for the HGX Vera Rubin NVL8 and HGX Rubin NVL8 systems to feature a combined memory pool of 2TB HBM4. This total capacity relies on 8Hi 32GB memory stacks distributed across the accelerator platform. Prioritizing 8Hi hardware gives Nvidia a steadier supply of high bandwidth memory as enterprise demand for artificial intelligence infrastructure scales into 2027.
