NVIDIA reportedly funds additional CoWoS capacity - igor´sLAB

The binding constraint in AI compute has shifted from transistor density to 2.5D/3D integration. NVIDIA's decision to directly fund CoWoS capacity expansion is a tacit admission that die performance is no longer the bottleneck—packaging is. The ripple effects extend well beyond TSMC's fabs: ABF substrate lead times (Ibiden, Unimicron) and HBM stacking schedules will compress further upstream, while AMD's MI300 roadmap and hyperscaler ASIC programs face tighter allocation at the same packaging nodes in Taiwan, China. Geopolitically, this customer-funded capacity lock is a hedging mechanism against potential US export-control expansion into advanced packaging. If Washington extends restrictions to CoWoS-class interconnects, pre-paid capacity contracts will become the industry's default risk-mitigation play. Over the next 18 months, I expect CoWoS-L (larger reticle) to become the default path for Blackwell Ultra and beyond. Intel's Foveros and Samsung's X-Cube are losing their catch-up window. Packaging is no longer a back-end process—it is the new process node, and whoever controls it controls the AI supply chain.