Massive Shortage Hits Advanced AI Chips

- The primary cause of the shortage is a deliberate production shift by major memory manufacturers like Samsung, SK Hynix, and Micron, who are reallocating manufacturing capacity away from conventional DRAM used in PCs and smartphones to more profitable HBM for AI data centers. - Producing HBM is far more resource-intensive; each gigabyte of HBM consumes approximately three times the silicon wafer capacity of a standard DDR5 gigabyte, severely impacting the supply of memory for consumer electronics. - A critical bottleneck exists in advanced packaging, where demand for TSMC's "CoWoS" (Chip-on-Wafer-on-Substrate) technology, essential for integrating HBM with GPUs, significantly outstrips supply from major clients like NVIDIA, Google, and Amazon. - Despite TSMC planning to double its CoWoS capacity annually and targeting a monthly output of over 140,000 wafers by the end of 2026, demand continues to exceed this rapid expansion. - The insatiable demand for HBM has altered the market landscape, helping SK Hynix to overtake Samsung as the top DRAM producer for the first time in Q1 2025, with SK Hynix now controlling over half the HBM market. - The supply is so constrained that lead times for key components like Nvidia's H100 GPUs have stretched to six months, and SK Hynix has reportedly sold its entire HBM chip output for 2026. - The ripple effect extends to corporate strategy, with memory maker Micron announcing it will exit its consumer-facing brand, Crucial, to free up wafer supply for its strategic data center clients. - The shortage is forecast to persist through 2026 and possibly into 2027, threatening to delay AI infrastructure projects and increase prices for consumer devices.