深度图解｜AI 巨头们正遭遇一堵怎样的「内存墙」？

The first lock: HBM is extremely capacity-intensive. According to Micron's estimates, the wafer capacity required to produce one HBM chip is roughly equivalent to that of three standard DDR5 chips. You might think "HBM's share of the memory market" isn't that large, but if you calculate based on the actual wafer starts it consumes: from 2025 to 2027, while the bit share rises gently from about 8% to 13%, the share of wafer starts surges from 18% to 30%. Standard memory is thus being physically squeezed out of the factories.

The second lock: Expanding capacity is difficult and slow, and new capacity will be prioritized for HBM first. An EUV lithography machine costs $200 million, a fab costs tens of billions, and it takes several years from groundbreaking to production. Micron's Idaho fab won't start production until the second half of 2027, with Samsung's and SK Hynix's new production lines scheduled for their first shipments in 2027 and 2028, respectively. Furthermore, the management of all three companies has made it clear: until 2027, new capacity will be prioritized for HBM and high-margin enterprise memory.

The third lock: The three major manufacturers have 'learned their lesson' this time and are deliberately avoiding a massive capacity expansion. This is the most significant difference from past cycles. Previous market crashes were self-inflicted by the manufacturers—they would frantically expand capacity as soon as prices rose, only to end up with a glut of unsold inventory. This time, Samsung and SK Hynix (who together control about two-thirds of the global memory market) have explicitly told investors: they will no longer expand standard memory capacity recklessly, instead reserving capacity for the more profitable HBM, prioritizing profitability over shipment volume.