SK hynix and Samsung are pushing 16-layer HBM4 while Rubin favors 12. Memory makers are running ahead of the accelerator roadmap — packaging ambition versus what the compute side actually specifies. I

What happened: Counterpoint Research said most Nvidia Vera Rubin products will likely use 12-layer HBM4, not the eight-layer variant it cited in August. SK hynix and Samsung continue work on 16-layer HBM4. Key numbers: SK hynix 16-layer HBM4 offers 48GB, one-third more than its 36GB 12-layer version. Samsung 12-layer HBM4 ships up to 36GB. Why it matters: The 12-layer outlook eases demand concerns for the Korean makers, but 16-layer adoption still faces yield and advanced packaging hurdles. Packaging capacity at TSMC could constrain AI accelerator supply.

Counterpoint Research said most Nvidia Vera Rubin products will likely use 12-layer HBM4, not the eight-layer variant it cited in August. SK hynix and Samsung continue work on 16-layer HBM4.

The 12-layer outlook eases demand concerns for the Korean makers, but 16-layer adoption still faces yield and advanced packaging hurdles. Packaging capacity at TSMC could constrain AI accelerator supply.