TSMC's Chokehold on AI: How Foundry Power Shapes the Capex Cycle

The semiconductor ecosystem supporting artificial intelligence is best understood as a connected industrial system rather than as a sequence of isolated company results. NVIDIA’s product cadence depends upon advanced fabrication, advanced packaging, and high-bandwidth memory; each of these, in turn, depends upon large and sustained investments by foundries, hyperscalers, memory manufacturers, and equipment suppliers. TSMC occupies the system’s central position, while the durability of the current expansion depends on whether AI infrastructure spending continues to generate sufficient economic returns.

The appropriate analytical distinction is between the short run, in which leading-edge capacity and supplier relationships are largely fixed, and the long run, in which new fabs, packaging facilities, memory capacity, and competing architectures can gradually alter the equilibrium. Under present conditions, TSMC’s technological lead and NVIDIA’s demand visibility are substantial. They are not, however, independent of capacity allocation, geopolitical exposure, or the cyclicality of AI capital expenditure.

TSMC’s Entrenched Foundry Position

TSMC’s advantages in advanced manufacturing are both technological and organizational. The company holds a near-monopoly in sub-3-nanometer contract manufacturing ^39,48 and operates a pure-play foundry model that avoids competing directly with its fabless customers ^20,50. High-volume production at the 3nm and 5nm nodes ²⁰ supports gross margins above 60%, materially exceeding the industry average ^2,3,50. Its proprietary advanced-packaging capabilities are even more profitable, with margins reported near 80% ¹⁷.

The financial evidence reflects the strength of this position. TSMC’s second-quarter 2026 net profit increased 61% ³⁷, first-half revenue in New Taiwan dollars rose 35.6% ³¹, and July revenue advanced 44.7% year over year ⁴², with month-to-date growth of 5.6% ⁴³. AI-related demand has become a central growth driver ²⁰, linking TSMC’s current performance directly to the investment decisions of hyperscalers and other large purchasers of computing infrastructure.

The company is also expanding capacity on a considerable scale. TSMC plans 13 new plants in Taiwan ²² and has outlined a capital-expenditure program of approximately $60 billion to $64 billion ²². A potential $265 billion investment in Arizona would extend this expansion geographically ²¹. Public support is an important counterweight to the associated cost: the U.S. CHIPS Act provides a $6.6 billion grant ^39,50, while Japanese government support ⁵⁰ and backing under the EU Chips Act ⁵⁰ assist expansion in other jurisdictions.

These investments should improve resilience over time, but they do not remove short-run constraints. TSMC is targeting 3nm output of 180,000 wafers per month by late 2026 ²⁹ and 2nm output of 100,000 wafers per month ²⁹. The relevant question for NVIDIA is not simply whether aggregate capacity is rising, but whether suitable capacity becomes available at the moment its product roadmap requires it.

NVIDIA’s Dependence on Capacity and Packaging

TSMC remains the primary manufacturer of NVIDIA’s advanced GPUs, making the foundry’s yields, scheduling, and packaging execution integral to NVIDIA’s commercial position. TSMC’s relationships with other major customers illustrate the allocation problem. Apple accounts for approximately 25% of TSMC’s revenue ⁵⁰, while Microsoft has reportedly ordered 300,000 Maia 300 chips ⁴¹. Microsoft’s decision to secure wafer supply directly ⁴¹ is evidence of the competition for leading-edge capacity.

TSMC’s neutral foundry model should, in principle, treat customers equally ⁵⁰. In practice, however, neutrality does not imply unlimited elasticity of supply. When several hyperscalers and fabless designers seek the same advanced nodes, one customer’s incremental demand can reduce the available margin of capacity for another. This competition is particularly relevant as Microsoft, Google, and other hyperscalers increase internal chip development ¹⁴. The resulting pressure need not displace NVIDIA abruptly; it may instead appear gradually through tighter allocation, longer qualification periods, or less favorable economics.

Packaging is a related source of both strength and dependence. TSMC is reportedly developing an advanced-packaging architecture similar to Intel’s EMIB ²². For NVIDIA, such an architecture could complement multi-die designs and potentially reduce reliance on Intel’s technology. TSMC’s manufacturing lead and high-margin packaging position remain central advantages ^20,23, but the same concentration creates a structural vulnerability: a yield problem, geopolitical disruption, or shift in capacity allocation could affect both wafer supply and the assembly of increasingly complex products.

Memory: A Necessary but Cyclical Complement

The AI infrastructure build-out has produced a powerful memory cycle. Samsung’s operating profit has been driven largely by AI-memory sales, with demand for HBM4, DRAM, and NAND increasing sharply ^8,16. AI-related tightening in memory supply has lifted prices and improved profitability ³⁰, while the results have challenged earlier fears of an immediate slowdown in AI expenditure ¹¹. Samsung’s memory revenue grew 471% year over year ⁵, and the company expects AI and high-performance-computing foundry revenue to exceed 30% of its business in 2026 ⁶.

SK Hynix has benefited from the same demand, with AI-related cash flow contributing to approximately $500 billion in SK Group business activity ⁴⁵. This concentration is productive while demand remains strong, but it also increases cyclicality. Excess investment or growing skepticism about the returns on AI infrastructure could reduce growth abruptly ^12,49. Semiconductor earnings remain highly cyclical ^8,36, and the possibility of oversupply is therefore more than a theoretical concern ²⁷.

For NVIDIA, HBM availability is an essential complement to next-generation GPU shipments. Samsung’s effort to regain ground in AI accelerators ³² and its ambitions in 2nm foundry manufacturing ⁶ could eventually create a second source of competitive supply. Nevertheless, TSMC’s yield performance and broader ecosystem advantages remain considerable ^6,25. The near-term equilibrium therefore favors the established suppliers, while the long-run possibility of substitution warrants continued observation.

AI Capital Expenditure: Demand Evidence and Cyclical Risk

AI capital expenditure is generating demand across the physical infrastructure chain. Data-center investment supports power semiconductors ^26,46, optical transceivers ^34,35, and silicon wafers ¹⁵, in addition to processors and memory. Several operating indicators show that this is not merely an expectation embedded in equity prices. Microsoft’s AI spending is converting into revenue ^10,13, and Palantir’s commercial growth provides further evidence of expanding AI demand ⁴⁷.

The supply chain is exhibiting corresponding strength. Qnity Electronics’ advanced-packaging and thermal-management businesses grew by more than 50% year over year ²⁸, while its Semiconductor Technologies segment outpaced the broader market ²⁸. TTM Technologies expects AI-related revenue to more than double in fiscal 2026, supported by hyperscaler PCB demand ³³. Applied Optoelectronics’ 800G revenue increased nearly fivefold sequentially ³⁴, and orders for 1.6T transceivers exceeded $200 million ³⁵. Strong bookings at ASM ⁴⁴ and record revenue at ChipMOS ^1,18 further indicate that the physical construction of AI data centers is proceeding at a substantial pace.

Yet evidence of current demand does not settle the question of long-run sustainability. The durability of AI spending remains contested ^4,13. A normalization of hyperscaler capital expenditure ¹², greater reliance on internally developed chips ¹⁴, or a failure of AI applications to generate adequate returns could transmit a slowdown through the entire ecosystem. The concern is not that demand must suddenly disappear, but that a marginal reduction in spending could expose the fixed-cost structure and inventory sensitivity of suppliers throughout the chain ^24,27.

Geographic Concentration and the Cost of Resilience

Taiwan remains the pivotal geographic node in this system. Integrated-circuit shipments increased 52.3% year over year ⁴⁰, and the sector accounted for 78.5% of Taiwan’s total exports ⁴⁰. This concentration demonstrates the economic importance of the semiconductor cluster, but it also makes any disruption to the island’s production system consequential. A blockade or comparable interruption could halt TSMC’s output catastrophically ⁵⁰. Taiwan’s strategic importance may deter military action ⁵⁰, but the underlying concentration risk remains material for companies dependent on its fabs ⁵⁰.

TSMC’s expansion in Arizona, Kumamoto, and Dresden ^20,50 provides a partial long-run adjustment mechanism. It does not constitute a simple geographic substitute for Taiwan. Construction costs in Arizona and Kumamoto are reportedly 30% to 40% higher than in Taiwan ²⁵, and advanced-packaging capacity outside Taiwan remains years away ⁴⁰. CHIPS Act grants offset some of the additional expense ²⁵, but depreciation from the broader capital program will place pressure on TSMC’s margins ²⁵. Higher depreciation could eventually appear in the form of higher wafer prices or reduced supply flexibility for customers such as NVIDIA.

This is the central trade-off in diversification. Additional facilities improve resilience and reduce the severity of a single-location shock, but they do so at a higher cost and with a long adjustment period. TSMC’s global expansion is therefore best understood as a gradual change in the industry’s risk profile, not as an immediate removal of Taiwan exposure ⁵⁰.

Implications for NVIDIA

NVIDIA occupies a position of considerable strength within this structure, but its strength is interdependent. TSMC’s advanced-node leadership and packaging capability underpin NVIDIA’s ability to maintain an aggressive product cadence. The AI memory cycle currently provides a favorable complement, while the broader infrastructure chain offers evidence of real and continuing demand. Financial results and cash flows across the ecosystem suggest that the present expansion is supported by operating activity rather than by speculation alone ³⁸.

The principal risks are concentrated at the margins. Hyperscaler competition for advanced capacity ^9,41 could reduce NVIDIA’s supply assurance as internal silicon programs expand. The concentration of HBM production among a small number of Korean suppliers creates an additional choke point ²⁹. A geopolitical shock centered on Taiwan ⁵⁰ would be more severe than an ordinary supply interruption, while the high fixed cost of overseas fabrication limits how quickly the system can adapt ²⁵.

NVIDIA’s valuation and investor sentiment will consequently remain sensitive to evidence concerning the return on AI capital expenditure ⁴. Microsoft’s results are an important bellwether ⁷, and the KOSPI’s 16% rebound following Microsoft’s AI earnings illustrates how quickly markets respond to evidence of returns ⁷. Such movements should not be confused with a permanent change in equilibrium; they are signals about the market’s current assessment of demand durability.

What to Monitor

Several indicators are particularly useful for assessing whether the present configuration is strengthening or beginning to adjust:

• TSMC’s monthly revenue reports ²⁰, capacity additions, and execution against its record capital-expenditure program ²².
• Hyperscaler announcements concerning custom silicon and direct wafer procurement, including the implications of Microsoft’s Maia 300 activity ⁴¹.
• HBM supply, pricing, and the progress of Samsung and other suppliers as potential sources of increased substitution.
• The development of overseas fabs, packaging capacity, and the effects of higher depreciation on TSMC’s margins ²⁵.
• U.S. CHIPS Act developments ²⁰ and Taiwan’s export data ¹⁹ as forward indicators of supply-chain stability.

Under current conditions, the evidence supports a strong but conditional conclusion. TSMC’s foundry dominance, NVIDIA’s product position, and the wider AI infrastructure cycle reinforce one another. TSMC’s record investment is adding long-run capacity and geographic resilience, but at higher cost; the memory boom is supplying an important near-term tailwind, but with pronounced cyclicality; and hyperscaler custom silicon introduces a gradual source of substitution. NVIDIA’s trajectory therefore depends not only on the scale of AI demand, but on the industry’s ability to expand capacity, preserve yields, and distribute risk across an evolving semiconductor ecosystem.