Memory: Datacenter HBM unlikely to move to 4-Hi
🔑 We share TrendForce’s view that the chance of de-spec to 4-Hi is low, vs worries over 4-Hi which could deliver same bandwidth with fewer dies, cut inference costs & stretch limited DRAM supply
• Two key reasons 4-Hi unlikely fly in datacenter AI:
– Per-stack capacity is too small (~16GB for HBM5/HBM4E) → hyperscalers would need more chips, racks & scale-up networking
– Higher volume still increases base-die demand on tight advanced logic nodes (esp. N2 for HBM5) → little to no supply relief
• Supply-chain feasibility low: mainstream production is focused on 8-Hi & 12-Hi, vs development cost and low-volume 4-Hi flows
• Cost per Gb is higher on 4-Hi vs 8-Hi, with similar total cube cost based on current price quote.
• Nvidia’s Rubin Ultra dropping to 8-Hi is well expected (our note on Aug 6)
• For Rubin Ultra we expect parallel development of HBM4 8-Hi 192GB / HBM4E 8-Hi 256GB / HBM4E 12-Hi 384GB, with higher chance of HBM4 8-Hi first, then HBM4E 8-Hi as the mainstream SKU (consistent with our earlier AI Thermal report of maintain 2600W)
• Reiterate our positive view on the DRAM/HBM sector
#
HBM# #
DRAM# #
MU#