Variant Perception

Figures converted from New Taiwan Dollars at historical FX rates — see data/company.json.fx_rates for the rate table. Ratios, margins, and multiples are unitless and unchanged.

Where We Disagree With the Market

The sharpest disagreement: consensus is underwriting TSMC's downcycle behaviour using the FY2023 stress test, but the company that printed a 42.6% trough operating margin in FY2023 no longer exists at today's customer-and-mix concentration. The market reads HPC at 61% of Q1 FY26 revenue as resilience because the last cycle barely scratched margins; we read it as the variable that will be the next cycle, because FY2023's protection came from HPC being secular while smartphone cycled — that arithmetic inverts when HPC is the cycle. A second disagreement: management has already disclosed an explicit 5–7 percentage point gross-margin headwind stack (N2 dilution + overseas-fab dilution widening through 2028) that the FY26 forward P/E of ~24× does not appear to compute. A third: the through-cycle 56%+ GM anchor was raised in 2026 — at peak — and the market is treating an uncalibrated commitment as a tested floor. Each disagreement is observable, narrow, and resolves on a published number.

Variant Perception Scorecard

The variant view scores in the high-medium band rather than the extreme because each disagreement is narrow and near-term resolvable, not a philosophical reframe. Consensus clarity is high (98% buy on 49 analysts; ~24× forward P/E; explicit FY26 USD growth guide "above 30%"), and the evidence we disagree on is published — management's own dilution math, the company's own customer-mix progression, and the company's own raised 56%+ through-cycle anchor. There is no edge in claiming AI is real or the moat is wide — consensus already prices both — so this page does not argue either. The edge sits in three specific places where the upcycle math and the downcycle math are being treated as the same calculation.

Consensus Map

The map exists so that the disagreements that follow are not arguments with shadows. The market view on every row above is documented in a sell-side rating distribution, a price reaction, a guidance number, or a directly-quoted research note. Where the market is mixed — geopolitical risk pricing in row 5, dilution-stack absorbability in row 6 — confidence in the consensus is marked accordingly, and we narrow the disagreement to the most observable parts.

The Disagreement Ledger

Disagreement 1 — the FY23 protection cannot be assumed forward. Consensus says FY23 proved TSMC is non-cyclical because revenue fell only 4.5% and operating margin held above peers' best ever. That observation is real, but it was generated by a foundry whose HPC mix was ~43% and whose Top-2 customer share was ~26%; Q1 FY26 mix is 61% HPC and Top-2 is ~36%. The protective dynamic in FY23 was that HPC was secular while smartphone (then 39%) cycled; in FY26 HPC is the largest segment of revenue and AI accelerator capex is the cycle itself. If hyperscaler capex digests even one normal cycle, the inventory-correction playbook does not protect — and the FY26 capex bill of $52-56B is sized to AI continuation. The market would have to concede that the right downside math for the new TSMC is something it has never observed. The cleanest disconfirming signal is two consecutive quarters of aggregate MSFT+GOOG+META+AMZN+ORCL capex sustaining 25%+ YoY while TSMC GM holds ≥62%; either condition failing flips weight to the variant.

Disagreement 2 — the dilution stack consensus has not absorbed. This one does not require any new event — every input is already disclosed by management. N2 ramp dilution at 2-3pp through 2026; overseas-fab dilution 2-3pp early widening to 3-4pp later through 2028; possible material/gas cost. Stacked, this is a 5-7pp peak GM headwind that lands across FY26-FY28, and Q1 FY26's 66.2% print absorbed only the first portion. Consensus forward P/E of ~24× implies flat-to-modestly-declining margins; if mgmt's own dilution math plays through and revenue stays at consensus levels, FY26 EPS comes in below the $12-13 per ADR consensus range and FY27 below that. The market would have to concede that what it is reading as "Q1 was the regime" was actually "Q1 was the start of the absorption window." Cleanest disconfirm: Q2 FY26 GM prints ≥66% with HPC mix ≥60% and management commentary explicitly keeps the dilution stack inside the 2-3pp range through 2028.

Disagreement 3 — the raised through-cycle anchor is uncalibrated. This is the subtlest. Management raised the through-cycle GM anchor from 53% to 56%+ in 2026 — at peak conditions, after Q1 FY26's 66.2%. The prior 53% anchor was tested in FY23 (cycle trough GM 54.4%). The new 56% anchor has never been tested in the AI-mix regime, because the AI-mix regime has not seen a downcycle. The market is reading the 10pp gap between Q1 GM and the new anchor as conservatism; we read it as a commitment without observation. If the next correction breaks below 56% GM at >85% utilisation, the market loses both the print and the credibility on every other raised anchor (FY24-29 USD CAGR raised from 15-20% to ~25%; 2030 TAM raised $1T → $1.5T) — a compound de-rating. Cleanest disconfirm: the first real cycle test post-anchor raise bottoms with GM >56% on healthy utilisation.

Evidence That Changes the Odds

How This Gets Resolved

Each signal above is observable in a published filing, an earnings call, a balance sheet line, or a public tape-out announcement. Two clusters of timing matter. The near-term cluster (Q2 FY26 print, hyperscaler Q2 commentary, May/June monthlies) resolves Disagreement 2 and starts tightening Disagreement 1 within ~90 days. The structural cluster (first real cycle test of the new GM regime, flagship multi-sourcing decisions, Samsung 2nm with a named customer) resolves Disagreements 1 and 3 across 4-12 quarters and is the harder underwrite — but is also where the asymmetry of being wrong is largest.

What Would Make Us Wrong

The cleanest case against this variant view starts with the company's own execution record. Management has hit 12 of 13 valuation-relevant promises since 2018 and raised through-cycle anchors twice on hard evidence rather than aspiration. If the raised 56%+ GM anchor and the ~25% USD revenue CAGR are reliable — and the historical hit rate says they probably are — then Disagreement 3 will resolve consensus-side at the first stress test, and the read on Disagreement 2 collapses because mix-and-pricing (the offset mgmt has been demonstrating quarter after quarter) will close the dilution gap on its own. A clean Q2 FY26 print at GM ≥66% with HPC mix ≥60% and unchanged dilution commentary would be one piece of evidence that this variant view is overreaching.

Disagreement 1 — the FY23-mis-applied argument — has a real counter that we should be honest about. The asymmetry that protected TSMC in FY23 was HPC being secular while smartphone cycled. If hyperscaler AI capex turns out to genuinely not be a cycle — if inference-layer silicon migration sustains AI accelerator demand at a structurally faster rate than the historical end-market cycle, as TSMC's own raised 2030 TAM to $1.5T implies — then "HPC is now the cycle" never tests, and the FY23 protection generalises. The cleanest evidence for this counter is what we are already seeing: capacity-bound rather than demand-bound supply chain (Broadcom March 2026, Google TPU cut), multi-year customer wafer commitments at N2/A16, CoWoS sold out through 2026. The variant requires AI capex to actually cycle; if it doesn't, the variant is wrong.

The third concession is on positioning. Crowded long positioning has been a feature of high-quality compounders for cycles before any unwind — Microsoft and Apple traded with similar consensus distributions for years. The "98% buy on 49 analysts" observation is real but not predictive on its own; it amplifies any narrative crack but does not create one. If the narrative cracks we identify (Q2 GM walk-down, hyperscaler capex digestion, Samsung HPC win) don't materialise, positioning alone won't move the stock.

The forth concession — and the largest open question on the variant case — is the implicit timing assumption. Disagreements 1 and 3 only resolve when a cycle arrives. If a cycle is years away (i.e., AI capex compounds through FY30 without a digestion quarter), the variant is right in theory but wrong in payoff for institutional money. Holding a variant view that resolves on a multi-year arc is not the same as making money on it. The disagreement on the dilution stack (Disagreement 2) is the only one with a near-term forcing event, which is why it leads.

The first thing to watch is Q2 FY26 gross margin and the explicit Q3/Q4 dilution walk — that single July print is what tells you whether Disagreement 2 has begun to land, and the dilution-walk commentary is what tells you whether the market and management are reading the dilution stack the same way.