Taiwan Semiconductor Manufacturing announced a second $100 billion capital commitment to U.S. fabrication capacity, doubling its Arizona footprint to six fabs producing sub-3nm process nodes. The decision comes sixteen months after the initial $40 billion Phoenix expansion and shifts the world's largest contract chipmaker toward a 25-30% U.S. capacity allocation by 2030.
The commitment locks in advanced packaging lines and 2-nanometer production at scale, targeting AI accelerator demand from Nvidia, AMD, and Apple. TSMC's Arizona operations will run the same N2 and N2P nodes scheduled for Taiwan's Fab 20, removing the 18-24 month lag that historically separated offshore from home-island tooling. First wafers are expected in late 2025, with volume production in 2027. The company declined to specify subsidy capture under the CHIPS Act but confirmed $6.6 billion in federal grants already allocated to the Phoenix site will apply toward the expanded footprint.
This move is margin arbitrage disguised as supply-chain resilience. TSMC's leading-edge nodes command 60-65% gross margins in Taiwan; U.S. fabs historically run 10-12 percentage points lower due to labor, permitting, and material costs. The company is betting that AI inference chips—where compute density and power efficiency justify premium pricing—can absorb the delta. Hyperscalers are paying 15-20% premiums for domestically fabbed silicon under classified government contracts, and TSMC is repositioning to capture that spread at volume.
The geopolitical read is secondary but material. A $100 billion build-out in Arizona is a hedge against Taiwan Strait interdiction risk, but it is not an insurance policy. Six fabs cannot replicate the 12 million wafer annual capacity TSMC runs in Hsinchu and Tainan. The U.S. footprint is a carve-out for AI and defense-critical SKUs, not a wholesale pivot. The company still plans $30-35 billion in Taiwan capex annually through 2026, concentrating 1.4nm and gate-all-around transistor R&D on the island.
Allocators should watch ASML's high-NA EUV delivery schedules. TSMC's 2nm nodes require the $380 million Twinscan EXE:5000 tools, of which fewer than 20 units exist globally. Any delay in Arizona tooling pushes first silicon into 2028 and narrows the window before Samsung's Texas fabs come online with competing 2nm capacity. Track permitting timelines in Maricopa County; construction delays have already pushed the first Phoenix fab back nine months.
Watch also for customer prepayments. Apple and Nvidia typically front 20-30% of fab costs in exchange for capacity guarantees. If TSMC announces a $20-25 billion prepayment tranche in the next 90 days, the U.S. buildout is already sold. If not, the company is eating the carry cost and betting on 2027 AI demand that does not yet have contracts behind it.