Micron Technology(US:MU) 财联社·2025-10-03 06:43Core Viewpoint - The article emphasizes that in the AI era, storage chips, particularly HBM (High Bandwidth Memory), have transitioned from a supporting role to a critical bottleneck and breakthrough point due to the explosive growth in model parameters and training data, making traditional memory technologies inadequate [1] Group 1: HBM Demand and Supply - Major tech companies are positioning HBM as a strategic asset, with Micron highlighting an expected worsening supply-demand imbalance for HBM and other semiconductor chips [2] - Micron's CEO noted that DRAM inventory is below target levels while NAND inventory continues to decline, with HBM demand significantly increasing and expected to outpace overall DRAM growth by 2026 [2] - Huawei announced that its Ascend AI chips will utilize self-developed HBM starting with the Ascend 950PR, indicating a shift towards proprietary HBM solutions [2] Group 2: Custom HBM and Cost Implications - Analysts point out that customized HBM (cHBM) has evolved from a passive component to an active one with logic capabilities, reshaping the role of storage in AI infrastructure and increasing total cost of ownership (TCO) [3] - The integration of different functionalities and logic designs in cHBM is seen as a key differentiator in performance [3] Group 3: Energy Efficiency and Full-Stack Solutions - Storage manufacturers are now offering full-stack solutions that include HBM, logic die, LPDDR, and PIM, focusing on customized HBM collaborations with clients [4] - SK Hynix predicts that a 10% improvement in HBM energy efficiency could lead to a 2% energy saving at the rack level, highlighting the importance of HBM in energy conservation [4] Group 4: AI Inference Growth - The rapid rise of AI inference is driving a surge in storage demand, with expectations of a billion-fold increase in inference capabilities [5] - The need for high-performance memory and tiered storage is becoming increasingly critical as AI inference applications grow, with HBM, DRAM, SSD, and HDD playing vital roles in handling extensive data [5][6] Group 5: Potential for "Storage as Computation" - The supply constraints of HBM may lead to the emergence of "storage as computation" technologies, which aim to shift vector data from expensive DRAM and HBM to more cost-effective SSDs [7] - This approach is expected to significantly reduce latency and improve throughput in AI inference, providing a feasible path for large-scale AI deployment [7]