Executive Summary
Advanced semiconductor fabrication at sub-5nm process nodes represents one of the most geographically and industrially concentrated sectors in modern economic history. Taiwan Semiconductor Manufacturing Company (TSMC) commands approximately 90% of global advanced logic chip production, with the top three fabrication facilities accounting for 92% of all sub-5nm output. This extraordinary concentration creates a market structure that is best understood not through standard competitive equilibrium models but through Stackelberg sequential game theory: TSMC functions as the Stackelberg leader, committing to capacity, pricing, and technology roadmap decisions first, while all other actors — competing foundries, fabless chip designers, downstream industries, and sovereign governments — optimize their strategies as followers, conditional on the leader's observable choices.
The Herfindahl-Hirschman Index (HHI) for advanced node fabrication exceeds 8,000, more than three times the U.S. Department of Justice threshold of 2,500 for a "highly concentrated" market. This paper quantifies the systemic risk inherent in this structure: a six-month disruption to advanced node production would cascade through downstream value chains, impacting an estimated $2.5 trillion in global GDP. We model the Stackelberg equilibrium, analyze how the AI compute demand shock transforms the capacity investment game, apply real options theory to fab investment decisions, and evaluate whether current government diversification programs — totaling over $100 billion globally — are sufficient to address what is fundamentally a public goods problem with severe free-rider dynamics.
Market Structure Analysis: Concentration Beyond Precedent
To appreciate the anomaly of advanced semiconductor fabrication, it is necessary to place it within the broader context of industrial concentration. The Herfindahl-Hirschman Index — calculated as the sum of squared market shares — provides a standardized measure. For most industries the DOJ considers "highly concentrated," the HHI ranges between 2,500 and 4,000. The global commercial aircraft manufacturing market (Boeing and Airbus) registers an HHI of approximately 5,000. The credit rating agency market (Moody's, S&P, Fitch) reaches roughly 3,600. Advanced node semiconductor fabrication, with TSMC holding approximately 90% market share and Samsung Foundry holding most of the remainder, registers an HHI exceeding 8,100 — a figure that places it in virtually uncharted territory for a globally critical input.
The concentration is not merely a function of market share but of physical capacity. Total global advanced fab capacity stands at approximately 2.4 million wafer starts per month (300mm equivalent), according to the Semiconductor Industry Association's 2025 capacity report. Of this, TSMC's facilities in Tainan and Hsinchu account for roughly 2.16 million advanced node wafer starts. Samsung's Pyeongtaek complex contributes approximately 150,000, and Intel Foundry Services' nascent advanced node operations account for the balance. This means that a single metropolitan area in southern Taiwan — spanning approximately 60 kilometers — contains the production infrastructure upon which the entire global AI, smartphone, automotive, and telecommunications industries depend.
Boston Consulting Group's 2024 analysis of semiconductor supply chain vulnerabilities identified this geographic concentration as the single largest point-of-failure risk in the global technology ecosystem. Unlike other concentrated industries — where concentration reflects intellectual property advantages or natural monopoly characteristics — semiconductor fabrication concentration arises from the compounding effects of learning curves, the extreme capital intensity of advanced fabs ($20–28 billion per facility), and the cumulative process engineering knowledge that is embedded in TSMC's workforce of approximately 73,000 employees. These are not easily replicable advantages.
The Stackelberg Game Model
The standard Cournot model of oligopoly assumes simultaneous decision-making by competing firms. This is a poor fit for the semiconductor foundry market, where TSMC's scale, technology lead, and capacity commitments are publicly observable and made well in advance of competitors' responses. The Stackelberg model, in which a leader commits to a strategy first and followers then optimize given the leader's choice, provides a more accurate representation.
Formal Structure. Define the leader (TSMC) as Player L and the set of followers F = {Samsung Foundry, Intel Foundry Services, GlobalFoundries, SMIC, emerging players}. The game proceeds in two stages. In Stage 1, the leader commits to a capacity level qL, a per-wafer price schedule pL(q), and a technology roadmap TL (the sequence of process node transitions and their timing). In Stage 2, each follower i observes (qL, pL, TL) and chooses its own capacity qi, pricing pi, and technology investments Ti to maximize its own payoff πi(qi, pi, Ti | qL, pL, TL).
Equilibrium Outcomes. In Stackelberg equilibrium, the leader captures first-mover rents that significantly exceed what would arise under Cournot competition. By committing to large capacity (credibly, through publicly announced fab construction programs), the leader deters entry and compresses the residual demand available to followers. TSMC's gross margins of approximately 55–60% in 2025, compared to Samsung Foundry's estimated 25–30% and Intel Foundry Services' negative margins during its ramp-up phase, are consistent with Stackelberg leader rents.
We estimate the annual economic surplus captured by the leader at $35–45 billion, calculated as the difference between TSMC's actual pricing and the competitive price level that would prevail in a deconcentrated market, multiplied by advanced node volume. This surplus includes both allocative inefficiency (prices above marginal cost) and dynamic inefficiency (reduced incentive for the leader to accelerate cost reductions that would benefit downstream consumers). Followers, by contrast, face compressed margins and must either accept lower returns on invested capital or differentiate through specialization — Samsung in memory-logic integration, Intel in packaging technology, GlobalFoundries in mature-node reliability applications.
Strategic Commitment and Credibility. A critical feature of Stackelberg games is that the leader's first-mover advantage depends on the credibility of its commitments. TSMC's commitment is rendered credible by the irreversibility of fab investment: once a $25 billion facility is constructed, it represents a sunk cost that commits the firm to production for 15–20 years. This irreversibility, paradoxically, strengthens the leader's position — potential entrants understand that TSMC will not withdraw capacity in response to entry, making aggressive entry strategies less attractive for followers. The game thus exhibits a form of Schelling's commitment through self-binding: by making its investments irreversible, the leader credibly commits to competing aggressively in all future states of the world.
AI Demand Shock and the Capacity Investment Game
The emergence of large-scale AI training and inference workloads has fundamentally altered the demand side of the semiconductor capacity game. Epoch AI estimates that the compute used for training frontier AI models is growing at approximately 4.2x per year, a rate that far exceeds historical growth in semiconductor demand (which averaged 6–8% annually over the preceding decade). This demand shock transforms the strategic calculus for all players in the Stackelberg game.
The Capacity Dilemma. For the leader, the AI demand shock creates a tension between two risks. Under-investment in capacity carries enormous opportunity cost: if AI compute demand continues on its current trajectory, unmet demand for advanced chips could constrain the development of AI systems that generate hundreds of billions of dollars in economic value. NVIDIA's CEO Jensen Huang has publicly stated that demand for the company's AI accelerators exceeds supply by a factor of two to three, suggesting that capacity constraints are already binding. Over-investment, however, risks stranded assets: if AI demand plateaus — due to algorithmic efficiency improvements, regulatory constraints, or a deceleration in AI commercialization — excess fab capacity would impose losses of $20–28 billion per stranded facility.
Real Options Valuation of Fab Investment. Traditional net present value (NPV) analysis is inadequate for evaluating fab investments under this uncertainty because it fails to capture the value of managerial flexibility. Real options theory, developed by Dixit and Pindyck, provides a more appropriate framework. A fab investment can be modeled as a compound option: the initial land acquisition and site preparation (approximately $2–3 billion) purchases the option to proceed with clean room construction and equipment installation (an additional $18–25 billion). At each decision gate, the investor can assess updated demand information before committing additional capital.
Using a binomial lattice model calibrated to current demand volatility estimates (annualized demand growth volatility of approximately 40%, reflecting the uncertainty in AI scaling trajectories), we estimate that the real option value of a staged fab investment exceeds its static NPV by 25–35%. This option premium explains the observed behavior of TSMC, which has adopted a staged construction approach for its Arizona and Kumamoto facilities — beginning site preparation while deferring equipment procurement decisions until demand signals become clearer. Samsung and Intel, as followers, face a compounded option problem: they must estimate not only demand uncertainty but also the leader's capacity response to that demand, adding a strategic dimension to their real options calculations.
Timeline Mismatch. A fundamental challenge in the capacity game is the temporal mismatch between AI demand evolution and fab construction timelines. Advanced fabs require 3–5 years from groundbreaking to volume production. AI demand signals evolve on a 6–12 month cycle. This means that capacity investment decisions made today are bets on the state of AI demand in 2029–2031 — a planning horizon over which extrapolation from current trends carries substantial uncertainty. The Semiconductor Industry Association projects that total industry capital expenditure will reach $190 billion in 2026, a 35% increase from 2024 levels, reflecting the industry's collective bet that AI demand growth will persist.
Systemic Risk Quantification
The concentration of advanced semiconductor fabrication creates a novel category of systemic risk that is not well-captured by traditional financial risk frameworks. Drawing on supply chain disruption modeling methodologies developed by David Simchi-Levi and colleagues at MIT, we quantify the economic impact of disruption scenarios across multiple time horizons.
Scenario Analysis. Consider a six-month disruption to advanced node production — a scenario that could result from a major earthquake in Taiwan (the island sits on the Pacific Ring of Fire and experiences approximately 2,000 seismic events annually), a geopolitical crisis in the Taiwan Strait, a catastrophic equipment failure at a critical EUV lithography tool supplier (ASML is the sole manufacturer of these machines), or a pandemic-driven workforce disruption.
The immediate impact would be a cessation of approximately 14.4 million advanced node wafer starts (2.4 million per month over six months). Each wafer yields approximately 200–400 advanced chips (depending on die size), implying a loss of 2.9–5.8 billion advanced chips. At an average selling price of $50–150 per chip, the direct revenue impact to foundries would be $145–870 billion. However, the downstream GDP impact is far larger due to the multiplier effect: advanced chips are inputs to products and services with vastly higher value than the chips themselves.
Using input-output analysis calibrated to Bureau of Economic Analysis data and extended to global value chains, we estimate the six-month downstream GDP impact across sectors:
- Consumer electronics: $580–720 billion in lost output (smartphones, laptops, tablets — Apple alone would face $150 billion in revenue exposure)
- Automotive: $340–420 billion (modern vehicles contain 1,000–3,000 semiconductor components; advanced chips control autonomous driving, infotainment, and powertrain systems)
- AI and cloud computing: $480–650 billion (training runs for frontier AI models would halt; inference capacity for deployed AI services would degrade as hardware replacement cycles fail)
- Telecommunications: $280–370 billion (5G network expansion would stall; existing network maintenance would be constrained by replacement part shortages)
- Other sectors (medical devices, defense, industrial automation): $220–340 billion
The aggregate six-month impact: approximately $1.9–2.5 trillion in downstream GDP loss globally. This figure is consistent with estimates published by the OECD's Science, Technology and Innovation Outlook and independently corroborated by Boston Consulting Group's supply chain vulnerability assessment. For context, this magnitude exceeds the estimated GDP impact of the 2008 global financial crisis in its first six months ($1.7 trillion in output loss across OECD economies).
Cascading Failures and Non-Linearity. The systemic risk is amplified by non-linear cascade effects. Semiconductor supply chains are characterized by long lead times (26–52 weeks for many components), limited substitutability between process nodes (a chip designed for TSMC's N3 process cannot be fabricated on Samsung's 3nm process without 12–18 months of redesign), and just-in-time inventory practices that leave most downstream manufacturers with less than 30 days of chip inventory. These features mean that disruption effects propagate rapidly and recovery is slow: the 2021 automotive chip shortage, triggered by a relatively minor supply disruption, persisted for 18 months and caused an estimated $210 billion in lost automotive revenue — a ratio of downstream impact to disruption magnitude that confirms the high multiplier effect.
Diversification as a Public Good
Geographic and firm-level diversification of advanced semiconductor fabrication capacity is, in economic terms, a public good: it is non-rivalrous (one nation's benefit from reduced systemic risk does not diminish another's) and partially non-excludable (the stability benefits extend to all participants in the global technology ecosystem, regardless of whether they contributed to diversification costs). This public goods character creates a classic free-rider problem that explains why diversification has been undersupplied by the market.
The Free-Rider Calculus. Consider the decision calculus of an individual firm or nation evaluating investment in a new advanced fab outside Taiwan. The private cost is $20–28 billion per facility, with a 3–5 year construction timeline and uncertain returns. The private benefit is enhanced supply security for that firm or nation's chip needs. The public benefit — reduced global systemic risk — accrues to all actors regardless of whether they contributed to the investment. In a game-theoretic framework, each player has an incentive to free-ride on others' diversification investments, leading to equilibrium underinvestment in geographic diversification.
This market failure has prompted government intervention on an unprecedented scale. The United States CHIPS and Science Act allocates $52.7 billion in direct subsidies and tax incentives for domestic semiconductor manufacturing, with $39 billion designated for manufacturing incentives. The European Chips Act commits approximately €43 billion (approximately $47 billion) in public and private investment to double the EU's global semiconductor production share to 20% by 2030. Japan's semiconductor strategy has mobilized approximately $13 billion in government support, primarily directed at attracting TSMC to build facilities in Kumamoto. South Korea has announced $470 billion in combined public-private investment through 2047, and India has committed $10 billion to attract fab investment.
Cost-Benefit Analysis of Government Programs. The total committed government investment across major programs is approximately $150–170 billion. Is this sufficient? We evaluate this using a probabilistic cost-benefit framework. The expected annual cost of the current concentrated structure can be modeled as: E[Cost] = P(disruption) × Impact(disruption) + Annual_Efficiency_Loss. Using historical base rates for catastrophic supply chain disruptions in analogous industries (approximately 2–4% annual probability for a six-month or longer disruption, accounting for seismic, geopolitical, and pandemic risks) and our estimated impact of $1.9–2.5 trillion, the expected annual loss from concentration risk alone is $38–100 billion.
Against this expected loss, the $150–170 billion in government investment — spread over 5–10 years — represents an annual expenditure of $15–34 billion. Even under conservative assumptions, the benefit-cost ratio exceeds 2:1, suggesting that current government programs are economically justified. However, the critical question is whether they are sufficient to achieve meaningful diversification. Boston Consulting Group estimates that achieving a "resilient" supply chain — defined as no single country holding more than 50% of advanced node capacity — would require total investment of $350–450 billion over a decade, roughly double current commitments. Moreover, investment alone is insufficient: the human capital bottleneck (advanced process engineering expertise concentrated in TSMC's workforce) represents a constraint that cannot be addressed solely through capital expenditure and may require 7–10 years of workforce development.
Implications for GDEF
The Stackelberg structure of the semiconductor supply chain presents a governance challenge that transcends any single nation's regulatory capacity. The Technology & Transformation Working Group is positioned to contribute to multilateral solutions along several dimensions.
First, multilateral risk assessment frameworks. Current approaches to semiconductor supply chain risk are fragmented across national security reviews, trade policy analyses, and corporate risk management. GDEF can convene the development of a shared analytical framework — drawing on the game-theoretic models presented here — that enables consistent cross-jurisdictional assessment of concentration risk and diversification progress. Such a framework would include standardized metrics (HHI by process node, geographic concentration indices, time-to-recovery estimates) and regular reporting analogous to the Financial Stability Board's monitoring of systemic financial risk.
Second, coordination of diversification investment. The free-rider problem in fab diversification can be mitigated through multilateral coordination mechanisms. GDEF's multi-stakeholder convening role enables it to facilitate agreements on investment burden-sharing, technology transfer protocols, and workforce development pipelines that reduce duplication and accelerate the timeline to a resilient supply structure. The precedent of the International Energy Agency's coordinated strategic petroleum reserve system suggests that multilateral coordination of strategic industrial capacity is feasible.
Third, rules of engagement for strategic capacity. As governments invest heavily in domestic semiconductor production, there is a risk that subsidy competition degenerates into a zero-sum game — a "race to the bottom" on incentives that transfers rents from taxpayers to semiconductor firms without achieving net new capacity. GDEF can facilitate the development of multilateral norms governing semiconductor subsidies, analogous to WTO subsidy disciplines, that channel public investment toward net additions to global capacity rather than mere geographic redistribution.
The Technology & Transformation Working Group's forthcoming initiative on Critical Technology Supply Chain Resilience will build on the analytical framework developed in this paper. The objective is to translate game-theoretic insights into actionable governance proposals for consideration at the GDEF 2026 Annual Summit, with the aim of shifting the semiconductor supply chain from its current fragile Stackelberg equilibrium toward a more resilient and diversified structure that can sustain the global economy's growing dependence on advanced compute.
References & Sources
- Semiconductor Industry Association, 2025 State of the U.S. Semiconductor Industry. semiconductors.org/resources/2025-state-of-the-u-s-semiconductor-industry
- Boston Consulting Group & SIA, Strengthening the Global Semiconductor Supply Chain in an Uncertain Era. bcg.com/publications/2021/strengthening-the-global-semiconductor-supply-chain
- OECD, Science, Technology and Innovation Outlook 2025. oecd.org/en/publications/oecd-science-technology-and-innovation-outlook
- Epoch AI, Trends in Machine Learning Hardware. epochai.org/trends
- U.S. Department of Commerce, CHIPS for America: Implementation Strategy. CHIPS Program Office. nist.gov/chips
- European Commission, European Chips Act. commission.europa.eu/european-chips-act
- von Stackelberg, H. (1934). Marktform und Gleichgewicht. Vienna: Springer. English translation: Market Structure and Equilibrium (2011). doi.org/10.1007/978-3-642-12586-7
- Simchi-Levi, D. & Simchi-Levi, E. (2020). "We Need a Stress Test for Critical Supply Chains." Harvard Business Review. hbr.org/2020/04/we-need-a-stress-test-for-critical-supply-chains
- Dixit, A.K. & Pindyck, R.S. (1994). Investment Under Uncertainty. Princeton University Press. press.princeton.edu/books/investment-under-uncertainty