3 6 Ke·2025-07-24 10:12Core Insights - Goldman Sachs analysts predict that HBM prices may drop by 10% by 2026 due to oversupply in the market [1][3] - The pricing power in the HBM market is shifting from manufacturers to customers, particularly represented by Nvidia, indicating a loss of SK Hynix's previous dominance in securing 80%-90% of Nvidia's orders [2][3] Supply and Demand Dynamics - The primary reason for the anticipated price drop is "oversupply," as major players like Samsung, SK Hynix, and Micron accelerate production, leading to a significant increase in HBM supply that will exceed demand [3][4] - Goldman Sachs has revised its total addressable market (TAM) for HBM in 2026 down by 13%, from $51 billion to $45 billion, and reduced growth expectations from 45% to 25% [3] - By the end of 2025, SK Hynix plans to increase its monthly production capacity from 65,000 wafers to 150,000 wafers, while Samsung's capacity is expected to reach 150,000 wafers, and Micron aims for 90,000 wafers [5] Market Competition and Product Trends - SK Hynix currently holds about 50% of the HBM market share, significantly higher than Samsung's 30% and Micron's 20% [8] - The demand for HBM products is expected to be driven by the ASIC market, particularly from companies like Amazon, Meta, and Google, which are rapidly increasing their demand for custom semiconductors [6][7] - The introduction of HBM4 in 2026 is anticipated to be a significant focus, with SK Hynix and Samsung both ramping up their production efforts [15][16] Future Outlook - UBS suggests that despite short-term noise in the memory market, SK Hynix is likely to maintain its leadership position in the HBM sector, especially with the growing demand from AI platforms [14] - The competitive landscape for HBM4 will be crucial, as Samsung aims to regain market share after setbacks in previous product launches [15][17] - The technological roadmap for HBM from 2025 to 2040 indicates a shift towards higher performance and efficiency, with HBM4 expected to play a pivotal role in the AI chip storage market [19][20]