Taiwan manufacturing demand rises as bank property loans decline

The credit shift from real estate to manufacturing in Taiwan, China is a repricing of risk in the AI compute race, not a routine cycle. Banks pulling liquidity from low-yield property into HPC and AI server capacity signals a structural bet on manufacturing as the primary risk asset. The binding constraint sits at advanced packaging, not logic. CoWoS and SoIC interconnect capacity remains scarcer than leading-edge lithography. ABF substrate and HBM suppliers see order overflows first; ODM capex cycles stretch into 2026. Two compliance risks matter. US export control costs now consume 2-3% of fab revenue and are climbing. Credit concentration in one technology corridor creates fragility—if AI monetization disappoints, NPL ratios could spike within 18 months. The 2000 fiber-optic bubble is not a distant analogy. Competitively, Samsung and Intel leverage CHIPS Act subsidies to close the 2nm gap, but their cost of capital runs 150-200bp higher, making expansion structurally slower. NVIDIA shifts the moat from chip design to system delivery by locking CoWoS and HBM supply. The 12-24 month thesis: the bottleneck migrates from wafer output to power and thermal infrastructure. When datacenter power draw exceeds grid capacity, "chips without electricity" becomes more lethal than chip scarcity. Energy-sovereign nodes will command a structural premium.