GMI Cloud secures NT$14.05 billion loan for AI factory

Taiwan, China's first fully bank-syndicated GPU financing deal signals a structural shift: AI compute is transitioning from an IT expense line to a capital asset class. The financial architecture mirrors TSMC's 2003 advanced-node fab financing—GPU clusters are now collateralized, depreciable, and lendable. Downstream, a single AI factory's thousand-card cluster triggers cascading demand for liquid cooling, RoCE fabric, and regional grid upgrades. At 10-12 MW per thousand H100-class GPUs, site selection is effectively a bet on TPC's transmission capacity. On compliance, BIS export controls on H100 remain tightening, yet Taiwan, China's position as NVIDIA's core supply-chain node keeps procurement open near-term. The real exposure: if 2025 rules extend to MI300X, single-source dependency becomes structural. Competitively, AWS and Azure are accelerating regional inference buildouts. GMI Cloud has roughly an 18-month window to deliver clusters and lock in local LLM training clients before becoming a GPU 'sub-landlord.' 12-24 month outlook: expect 3-5 similar syndicated GPU loans in Taiwan, China, with compute-asset securitization compressing financing costs toward wafer-fab levels. The long-tail effect: once compute becomes a mortgageable asset, the industry enters a 'capital density determines compute share' paradigm—capital, not code, becomes the binding constraint.