Another HBM killer is exposed. - 36氪

High Bandwidth Memory (HBM) has become one of the most prominent keywords in the semiconductor industry, driven by surging demand in AI large models, high-performance computing, and data centers. With... Read original →

HBM’s dominance is facing structural disruption. ZAM and 3D DRAM aren’t mere substitutes—they redefine memory-compute coupling, forcing rapid evolution in TSV and hybrid bonding. Upstream, EUV and 3nm foundry capacity may shift to accommodate novel stacking demands; downstream, AI accelerators must be redesigned for ZAM’s ultra-high bandwidth and low latency. For Intel, this move is both a technical pivot and geopolitical play—bypassing Samsung/Micron’s HBM hegemony to build a U.S.-aligned supply chain backed by SoftBank. Yet the current 10GB capacity caps ZAM to inference workloads, leaving training firmly in HBM’s grip. TSMC will likely accelerate CoWoS-R and SoIC integration to retain NVIDIA. Over the next 24 months, if ZAM yields stay below 60%, it remains niche; if not, it could trigger a global DRAM realignment, directly threatening Korea’s pricing power.