Phison CEO Warns of Severe Memory Supply Crisis

- The memory supply crunch is a structural shift, not a temporary shortage, as major manufacturers like Samsung, SK Hynix, and Micron redirect production from consumer-grade DRAM and NAND to more profitable, high-margin HBM for AI accelerators. This reallocation is creating a zero-sum game where every wafer used for an HBM stack for an AI server is one less for a consumer device. - Price increases have been dramatic, with some DRAM types soaring 75% from December 2024 to January 2025, and NAND flash prices are also rising sharply. Some specific components, like 8GB eMMC modules used in consumer electronics, have seen a price increase of over 13 times in the past year. - The impact extends beyond just price, with lead times for memory components stretching to 32-40 weeks or more. Memory manufacturers are reportedly demanding up to three years of prepayment for orders, a move unprecedented in the electronics industry. - Hyperscalers are aggressively developing and deploying their own custom silicon (ASICs) to optimize for specific AI workloads and reduce reliance on general-purpose GPUs. Shipments of these custom AI server ASICs are projected to triple between 2024 and 2027, with players like Google, AWS, Meta, and Microsoft ramping up their internal silicon. - This "build vs. buy" decision for hyperscalers is driven by the need for greater performance-per-watt, lower total cost of ownership (TCO), and more control over their hardware and software stacks. However, due to long internal build cycles of 4-5 years, they often lease capacity from third-party providers to meet immediate, uncertain demand fueled by the AI race. - The shortage is forcing a strategic response from companies outside the AI sector, including redesigning products to use less memory, securing long-term supply agreements, and in some cases, exiting product lines altogether. Apple and Samsung are better hedged due to their scale and long-term agreements, while smaller manufacturers with thin margins are most at risk. - While the immediate crisis is severe, new memory fabrication plants are under construction, though significant new capacity is not expected to come online until 2027 or later, suggesting the tight supply conditions will persist. Samsung has begun mass production of HBM4, which offers significant performance and efficiency gains, but the overall supply will remain constrained. - For AI startups and ML teams, the rising cost of memory directly impacts training and inference economics, making cost optimization even more critical. The high cost and scarcity of HBM could slow down the expansion of data centers and impact the margins of cloud service providers.