GigaDevice director maps memory's next three years: capacity, AI demand, and the race for new applications

The AI compute arms race is transforming memory from a cyclical commodity into structurally constrained supply. For firms like GigaDevice, merely scaling capacity is insufficient: the convergence of HBM and CXL interfaces forces parallel upgrades in controller IP, advanced packaging, and test equipment—laggards risk exclusion from high-end supply chains. Tightening export controls on ALD tools by the U.S. and Netherlands have already raised compliance costs for Taiwan, China and mainland China fabs by over 15%. With Samsung accelerating 2nm DRAM trials and Micron betting on LPDDR6-AI custom designs, domestic players must move beyond generic NOR Flash toward AIoT-optimized memory architectures. Over the next 18 months, edge AI inference chips will drive demand for 'small-capacity, high-bandwidth, ultra-low-power' memory; early movers in compute-in-memory co-design will capture the first wave of long-tail returns.