High-bandwidth memory: memory chips stacked on top of each other and placed beside a processor, so data reaches it much faster.
AI accelerators depend on it, and only a few manufacturers make it, which has made HBM one of the tightest bottlenecks in AI hardware.
Because rules judge only the processor, memory worth about $4,000 a unit in China is moving through Hong Kong and Shenzhen, with enforcement expected to tighten.
The talks come as chip makers warn that DRAM supply could stay short of demand into 2028.
The stoppage could spread to Micron's largest Taiwan plant, where a union has not yet authorized action and is due in mediation the same day.
The roadmap points to bigger multi-chiplet systems with tighter interconnects and stronger power delivery, extending EMIB into larger package classes.
Most of that agent traffic is cached prompt activity, which shifts the pressure from billing to memory capacity.