Beyond the Memory Cycle: AI, HBM, and the New Semiconductor Shortage

Center for Strategic and International Studies | Sujai Shivakumar, Charles Wessner, and Thomas Howell

Global memory manufacturers Samsung, SK Hynix, and Micron are reallocating fabrication capacity toward High-Bandwidth Memory (HBM) to meet surging artificial intelligence data center demand, triggering a structural global semiconductor shortage projected to persist into 2028. This strategic pivot starves conventional dynamic random-access memory (DRAM) and NAND flash markets, elevating critical supply chain security risks for consumer electronics, automotive manufacturing, and military platforms.

High fixed operational costs and multi-year fabrication facility lead times prevent rapid production adjustments, leaving global markets vulnerable to sudden structural demand shocks. Capital concentration reinforces this extreme instability. With the Big Three controlling 90.5 percent of DRAM sales in late 2025, advanced stacking and packaging requirements keep commercial HBM manufacturing heavily concentrated in East Asia. Consequently, tighter conventional supplies create expanding market opportunities for Chinese producers like ChangXin Memory Technologies and Yangtze Memory Technologies, deepening Western reliance on foreign suppliers for foundational digital hardware.

Comment

Concentrated advanced packaging for Nvidia H100 processing units creates a single-point industrial vulnerability in global defense microelectronics. These AI accelerators require high-bandwidth memory dies stacked up to 16 layers deep via copper through-silicon vias before system integration. Without domestic packaging ecosystems, expanded United States silicon wafer fabrication under the CHIPS Act cannot yield operational AI processing hardware.

A similar structural realignment occurred as ChangXin Memory Technologies expanded its share of global dynamic random-access memory production to 8 percent. That expansion positioned Hefei-based foundries as dominant suppliers for conventional automotive and industrial microcontrollers when SK Hynix and Micron prioritised high-bandwidth memory. Consequently, this manufacturing shift leaves Western defense subcontractors reliant on ChangXin Memory Technologies for legacy DRAM chips used in military logistics systems.

Strategic Question for Discussion

Which factor poses a greater long-term supply vulnerability for Western military hardware — the concentration of Nvidia H100 advanced packaging in East Asia or growing defense contractor reliance on ChangXin Memory Technologies for legacy DRAM?

The available evidence points toward reliance on ChangXin Memory Technologies as the more immediate operational vulnerability, given that legacy DRAM components are already embedded throughout defense supply chains with limited non-Chinese alternative sources. While advanced packaging bottlenecks for Nvidia H100 accelerators restrict future AI capacity building, those bottlenecks remain concentrated among allied firms susceptible to coordinated policy interventions. In contrast, replacing widespread Chinese commodity memory in legacy defense logistics platforms requires costly, multi-year supply chain audits and requalification processes.

Share your assessment in the comments below.

To read in details click here