Column: Compute is no longer the AI bottleneck. Memory is — and suppliers know it

The AI hardware race has shifted from compute to the memory bottleneck. The 2026 semiconductor index surge—and subsequent correction—reveals deep skepticism about AI-driven demand sustainability. Technically, HBM and CXL are reshaping server architectures, forcing co-evolution in EDA, advanced packaging, and test equipment. Regulatory fragmentation—U.S., Dutch, and Japanese export controls combined with China’s localization push—has bifurcated the memory supply chain, raising operational costs by 15–20%. Samsung and SK Hynix are betting on HBM4, Micron leverages CHIPS Act subsidies for capacity expansion, while TSMC integrates ASICs with HBM via CoWoS to lock in performance leadership. Over the next 12–24 months, memory bandwidth will dictate AI cluster efficiency; vendors lacking 3D stacking or silicon photonics capabilities face marginalization. Geopolitical decoupling will accelerate regional 'trusted supply chains,' especially creating a technology firewall between Taiwan, China and mainland China.