This instructor report opens only after the May 14, 2002 learner packet is frozen. It evaluates the quality of the cutoff decision process against later evidence; it does not assume that a favorable endpoint made every capital project correct.
What TSMC did
The public record is consistent with TSMC preserving the dedicated-foundry model while continuing advanced-process and 300mm investment, but it does not reveal the exact May 15, 2002 board resolution or project-by-project release sequence [claim.tsmc.outcome.actual-action; assumption.tsmc.outcome.board-action-inference]. That distinction matters: realized spending can test the learner recommendation only imperfectly.
In 2002, consolidated capital expenditures were NT$55.236 billion, down from NT$70.201 billion in 2001 and below the approximately US$2.5 billion plan disclosed near the cutoff [fact.tsmc.outcome.capex.2002; evidence.tsmc.outcome.2002.capex; evidence.tsmc.cutoff.2002-q1.capex]. Year-end monthly capacity fell to 350,735 wafers from 381,044 as older 0.35- and 0.25-micron capacity was upgraded to 0.18- and 0.15-micron technology [fact.tsmc.outcome.monthly-capacity.2002; evidence.tsmc.outcome.2002.utilization-capacity]. Average utilization recovered from 51% in 2001 to 73% in 2002, but the quarterly path of 67%, 85%, 79% and 61% remained volatile [fact.tsmc.outcome.utilization.2002; evidence.tsmc.outcome.2002.utilization-quarterly].
The next two years show a staged aggregate path rather than a simple freeze or an uninterrupted spend ramp. Capex declined again to NT$37.871 billion in 2003 and then increased to NT$81.095 billion in 2004; Fab 14 began production in the fourth quarter of 2004 [claim.tsmc.outcome.capital-path; evidence.tsmc.outcome.2004.capex-path; evidence.tsmc.outcome.2004.fab14]. This sequence is consistent with technology-first, capacity-gated investment, but it cannot prove that interpretation. Equipment delivery timing, demand recovery or unobserved project economics could produce a similar aggregate pattern [judgment.tsmc.outcome.process-quality].
Near-term financial scorecard
Under consolidated ROC GAAP, net sales rose from NT$125.885 billion in 2001 to NT$162.301 billion in 2002 and net income from NT$14.483 billion to NT$21.610 billion [claim.tsmc.outcome.first-year-gating; table.tsmc.outcome.2002-scorecard; @src.tsmc.outcome.2002-20f]. Operating cash flow was NT$98.507 billion in 2002, exceeding the NT$55.236 billion of reported capex, but no free-cash-flow number is manufactured because the filing categories and maintenance/growth split are not reconstructed [fact.tsmc.outcome.ocf.2002; table.tsmc.outcome.capex-cash-flow].
By 2004, net sales were NT$257.213 billion, net income NT$92.316 billion and operating cash flow NT$153.151 billion [fact.tsmc.outcome.sales.2004; fact.tsmc.outcome.net-income.2004; fact.tsmc.outcome.ocf.2004; evidence.tsmc.outcome.2004.results; evidence.tsmc.outcome.2004.ocf]. Filing-defined utilization averaged 100%, versus 89% in 2003, although TSMC changed the calculation from the first quarter of 2003 to exclude engineering wafers and Vanguard figures [fact.tsmc.outcome.utilization.2004; evidence.tsmc.outcome.2004.utilization]. The utilization rebound is therefore economically relevant but not fully comparable to the cutoff series.
Long-run operating evidence
TSMC's as-reported nominal scale expanded materially at every selected horizon. Net revenue and net income were NT$419.538 billion and NT$161.605 billion in 2010, NT$843.497 billion and NT$302.851 billion in 2015, NT$1.339 trillion and NT$510.744 billion in 2020, and NT$2.894 trillion and NT$1.158 trillion in 2024 [claim.tsmc.outcome.long-run-scale; table.tsmc.outcome.longitudinal-results]. The underlying filings are the sources of record for those observations source · tsmc.outcome.2010-20f source · tsmc.outcome.2015-20f source · tsmc.outcome.2020-20f source · tsmc.outcome.2024-20f.
The physical and technology record moved in the same direction. TSMC reported approximately 9 million 300mm-equivalent wafers of annual capacity in 2015, approximately 13 million 12-inch-equivalent wafers in 2020 and approximately 17 million in 2024 [fact.tsmc.outcome.capacity.2015; fact.tsmc.outcome.capacity.2020; fact.tsmc.outcome.capacity.2024; table.tsmc.outcome.technology-capacity]. Advanced technologies defined as 16 nanometers and below supplied 58% of 2020 wafer revenue; technologies at 7 nanometers and below supplied 69% in 2024 [fact.tsmc.outcome.advanced-node-share.2020; fact.tsmc.outcome.advanced-node-share.2024]. Those thresholds are different and must not be treated as one unchanged KPI.
Capital intensity remained central. Reported capex was NT$186.944 billion in 2010, NT$257.517 billion in 2015, NT$507.239 billion in 2020 and NT$956.007 billion in 2024 [table.tsmc.outcome.capex-cash-flow]. Operating cash flow at the same horizons was NT$229.476 billion, NT$529.879 billion, NT$822.667 billion and NT$1.826 trillion [evidence.tsmc.outcome.2010.ocf; evidence.tsmc.outcome.2015.ocf; evidence.tsmc.outcome.2020.ocf; evidence.tsmc.outcome.2024.ocf]. This is evidence of internally generated funding capacity, not a deterministic finding that every incremental fab earned its cost of capital.
The platform became broader than fabrication
The cutoff packet already showed that “pure play” included design, mask, probing, test and assembly services rather than wafer fabrication alone [claim.tsmc.cutoff.dedicated-model; claim.tsmc.cutoff.service-breadth]. In 2008 TSMC described an Open Innovation Platform built on design-enabling blocks and interfaces shared with EDA and other ecosystem partners [claim.tsmc.outcome.ecosystem-expanded; evidence.tsmc.outcome.oip.ecosystem; @src.tsmc.outcome.oip-2008]. The supported interpretation is that TSMC extended a neutral customer platform without entering customers' branded-chip markets; the public packet does not quantify a trust premium or isolate the platform's return [assumption.tsmc.outcome.neutrality-value].
Neutrality also did not eliminate strategic competition. Intel's 2021 IDM 2.0 announcement combined internal manufacturing expansion with plans for a major foundry service [claim.tsmc.outcome.idm-rival; evidence.tsmc.outcome.intel.idm2; @src.tsmc.outcome.intel-idm2]. Integrated device manufacturers could remain customers, capacity alternatives and competitors at the same time.
Causal assessment and rival explanations
The primary hypothesis is that dedicated-foundry focus, design enablement and sustained process/yield investment contributed materially by reducing customer conflict, improving design-to-manufacturing execution and compounding technology learning at scale [hypothesis.tsmc.focus-enablement-and-technology]. That mechanism is coherent with the cutoff model, the later ecosystem record, advanced-node mix and physical capacity growth. Confidence remains moderate because the evidence is observational [judgment.tsmc.outcome.causal-contribution].
Demand is the strongest rival and complementary cause. Fabless outsourcing was already visible at the cutoff, while later smartphone, mobile and AI/HPC demand could not be forecast with precision in 2002. In 2024, TSMC attributed NT$542.122 billion of year-over-year revenue growth to high-performance computing, a 58% increase, and contemporaneous Reuters reporting tied its market-value milestone to demand for AI-capable chips [claim.tsmc.outcome.ai-demand; evidence.tsmc.outcome.2024.platform-mix; evidence.tsmc.outcome.reuters.ai-demand]. A favorable demand mix may explain a large share of the endpoint regardless of board-process quality [hypothesis.tsmc.external-demand-and-scale].
Policy, geopolitics and competitor execution form a second rival and complementary bundle. The U.S. government awarded up to US$6.6 billion of direct funding and offered up to US$5 billion of proposed loans for a planned Arizona investment above US$65 billion, with disbursement tied to milestones [claim.tsmc.outcome.policy-co-cause; evidence.tsmc.outcome.commerce.award; evidence.tsmc.outcome.commerce.loans; @src.tsmc.outcome.commerce-2024]. Those terms show that later scale and geography were affected by state support and supply-security objectives, not only the 2002 business-model choice [hypothesis.tsmc.policy-geopolitics-and-rival-execution].
The July 2024 ADR-based US$1 trillion market-value milestone is context, not a causal bridge [claim.tsmc.outcome.market-cap-milestone; evidence.tsmc.outcome.reuters.trillion; @src.tsmc.outcome.reuters-2024]. No capitalization-complete model in this episode reconciles shares, dividends, dilution, exchange rates, valuation multiples or later shocks. The case therefore abstains from attributing all shareholder value or return to the 2002 decision [judgment.tsmc.outcome.value-attribution].
Persistent risks and counterfactuals
The fixed-cost risk identified before the decision never disappeared. Both later filings and the cutoff filing state that owned-fab costs do not fall proportionally when demand or utilization declines [claim.tsmc.outcome.fixed-cost-persistence; evidence.tsmc.outcome.2010.fixed-costs; evidence.tsmc.outcome.2024.fixed-costs]. Customer concentration also remained material: the top ten represented 76% of 2024 net revenue and the largest customer 22% [claim.tsmc.outcome.customer-concentration; table.tsmc.outcome.customer-concentration]. Scale improved economics in favorable demand, but it did not repeal cyclicality, customer power or irreversibility.
The relevant alternatives remain instructive. Full retrenchment could have protected cash during the downturn but risked forfeiting process learning and later design wins [counterfactual.tsmc.retrench-and-harvest]. Faster unconditional expansion could have captured recovery earlier but enlarged fixed-cost downside if forecasts failed [counterfactual.tsmc.execute-plan-without-gates]. Entering proprietary branded chips might have captured downstream margin but introduced direct customer conflict [counterfactual.tsmc.enter-branded-chips]. Customer-financed capacity could have improved demand evidence but increased dependence on a small set of buyers [counterfactual.tsmc.customer-financed-capacity]. The packet lacks the contracts, yields, equipment penalties and project cash flows needed to assign any of these alternatives a fabricated return.
Learnings for an analysis agent
First, separate continuous technology learning from divisible volume capacity. In a long-lead, fixed-cost business, stopping both can destroy relevance while funding both without gates can destroy liquidity. Release volume modules only against stable-scope utilization, verified yield economics, diversified binding demand and a downside liquidity floor [rule.gate-irreversible-capacity-protect-technology-learning].
Second, platform neutrality can coexist with a richer service layer. Design tools, interfaces, libraries and ecosystem partners may deepen customer value without competing for the customer's branded economics [rule.extend-neutral-platform-without-competing-with-customers]. The rule remains a candidate because this case does not identify the size of a neutrality premium.
Third, preserve measurement scope before comparing periods. This series crosses ROC GAAP and IFRS, net-income attribution definitions, utilization calculations, wafer-equivalence bases and period-specific convenience translations [claim.tsmc.outcome.comparability-limits; evidence.tsmc.outcome.2015.ifrs-transition]. Capital expenditure is not total investing cash flow, and NT-dollar values are not silently converted with a single exchange rate [rule.preserve-operating-scope-before-longitudinal-comparison].
Finally, distinguish a good process from a good outcome. The later evidence makes this a major business success, but public filings cannot recover the 2002 board's actual information set, authorization sequence or rejected projects. The strongest reusable lesson is disciplined asymmetry: protect hard-to-rebuild technology and customer trust while making irreversible capacity earn its release.