ITRI–TSMC semiconductor ecosystem

Public absorption preceded commercial selection

Taiwan's semiconductor policy began from a capability gap, not from an existing private manufacturing base. A peer-reviewed 1994 reconstruction says that in 1973 Taiwan's integrated-circuit activity consisted of assembly, after which the government selected IC technology, brought it from RCA through the Industrial Technology Research Institute, assimilated and improved it, and transferred the resulting capability to industry.¹ This sequence matters: the public intervention did not merely subsidize a firm that already knew how to produce chips.

ITRI's fiftieth-anniversary history supplies the participant account. It dates the institute's founding to 1973, describes a 1974 policy meeting that selected integrated circuits as a target, and records a 1976 technology-transfer and licensing agreement with RCA. Its account follows engineers through RCA training and the installation of a Taiwanese demonstration line.² The page is authoritative for how ITRI records its own program; the independent study is stronger evidence that selection, assimilation, improvement, and industrial transfer formed a connected mechanism.

A license could specify designs and processes, but the demonstration line made performance observable in local operation. People, equipment, production problems, and repeated adjustment converted an imported package into an organizational capability. This is the central inference: ITRI temporarily held technical and financial uncertainty that private firms had little reason to bear, while building a body of practice that could later be tested by commercial orders.

ITRI's history says it created United Microelectronics Corporation in 1980 and transferred a four-inch wafer process and an R&D team to the company. It then places Morris Chang's return and the six-inch VLSI project before TSMC's 1987 creation.³ A 2024 U.S. International Trade Commission staff working paper independently gives the same broad handoff—ITRI, the RCA transfer, and a 1987 TSMC joint venture with Philips—and describes TSMC as a pioneer of the foundry model.⁴ The precise “first” language remains a participant and later analytical characterization, not a claim that every possible predecessor has been exhaustively ruled out.

The state designed handoffs rather than one permanent hierarchy

An-Chi Tung interprets the policy as a catalytic sequence across pre-commercial research, cluster formation, venture capital, private spinouts, and traditional business-group participation. Her peer-reviewed account argues that the government addressed market failures without micromanaging the firms as bureaucratic state enterprises.⁵ That interpretation narrows “state support” into several institutional handoffs: public selection and absorption came before commercial ownership, market demand, and private operating judgment.

Hongwu Sam Ouyang adds a different organizational warning. His principal-agent analysis argues that successful technology policy depended on arrangements that controlled bureaucratic agents and mediated agency problems; state intervention's success should not be assumed.⁶ The two studies are compatible on the importance of design but emphasize different hazards. Tung foregrounds the boundary between catalysis and micromanagement; Ouyang foregrounds the political and organizational controls required inside the public apparatus.

Nor does the semiconductor result establish a portable formula. Tung calls the model apparently replicable elsewhere, while Gregory Noble's comparative working paper documents an ITRI-supported hard-disk-drive effort that did not keep pace with short product cycles, entry barriers, and scale. Noble also argues that firms and ITRI redirected capability toward related CD-ROM products.⁷ The disagreement is productive: public absorption and spinout can create options, but architecture, timing, scale, and market selection still determine whether a durable industry forms.

Pure-play manufacturing made noncompetition a boundary rule

TSMC's current annual report describes a commitment made when the company opened in 1987: it would not design, manufacture, or market integrated-circuit products under its own name. The company presents that noncompetition rule as the foundation of customer trust and contrasts it with integrated-device manufacturers that may prioritize their own products.⁸ This is company evidence for its stated business rule and rationale, not independent proof that every capacity decision treats every customer equally.

Tung's independent account describes the resulting system as virtual integration adapted to fabless and chipless operations, away from the older integrated-device model.⁵ Manufacturing remained capital- and process-intensive, but a design firm no longer had to own that hierarchy. It could contract across a standardized organizational boundary while the foundry pooled equipment, yield learning, and demand from many customers. The customer benefit is therefore narrower than “outsourcing”: access to specialized manufacturing lets a designer avoid many costs and operating risks of owning a fab while retaining responsibility for product design.⁸

The two company relationships in the structured record have different types. Intel is a comparison case for the integrated-device boundary; no historical influence or procurement relationship is asserted here. NVIDIA is a documented manufacturing relationship: NVIDIA's fiscal-2025 Form 10-K says it uses a fabless and contract-manufacturing strategy and names TSMC and Samsung as wafer foundries.⁹ NVIDIA also says this arrangement avoids many ownership and operating risks while reducing its direct control and creating geographic and supplier dependencies. That record supports contractual interdependence, not a claim that either company caused the other's organizational design.

Local specialization remained open to global networks

The cluster was neither one vertically integrated corporation nor a collection of anonymous spot markets. Sue-Ching Jou and Dung-Sheng Chen assembled organizational-event data for thirty-two large Taiwanese IC firms from 1976 to 1996. They found increasing local and international ties as the industry matured, with relationships spanning technology, finance, manufacturing, and other forms of cooperation.¹⁰ Their evidence supports a vertically specialized, mixed local-global network rather than the idea that Hsinchu succeeded through geographic proximity alone.

The study also establishes its own limits. The authors excluded many small firms because their relationships could not be traced reliably, attempted interview verification but could not verify every event, and stopped the series in 1996.¹⁰ It is strong historical evidence about the architecture and development of interfirm ties; it is not a current map of every supplier or an outcome evaluation for workers and communities.

The six primary idea relationships are analytical lenses, not documented lines of intellectual influence. Structure, hierarchy, and scale captures the move from public institute to spinout and from integrated firm to specialized network. Cooperation, incentives, and organizational equilibrium captures the noncompetition promise and reciprocal dependence across firms. Knowledge, expertise, and professional autonomy and learning, quality, and reliability capture absorption, pilot production, yield improvement, and cumulative process practice. Strategy, competition, and adaptation and innovation, entrepreneurship, and renewal capture institutional sequencing, commercial selection, spinout, and the possibility of redirecting capability after failure.

Those same knowledge, strategy, and innovation pages are related comparisons in the broader record. Executive attention, information, and organizational sensing is another comparison: officials, advisers, engineers, customers, and markets made different uncertainties visible at different stages. None of these links claims that a named framework was read by, influenced, or descended from ITRI or TSMC.

Concentrated capability created dated success measures and present burdens

The industrial concentration is measurable, but the dates matter. The May 2024 USITC staff working paper reports that Taiwan accounted for 60 percent of global foundry revenue and, in 2021, produced 92 percent of chips below 10 nanometers; it also reports that integrated circuits made up nearly 40 percent of Taiwan's 2022 exports by value.⁴ These figures establish substantial concentration for their stated periods, not a live market-share estimate. The paper also warns that changes in Taiwan's production position could affect its economic and national-security prospects, while explicitly stating that it is staff research rather than a Commission finding.

Concentration reaches customers and suppliers differently. Fabless customers gain access to costly manufacturing, but NVIDIA's filing also describes limited control, non-guaranteed supply, long lead times, and geographically concentrated suppliers.⁹ Suppliers and partners gain routes into a large design and manufacturing network. TSMC reports that its current purchasing program sets or supports supplier targets for greenhouse gases, energy, water, waste, and regulated materials.¹¹ Neither company reporting nor the historical network study measures how negotiating power, compliance cost, or surplus is distributed across the supplier base.

The resource record shows both mitigation and expansion pressure. TSMC reports that in 2024 renewable sources accounted for 14.1 percent of its power use and reclaimed water replaced 17 percent of tap-plus-reclaimed water at the covered sites. The same report says unit greenhouse-gas emissions were 19 percent above the 2020 base and unit water consumption 14.3 percent above the 2010 base, missing the stated reduction targets.¹¹ DNV gave the report limited assurance using sampling and the Chinese version; that review does not turn company-selected measures into a complete ecological or distributional assessment.

Independent Taiwanese research reaches beyond the company boundary. A 2025 report from the Research Institute for Democracy, Society and Emerging Technology and the Environmental Rights Foundation concludes that advanced-fab expansion intensifies regional water and electricity pressure under climate change. It credits reclaimed-water investment while warning that supply depends on local sewerage connections, public infrastructure, interregional power, and the timing of new plants.¹² The report is policy analysis, not representative testimony from every affected community. It supports the infrastructure burden and climate-risk claim but not a quantified net community effect.

Worker evidence is thinner still. TSMC reports more than 8.4 million employee training hours, an injury frequency rate of 0.26, no employee occupational fatalities, one contractor occupational fatality, and 6.3 percent of employees with reported high stress for 2024.¹¹ A 2008 study of 162 employees in Hsinchu semiconductor factories found that safety-law understanding, public policy, legal compliance, accident prevention, communication, and complete standards formed salient safety-management factors; perceptions differed by role and worker characteristics.¹³ That study did not identify TSMC as the employer and is too old to establish current conditions. The public record reviewed here cannot support the earlier broad claim that long hours and frontier pressure characterize workers across the ecosystem.

The coding emphasizes mechanisms, not a complete social balance sheet

The highest idea score, 3, goes to structure, hierarchy, and scale, cooperation, incentives, and organizational equilibrium, knowledge, expertise, and professional autonomy, learning, quality, and reliability, strategy, competition, and adaptation, and innovation, entrepreneurship, and renewal because the cited record directly develops institutional boundaries, network cooperation, technology absorption, production learning, competitive selection, and spinout.

A score of 2 marks substantial but secondary treatment of purpose, mission, and institutional legitimacy, coordination, communication, and common understanding, work design, productivity, and automation, and executive attention, information, and organizational sensing. A score of 1 marks contextual treatment of authority, legitimacy, and acceptance, delegation, decentralization, and responsibility, decision making, judgment, and bounded rationality, and measurement, accounting, and control. The zero scores for governance, stewardship, and accountability, culture, informal organization, trust, and voice, and organizational ignorance mean that the available account does not substantially develop those concepts; zero does not assert that they were absent.¹⁴

The organizational profile follows the same evidence. State bureaucracy, market capital, and professional expertise supplied different forms of authority. Decisions moved among central executives, specialist cells, and divisions; ownership moved among state sponsorship, a public corporation, and partnership networks. Planning, markets, standards, and modular interfaces coordinated the system, while specialist, peer-networked, and bidirectional knowledge flows supported formal research, experimentation, market feedback, and continuous improvement. Operational, quality, and financial measures made some outcomes visible. Central reconfiguration, modular recombination, and competitive selection supplied different adaptation modes.¹⁴

The affected-subject record is deliberately uneven. Customer access and partner opportunity are supported benefits. Public institutions receive strategic and economic capability while retaining agency, infrastructure, and geopolitical risk. Current company metrics show worker training and safety systems alongside stress and a contractor fatality, but representative worker voice is missing. Communities and ecosystems carry documented water, power, and climate-related burdens, while household distribution, consent, land effects, pollution, health, and net ecological outcomes remain unresolved. Industrial success is well sourced; a complete social balance sheet is not.