每堆栈和每 GB 的价格
HBM3 24GB 12 层堆栈: $199 per stack, $8.31/GB (报告值). HBM3E 36GB 12 层堆栈: $2,108 per stack, $58.56/GB (报告值). HBM4 48GB 16 层堆栈: $3,521 per stack, $73.35/GB (报告值).
HBM3 is a contract reference level. The HBM3E and HBM4 levels are scarce-lot spot indications reported by TrendForce in September 2026, citing Korean press. TrendForce put them at roughly four to five times long-term-agreement pricing. They are real reported prints, but they show what a buyer without allocation pays, not what Nvidia or a hyperscaler pays under contract.
几代人之间发生了什么变化
HBM3 (24GB, 12-Hi here) set the baseline for the first wave of large AI accelerators. HBM3E raises per-pin speed and capacity (36GB at 12-Hi) and is the volume part in current GPUs. HBM4 doubles the interface width to 2,048 bits, moves the base die to a logic process, and reaches 48GB at 16-Hi.
每个步骤都会增加堆叠层数和封装复杂性，从而降低产量。这不仅仅是原始芯片成本，也是每 GB 价格从一代到下一代不断上涨的原因。
几代人如何互动
Suppliers shift wafers and packaging lines toward the newest generation, so older HBM becomes scarcer before demand for it fades. The first sign of a softer HBM market usually shows up in HBM3: a sustained negative 30-day move there, before HBM3E or HBM4 react.
HBM4 还没有有意义的同比历史记录。它的图表线从当前报告的水平开始并向后重建，因此请谨慎阅读其长窗口形状。