Samsung and SK hynix are preparing to supply 8-layer HBM4 memory modules to Nvidia. Both companies aim to begin volume shipments for the Rubin Ultra NVL systems around the second quarter of 2027.
Samsung started mass shipments of 12-layer HBM4 memory for Nvidia's Vera Rubin AI accelerator in February 2026. These chips achieved speeds reaching 11.7 Gbps, exceeding the JEDEC industry standard of 8 Gbps.
The shift toward 8-layer stacks is driven by the need for better thermal management and higher manufacturing yields. Taller 12-layer configurations are more prone to defects and heat-related issues.
SK hynix is expected to provide approximately two-thirds of Nvidia's total HBM4 supply. Meanwhile, Samsung reports an 80% yield rate for its HBM4 production, which may help increase its market share to 28% or more.
By mid-2027, Samsung and SK hynix are projected to control over 80% of the HBM4 market. Nvidia utilizes both suppliers to mitigate manufacturing risks and ensure a stable supply chain for its AI hardware.
