This one is super hardcore and worth a careful read. Here's a personal summary of the key points: 1) In the AI inference era, memory (especially HBM) has become the decisive bottleneck and is transitioning from a traditional cyclical commodity to a structural growth cycle. 2) Token throughput = HBM size × HBM bandwidth, meaning HBM size and bandwidth must double with each generation. 3) HBM meets two and a half out of three conditions: 'partial customization + structural exponential demand + rapid tech iteration,' significantly weakening its cyclicality. 4) DRAM, driven by agentic AI (long context, stateful processes, sandbox isolation), is seeing exponential TAM growth for CPU servers + a sharp increase in DRAM usage per core, creating new structural demand. While not as transformative as HBM, it now exhibits growth cycle characteristics. 5) China's ChangXin expansion has limited impact on overall industry bit growth (CAGR only increases by about 1.5%), making it unlikely to repeat the traditional cyclical pattern. 6) NAND/SSD also benefits from structural growth drivers like KV cache offload, AI video, and agent sandbox, resonating with HBM/DRAM to form demand synergy within the same memory hierarchy.