Silicon Motion's CEO has issued a stark warning that the AI-driven memory semiconductor shortage will likely persist until 2028.
This isn't just one executive's cautious outlook; it's a conclusion supported by a cascade of evidence across the entire semiconductor supply chain. The situation is a perfect storm where surging AI demand meets deeply entrenched supply constraints, creating a bottleneck that affects everything from high-end HBM to conventional DRAM and NAND flash memory.
Let's break down the causal chain. First, the world's leading memory suppliers are running at full tilt. Recent reports indicate that a major player like SK hynix has essentially zero spare capacity. They are receiving unprecedented offers from clients willing to co-fund new production lines—a highly unusual move in this industry. This is corroborated by Samsung, which also projected significant shortages lasting until at least 2027.
Second, the problem isn't just about making more silicon wafers. A critical bottleneck exists in advanced packaging, like TSMC's CoWoS technology. This process, which is essential for assembling high-performance AI chips, is itself capacity-constrained. Even if memory makers could produce more chips, the means to package them into finished products are limited, delaying any potential relief.
Finally, the market is already adapting to this new reality by locking in supply for years to come. We are seeing a proliferation of Long-Term Agreements (LTAs), where cloud giants and other large customers secure memory and storage capacity through multi-year contracts extending to 2028 and even 2029. This trend isn't limited to advanced memory; it has spilled over to NAND flash and even traditional hard disk drives (HDDs), institutionalizing the shortage across the board.
Therefore, the warning of a shortage lasting until 2028 isn't speculation. It's the logical endpoint of a path defined by zero spare capacity, packaging bottlenecks, and a market solidifying scarcity through long-term contracts. Barring a sudden drop in AI demand or a breakthrough in manufacturing capacity, this extended period of tightness seems to be our new base case.
- HBM (High Bandwidth Memory): A type of high-performance memory stacked vertically, essential for powerful AI processors that need to access large amounts of data very quickly.
- NAND Flash: The most common type of non-volatile storage memory, used in SSDs, smartphones, and USB drives. It retains data even when power is turned off.
- LTA (Long-Term Agreement): A multi-year contract between a supplier and a customer to secure the supply of goods at predetermined terms, which helps stabilize supply chains.