Compared to 90 days ago, the storage market landscape has significantly strengthened, with the structural supply-demand imbalance continuing to worsen and the industry's business model transformation accelerating. On the demand side, hyperscale cloud vendors and AI capital expenditures are expanding rapidly (JPMorgan's model predicts that the data center CapEx of the top four US cloud vendors will grow by 85% year-on-year in 2026, and still exceed 70% year-on-year in 2027); this, coupled with the increased HBM usage per AI accelerator card and the explosive growth of agentic AI inference workloads (highly reliant on GPU-based HBM and CPU-based DRAM), is driving demand to continue far exceeding supply.
Meanwhile, due to the significantly higher requirements for DRAM capacity configurations placed on agentic AI workloads, server CPU demand has become a significant source of incremental growth. On the supply side, multiple structural bottlenecks continue to constrain bit shipment growth, including a slowdown in process node evolution, increased HBM wafer capacity loss (the wafer consumption ratio widens from approximately 3:1 to approximately 4:1 in the HBM4E stage), longer cleanroom expansion cycles, and the fading benefits of "tapping existing capacity" in 2025/2026. DRAM bit shipment growth is projected to be in the early 20% range in 2027-2028, while NAND bit shipment growth will be in the mid-20% range.
According to our global memory team's calculations, this will result in a persistent supply-demand gap of approximately 20%/19%/15% in 2026/2027/2028, respectively. By 2028, the cumulative HBM shortage may reach 23 weeks, and the total potential market size (TAM) of HBM in the industry is expected to jump from approximately $64 billion in 2026 to $160 billion in 2027 and $280 billion in 2028.
