AI server tracker: Taiwan's AI server supply shows no sign of cooling despite industry uncertainty

A 63.8% YoY revenue surge across 49 AI server supply-chain firms in Taiwan, China is not a linear demand extrapolation—it is the concentrated release of HBM3E capacity ramp and CoWoS packaging bottleneck easing. The ~$98B single-month figure signals a structural shift: global AI capex has moved from exploratory procurement to locked-in inventory positioning. Technically, this validates the acceleration of HBM stack migration from 8-layer to 12-layer and CoWoS-L expansion. Downstream TDP escalation will force liquid-cooling penetration past 60% by 2026. On risk, any extension of export controls to advanced packaging equipment would force a cost restructuring of Taiwan, China's 'middle-layer' role. The 2022 EUV restriction and the subsequent ASML-Nikon share reallocation serve as a cautionary precedent: single-point chokepoints inevitably accelerate alternative-path development. Competitively, hyperscaler ASICs (TPU, Trainium) are transitioning from supplementary to primary compute, compressing general-purpose GPU pricing power while simultaneously inflating the scarcity premium on HBM and advanced packaging. The critical variable over the next 18 months is not demand sustainability but whether power-grid infrastructure and data-center siting can keep pace with chip shipment velocity. The bottleneck is migrating from the wafer fab to the substation.