Nvidia is reportedly considering eight-layer HBM4 memory stacks as Samsung and SK hynix grapple with heat and manufacturing yields on taller versions.
According to ZDNet Korea, Nvidia entered discussions with both suppliers during the third quarter of 2026 about supplying eight-high stacks alongside existing 12-high products.
Samsung and SK hynix are reportedly preparing to increase eight-layer production volumes in the second quarter of 2027. Neither manufacturer is expected to abandon 12-layer HBM4 or HBM4E.
The discussions suggest Nvidia wants more flexibility in its memory supply, although it is still reviewing a mixed configuration. The report describes a possible option, but no final decision has been announced.
Taller stacks offer more memory capacity within the same package footprint, but manufacturing and cooling become more demanding as more DRAM dies are stacked.
Each extra layer complicates the thermal path and adds another bonding interface where a defect can spoil the finished stack. Eight-high products sacrifice some capacity per stack in exchange for potentially easier production and cooling.
The report points to thermal and yield concerns affecting 12-high HBM4 and HBM4E, without providing a detailed engineering explanation or figures quantifying the difficulties.
Samsung and SK hynix are reportedly already mass-producing 12-layer HBM4 that has passed Nvidia’s qualification process. The shorter version requires separate approval.
Both suppliers are said to be arranging components and materials while preparing quality testing for the eight-high products.
Twelve-layer HBM4 is expected to remain the main product this year, with the shorter stacks taking a larger share from mid-2027.
Some industry observers had expected eight-high supplies during the second half of 2026. That timetable has reportedly slipped, although an unnamed source expects qualification to be easier than for the taller stacks.
Earlier reporting suggested Rubin would retain 12-high HBM4. The latest account leaves room for continued use of that memory alongside a more readily manufactured alternative.
Rubin Ultra has faced separate reports of supply chain constraints, including a possible reduction from a four-die GPU design to two dies. Those claims remain unconfirmed.
Previous estimates put the cost of a Rubin Ultra rack at approximately $21 million, including about $1.5 million for HBM4E memory.
Using shorter stacks could change the memory bill, but the report provides no revised pricing or confirmed capacity specifications for an eight-high configuration.
