TSMC commits another $100B to U.S. fabs — Arizona megaproject reaches $265B

Taiwan Semiconductor Manufacturing announced an additional $100 billion commitment to its Arizona fabrication facilities, raising the total U.S. investment to $265 billion and cementing the site as the largest foreign direct investment in American manufacturing history. The expansion accelerates a timeline originally set for completion by 2030 and adds capacity targeting 2-nanometer and 1.4-nanometer process nodes, the bleeding edge of semiconductor physics where each wafer commands margins above 60 percent.

The announcement follows $6.6 billion in CHIPS Act subsidies awarded in April 2024 and reflects a structural shift in how hyperscalers procure compute. TSMC's Arizona site will house six fabs when complete, producing 600,000 wafer starts per month by 2028 — roughly 15 percent of the company's current global output. The first fab began trial production in Q4 2024, yielding 4-nanometer chips for Apple's A-series processors at parity with Taiwan yields, a milestone that took Samsung's Texas facility three years to reach. TSMC is now pulling forward its second and third Arizona fabs by 18 months, targeting 3-nanometer production by late 2026. The fourth through sixth fabs, previously unannounced, will focus on advanced packaging and chiplet integration, the architecture that lets Nvidia's H200 GPUs stack 141 billion transistors without thermal collapse.

The expansion matters because it rewrites the economics of semiconductor sovereignty. TSMC's Arizona cost structure runs 35 to 40 percent higher than Taiwan due to labor, permitting, and construction timing, but the margin sacrifice buys two things: access to $75 billion in projected U.S. defense and intelligence orders through 2030, and immunity from Taiwan Strait disruption risk that insurance underwriters now price at 8 to 12 basis points annually for semiconductor cargo. The Defense Department's Rapid Assured Microelectronics Prototyping program has already reserved 20 percent of Arizona Fab 2's capacity for secure-node production, paying a 15 percent premium over commercial rates. That contract structure, quietly finalized in November, signals Washington's willingness to subsidize onshore production at scale, not just for strategic rhetoric but for procurement economics that pencil at current AI infrastructure spending rates.

For allocators, the $265 billion figure represents the largest single-company bet on the AI infrastructure thesis outside hyperscaler capex itself. TSMC's trailing twelve-month capex through Q3 2024 was $30 billion, meaning Arizona alone now equals nearly nine years of historical global spend. The company is financing the expansion through a mix of retained earnings, $40 billion in committed credit facilities from a consortium led by JPMorgan and Citigroup, and the CHIPS Act subsidies. Free cash flow through 2024 was $38 billion, giving TSMC runway to self-fund roughly $150 billion of the Arizona commitment without raising equity or materially compressing the 3.2 percent dividend yield that has anchored the stock since 2018. The financing structure assumes AI accelerator demand grows at a 28 percent CAGR through 2030, below the 35 percent Wall Street consensus but above the 22 percent rate TSMC used for its 2023 long-range plan.

Operators should watch three catalysts over the next 18 months: Arizona Fab 2's 3-nanometer qualification timing in Q4 2026, which determines whether TSMC can capture Apple's M-series chip production currently locked in Taiwan; the Defense Department's fiscal 2026 budget language around semiconductor procurement set-asides, expected in March; and any move by Samsung or Intel to match the Arizona footprint, which would signal whether onshore FAB economics work without TSMC's 54 percent global foundry market share to smooth utilization risk. TSMC's Arizona hiring is running 2,400 engineers ahead of schedule, clustering talent in Phoenix that hasn't existed outside of Taiwan, South Korea, and Oregon since the 1990s.

The $265 billion is the number. The question is whether it's the floor or the ceiling for what AI infrastructure requires in atoms, not just bits.