Taiwan Semiconductor Manufacturing Company (TSMC)

In the AI Story

TSMC's Arizona expansion, announced in 2020 and expanded in subsequent years, represents the most significant geographic diversification of leading-edge chip manufacturing in a generation. The company is constructing multiple fabs in Phoenix with total investment exceeding $65 billion, supported by $6.6 billion in CHIPS Act grants and $5 billion in loans. The first fab, originally scheduled for 2024 production, experienced delays pushing initial output to 2025 and volume production to 2026-2027. The delays reflect not just construction complexity but workforce challenges—the facility requires thousands of experienced semiconductor engineers and technicians, a skillset concentrated in Taiwan and Korea. TSMC is training American workers and temporarily relocating Taiwanese engineers, but the knowledge transfer timeline extends beyond the physical construction timeline. The Arizona fabs will eventually produce significant capacity, but 'eventually' is measured in years, and the fabs will represent perhaps 10-15% of TSMC's total leading-edge capacity even when fully operational.

The comparison to Samsung illustrates why TSMC's lead is difficult to close. Samsung operates leading-edge fabs in Korea and has announced U.S. expansion (Texas, with Intel partnership discussions). But Samsung's yields at equivalent process nodes consistently trail TSMC's by margins that matter economically: if TSMC achieves 90% yield and Samsung achieves 70%, Samsung must process roughly 30% more wafers to deliver the same number of working chips—increasing costs for silicon, energy, time, and equipment utilization. Yield is not merely a technical metric; it is the accumulated result of process knowledge, equipment optimization, and operator expertise built through millions of wafer starts over years. TSMC's yield advantage compounds its technological lead, making it the preferred foundry even for customers who would prefer supply chain diversification for risk management. Capability trumps preference when the performance gap is large enough.

The geopolitical dimension became explicit when U.S.-China tensions intensified in the late 2010s and early 2020s. TSMC manufactures chips for both American (NVIDIA, AMD, Apple) and Chinese (Huawei historically, others) customers, occupying a position where it is both strategic asset and geopolitical pawn. U.S. export controls imposed in 2022-2023 prohibit TSMC from manufacturing advanced chips for certain Chinese entities; the controls put TSMC in the position of enforcing U.S. policy despite being a Taiwanese company. The situation is unstable: if cross-strait conflict disrupts TSMC's Taiwan operations, the global supply of advanced chips halts until alternative capacity comes online—a timeline measured in years even under crash-program conditions. The Arizona fabs provide partial insurance but not replacement; TSMC's Taiwan facilities will remain the majority of leading-edge capacity for at least the remainder of the 2020s.

The financial structure sustains TSMC's position: customers pre-pay for capacity years in advance, providing capital for fab construction and equipment purchases. NVIDIA's multi-billion-dollar orders commit to production slots at future nodes, guaranteeing revenue that justifies the $30-40 billion annual capital expenditures TSMC requires to maintain its technological edge. The financial model works because TSMC is trusted to deliver—forty years of execution history, industry-leading yields, reliable timelines. That trust is itself a form of capital, accumulated through decades and not easily transferred to competitors or new entrants. The combination of technological capability, financial strength, and institutional trust creates a moat that even well-capitalized competitors require years to challenge. The moat protects TSMC's business; it also concentrates global AI capability in a single firm operating primarily in a single geopolitical flashpoint.

Origin

Morris Chang founded TSMC in 1987 after a career at Texas Instruments and a stint leading Taiwan's Industrial Technology Research Institute. The pure-play foundry model was novel—existing semiconductor firms (Intel, IBM, Motorola) designed and manufactured their own chips, viewing fabrication as a proprietary advantage. Chang's insight was that design and fabrication could separate, and that a firm specializing exclusively in fabrication could achieve economies of scale, process excellence, and capital efficiency that integrated manufacturers could not match. The model succeeded beyond initial projections; by the 2000s, fabless design firms (Qualcomm, NVIDIA, AMD after spinning off its fabs) depended entirely on foundries, and TSMC had become the largest and most technically advanced.

TSMC's specific role in AI emerged in the 2010s as machine learning shifted from CPUs to GPUs (requiring TSMC's advanced processes for NVIDIA's chip designs) and intensified in the 2020s with large language models' extreme computational demands. The company's 3-nanometer and 2-nanometer processes, among the most advanced in production, are essential for the chip performance and energy efficiency that frontier AI models require. Smil's framework treats TSMC not as a company but as a critical infrastructure node whose geographic concentration and temporal inertia (four-year fab construction) constrain how fast global AI capability can physically scale. The company is simultaneously the enabler of the AI revolution and one of its most significant bottlenecks.

Key Ideas

Foundry monopoly at frontier. TSMC manufactures ~90% of the world's most advanced chips; no competitor matches its combination of process technology, yield, capacity, and execution reliability at leading nodes.

Arizona diversification timeline. TSMC's U.S. fabs represent significant geographic risk reduction but require four-plus years to reach volume production and will constitute only 10-15% of total leading-edge capacity even when complete.

Yield advantage as moat. Superior manufacturing yields—the percentage of functional chips per wafer—translate directly to cost advantages and capacity advantages that competitors cannot close through capital investment alone; the advantage rests on accumulated process knowledge.

Geopolitical concentration risk. Taiwan's production of 90% of advanced chips, situated 100 miles from mainland China, creates the supply chain's most significant single-point vulnerability—a risk that diversification efforts address but do not eliminate on policy-relevant timescales.

Customer prepayment model. Long-term capacity commitments from NVIDIA, Apple, AMD provide the financial foundation for TSMC's massive capital expenditures; the model requires customer trust built over decades and not easily extended to new foundries.