Weekly news roundup: AI chip race widens across silicon photonics, HBM, memory fabs, and frontier models

The AI chip race has evolved from raw compute competition into a full-stack control battle. Co-design of silicon photonics and HBM5 is forcing EDA toolchain overhauls, while TSMC and packaging partners in Taiwan, China grapple with dual bottlenecks in 3D stacking yield and thermal management. U.S. export controls on advanced packaging tools—now extended to Japan and South Korea—compel Samsung and SK Hynix to localize HBM production, inflating capex by over 15%. NVIDIA and AMD are forming patent cross-licensing alliances, pushing Chinese firms toward alternative Chiplet interconnect standards. Over the next 18 months, geopolitically driven supply chain redundancy will trigger structural oversupply in memory foundry capacity, yet companies mastering CoWoS or Foveros integration will sustain gross margins above 30%.