One hundred fifty billion dollars over four years is the number Nvidia put on Taiwan last week. The number is a physical claim. It commits the B200 NVL72 supply chain to TSMC's Hsinchu capacity, specifically to the CoWoS-S substrate packaging line that integrates HBM3e memory at yields no alternative foundry currently matches at volume. BIS export controls issued in October 2023 restrict who receives the finished product; they say nothing about where the foundry capacity sits. The $150 billion answers that: in a jurisdiction that three PLA encirclement exercises since August 2022 have treated as a rehearsal address.

Anthropic's confidential IPO filing and Alphabet's $80 billion capital raise, both announced this week, are bids on B200 and successor Nvidia hardware that does not exist without TSMC's CoWoS-S throughput at Hsinchu. On the PRC side, Huawei's Ascend 910B runs on SMIC's N+2 node (approximately 7nm-class) with HBM2e memory, against the B200's HBM3e at roughly 8TB/s per chip. The gap is process-node-imposed, not policy-imposed. SMIC has no publicly confirmed N+1 (approximately 5nm-class) volume production date. The N2 capacity ramp at TSMC's Hsinchu fabs is scheduled for volume production beginning Q4 2026, and Nvidia's $150 billion commitment means that capacity is already claimed, closing the redistribution window before it opens.