Physix FrontierCore Judgment: AI-Driven Memory Demand Enters Structural Shortage Phase; SK Hynix Predicts Doubling Capacity Won't Meet Future Five-Year Needs
I. From "Cyclical" to "Structural": The Paradigm Shift in the Memory Industry
Chey Tae-won gave an interview on the day SK Hynix's ADR listed on Nasdaq, emphasizing that "the memory industry has entered a phase of structural growth." This was no coincidence. Choosing to release this signal at the moment of US stock listing serves both as expectation management for Wall Street and as endorsement of their own technology roadmap.
Traditional memory industries (DRAM, NAND) are characterized by strong cyclicality: prices crash during oversupply and soar during undersupply, cycling every 3-4 years. But the demand logic brought by AI is completely different. Take HBM (High Bandwidth Memory) as an example; it is not general-purpose storage but a core component for "near-memory computing" specifically customized for AI accelerators (like NVIDIA's H100, B200). One H100 GPU requires about 144GB of HBM3, and a large language model training cluster often needs thousands of GPUs. This demand is exponential, not linear growth.
When Chey says "even doubling capacity in the next five years may not meet demand," there are two layers of meaning behind it: first, current capacity utilization is already near limits; second, the ramp-up cycle for new capacity (taking 2-3 years from factory construction to stable yield) simply cannot match the Moore's Law-like doubling of AI compute demand every two years. Benchmarking overseas cases, NVIDIA CEO Jensen Huang stated bluntly at the 2024 GTC conference that "HBM supply is extremely tight," and Samsung Electronics' catch-up speed on HBM3E also confirms the depth of technical barriers.
II. Reshaping Competitive Landscape Under Porter's Five Forces Framework
1. Supplier Bargaining Power: Constraints Across Equipment and Materials
Memory manufacturing is highly dependent on ASML's EUV lithography machines, Tokyo Electron's etching equipment, and Shin-Etsu Chemical's silicon wafers. As SK Hynix, Samsung, and Micron all expand production, equipment capacity becomes the bottleneck. The "capacity doubling" mentioned by Chey is actually constrained by equipment delivery cycles—ASML's 2024 EUV capacity is only about 60 units, and the three major manufacturers competing for HBM-specific equipment have extended lead times to over 18 months. This means that even with funding, capacity cannot be released quickly.
2. Buyer Bargaining Power: The "Dual Identity" of Cloud Providers
Buyers of AI chips (Microsoft, Google, Amazon) are also indirect purchasers of memory (via OEMs or self-developed chips). However, HBM customization is extremely high; NVIDIA's HBM3E specs are almost exclusively supplied by SK Hynix, leaving buyers with little bargaining power. This "seller's market" is historically rare in the memory industry, last seen during the Japanese semiconductor golden age in the 1980s.
3. Potential Entrants: High-Barrier "Capital Game"
Memory manufacturing is a typical "economies of scale + technology-intensive" industry. Building a new 12-inch wafer fab costs $10 billion, and yield ramp-up takes years. Chinese manufacturers like CXMT and YMTC have made breakthroughs in NAND/DRAM, but the gap in TSV (Through-Silicon Via) processes and CoWoS packaging required for HBM is huge. Chey admitted that "individual enterprises cannot meet demand," implying SK Hynix might attract entrants through tech licensing or joint ventures, but the oligopoly structure won't change in the short term.
4. Threat of Substitutes: "Far Water" from New Storage Types
New non-volatile storage technologies like MRAM (Magnetoresistive RAM) and PCM (Phase Change Memory) can theoretically replace DRAM, but speed, power consumption, and lifespan still fail to meet AI training demands. While CXL (Compute Express Link) technology can expand memory pools, it essentially optimizes existing architectures rather than replacing them. Therefore, in the next 5-10 years, HBM and DDR5 remain the core bottlenecks for AI compute.
5. Existing Competitors: Samsung's Catch-Up and Micron's Gamble
Samsung Electronics trails SK Hynix by about a year in HBM3E but is accelerating catch-up thanks to its massive DRAM capacity and financial strength. Micron has chosen an aggressive capacity expansion strategy, planning to increase its HBM share to over 20% by 2025. Chey's statement...
Original Link: https://www.ithome.com/0/975/364.htm