[News] AI Interposers Grow Larger, Raising HBM4E Cooling Challenges: How Samsung, SK hynix Are Responding - TrendForce

Larger interposers don't just add area—they multiply thermal flux density non-linearly. HBM4E pushing 16+ stacks past 400W per package exposes a structural flaw: copper interconnect thermal resistance in CoWoS-class packaging hasn't kept pace. The real bottleneck sits between TSV arrays and organic substrates, not in the memory die itself. Samsung's vertical thermal via approach and SK hynix's microchannel liquid-cooling represent fundamentally different engineering philosophies—one internalizes heat extraction, the other externalizes it to the system level. Upstream, ABF substrate and TIM suppliers face requalification cycles; panel makers in Taiwan, China will see utilization volatility from design pivots. Downstream is harsher: NVIDIA and AMD must reserve thermal headroom in next-gen accelerators, forcing die-area reductions that directly cap compute density. Competition-wise, if Micron can't resolve the HBM4E yield-thermal coupling by 2026, its roughly 20% NVIDIA supply share erodes further. Samsung's window isn't die performance—it's integrated package-thermal delivery speed. The 12-to-24-month tail: thermal management escalates from engineering parameter to commercial moat. Whoever locks in liquid-cooling interface standardization first owns pricing power over the next AI cluster generation.