Weekly news roundup: AI compute race intensifies as DRAM rivals diverge and chipmakers expand investments

The 4F-squared split is not a technical disagreement—it is the formal decoupling of China's and Korea-Japan's memory roadmaps. CXMT's 2027 capacity doubling bets on a closed domestic AI chip ecosystem; SK Hynix is pivoting capex from DRAM expansion into HBM and AI infrastructure, reasoning that NVIDIA-powered sovereign compute (Sferical One in Sweden as the template) locks in HBM premium cycles through 2027. These paths will not converge soon, but HBM4 in 2026 forces a head-on collision. The quieter signal sits in talent flows: MediaTek poaching Intel's packaging chief is not backfilling a role—it is buying time for 2.5D/3D stacking on in-house AI SoCs. Advanced packaging has graduated from a foundry afterthought to a core design moat, and the CoWoS bottleneck is pulling competition back from transistor scaling to process engineering. On risk, the Panasonic plant explosion in Taiwan, China, stress-tests single-node fragility. Under rigid AI demand, any line outage gets amplified into a quarterly supply shock. CXMT's expansion, if it triggers another export-control escalation, will reprice global DRAM bargaining power. 12–24 month call: HBM, not logic, becomes the #1 AI compute bottleneck; the 4F-squared dual-track hardens; packaging talent premiums rise another 40%.