Business

Know the Business

TSMC is what happens when the most cyclical industry in capitalism develops a near-monopoly inside it. The economic engine is per-wafer pricing on leading-edge process nodes that nobody else can produce at scale — Apple, Nvidia, AMD, Broadcom, and Qualcomm have no second source at 3nm and below, and that gives one company the right to set price for the most valuable real estate in computing. The market's frequent mistake is reading the headline 33% capex-to-revenue as a capital-destruction story; the numbers actually show a fixed-cost asset base earning 44% net margins, 35% ROE, and ~44% ROIC because the leading node has no economic substitute.

1. How This Business Actually Works

TSMC sells one thing: contracted fab capacity, priced per 12-inch-equivalent wafer, multiplied by node sophistication. Customers hand over a design; TSMC runs the wafer through several hundred process steps over weeks; TSMC bills per wafer started. Revenue this year was NT$3.81T on roughly 16 million wafer-equivalents — the unit is mundane; the economics are not.

The shape of the P&L flows mechanically from three levers:

Two readings flow from that table. First, gross margin is structurally a node-mix story, not a volume story. The 16 points of gross-margin expansion from FY2014 (49.5%) to FY2025 (59.9%) happened on revenue that grew 5x, but the lift came mostly from migrating wafer revenue up the node ladder — every move from 16nm to 7nm to 5nm to 3nm added pricing power that did not exist before. Q1 2026's 66.2% gross margin is the cleanest read on this: when the leading edge tightens and 3nm fills, the same fab footprint earns dramatically more.

Second, the capital intensity is the moat, not a weakness. TSMC spent NT$1.27T (33% of revenue) on capex in FY2025 — larger than the combined annual revenue of UMC, GFS, and SMIC. That spend funds the next two nodes, and only one other company on earth (Samsung) can match it. Intel is trying with NT$471B of FY2025 capex but losing money in foundry while doing so. The capex bill is the cost of the monopoly.

The same fab base used to be a smartphone-and-PC business with cyclical handset orders. As of Q1 2026, 61% of revenue is HPC — AI accelerators (Nvidia, AMD, Broadcom ASICs, Google TPU), CPUs (AMD, Intel server, Apple Silicon), and data-center networking. That re-platforming is what's defending margins against the historical handset-driven cycle.

2. The Playing Field

The foundry industry has one leader and a long tail. The peer table understates how lopsided that actually is — TSMC's FY2025 revenue is roughly 10x Samsung Foundry's, 13x SMIC's, and 16x UMC's. More striking, TSMC's operating margin exceeds every other foundry's gross margin.

The bubble chart frames the moat. TSMC is alone in the top-right quadrant — and the gap is not narrow. GlobalFoundries and UMC, the two well-run pure-play specialty foundries, earn one-third of TSMC's operating margin on one-twentieth the revenue. Intel runs the only other fab base of remotely comparable scale (US$210B of assets, US$15B of capex) and earns negative operating margin. SMIC has growth (FY2025 +21% revenue) but the gross margin is 21% — what a foundry earns when it can't price the leading edge.

The instructive comparison is not "TSMC vs Samsung Foundry" — it's TSMC vs Intel Foundry. Both compete for advanced-node logic customers. TSMC books US$50B+ of operating profit on that base; Intel loses US$13B trying to be a credible alternative. That gap is what a real-world leading-edge monopoly looks like in P&L terms.

3. Is This Business Cyclical?

TSMC has historically been cyclical. As of 2025–2026, the cycle has been partly converted into a secular — and that conversion depends on AI demand sustaining. The cycle hits foundries through utilization → ASP → margin, in that order, and three of TSMC's last four downturns were severe enough to compress operating margin by 7–10 points.

The chart tells two stories. The first is the classical foundry cycle: 2001–02 (tech bust dropped op margin to 21%), 2008–09 (revenue declined 11%, op margin held near 31% only because TSMC cut capex), and 2018–19 (handset glut took op margin to 34.8% on 3.7% revenue growth). Each lasted four to six quarters and reversed when handset and inventory cycles cleared.

The second story is what AI did to the cycle baseline. The 2023 correction was sharp at SMIC, UMC, GFS, and on TSMC's mature nodes — but TSMC's advanced-node revenue grew that year, and full-year operating margin still held at 42.6%, well above prior trough margins. FY2024 and FY2025 then ran 45.7% and 50.8% — territory only briefly visited in FY2022's peak. The structural read: a foundry whose advanced-node revenue is 74% of total has converted ~40% of the cycle into a secular. The remaining ~30% — mature node, handset, IoT — still cycles, but those segments are now too small to dominate the P&L.

The risk to this benign read is concentrated and identifiable: if hyperscaler AI capex flattens or rolls over and TSMC has built capacity for it, the company would re-enter a classical foundry downcycle with much higher capex commitments and lower depreciation cover. Management's guidance to spend US$52–56B in FY2026 — at the top of historical absolute capex — sizes the bet on AI demand continuing. Investors who think of TSMC as a non-cyclical compounder are right if AI demand stays where it is, and wrong by 15+ points of operating margin if it does not.

4. The Metrics That Actually Matter

Five numbers explain TSMC. Everything else is a derivative. Watching headline P/E or EPS without these will mislead in both directions.

Two readings sit in those tables. The dominant one: on the metrics that actually drive foundry value — node mix, gross margin, ROIC, asset productivity — TSMC isn't a leader, it's an outlier. No other foundry in the world earns positive return on its capital base after capex; TSMC earns 44%. The second reading is more subtle: capex/revenue is not the right comparison number. SMIC at 71% capex/revenue is building capacity it cannot earn returns on (4% ROE); TSMC at 33% capex/revenue is spending US$52–56B in FY2026 to defend a 44% ROIC base. The two numbers say opposite things despite looking similar.

The usual ratios — P/E, P/B, dividend yield — describe the shareholder claim on this engine but not the engine itself. P/E will move with cycle psychology far more than with the underlying margin trajectory. Watch the five operating metrics first; the multiple will follow.

5. What Is This Business Worth?

TSMC is best valued as one economic engine, not a sum of parts. The segments TSMC discloses (HPC, smartphone, IoT, auto, DCE) share the same fab base, the same depreciation pool, and the same leading-edge process technology — they don't have separable economics, and there are no listed subsidiaries or stakes that materially distort the picture. Value is determined by node-mix progression, through-cycle gross margin, ROIC, and the durability of the leading-edge monopoly. Forced SOTP would add false precision.

The valuation conversation reduces to four numbers: through-cycle gross margin, through-cycle ROE, capex intensity, and reinvestment runway. Management's published targets — 56%+ through-cycle gross margin, high-20s ROE through cycle, FY2024–FY2029 revenue CAGR ~25% in US$ — are the right anchor. Today's gross margin at 59.9% and ROE at 35% sit clearly above through-cycle, which is what should be true near a cycle peak.

Two takeaways from cash returns. FCF / net income runs only 50–60% over the cycle — this is a capex-heavy compounder, not a cash-rich compounder, and any FCF-based valuation needs a through-cycle assumption rather than a trailing one. Dividends grow but absorb only ~28% of net income; buybacks are negligible. Capital allocation is overwhelmingly reinvestment, and that is the correct decision so long as ROIC stays where it is. An owner of TSMC is implicitly betting on management continuing to recycle cash into the leading edge at 30%+ marginal ROIC. If that opportunity disappears (Samsung catches up, demand flattens), shareholders should expect the policy mix to swing toward higher dividends and buybacks. The current mix presumes the runway.

6. What I'd Tell a Young Analyst

Watch the leading edge, not the headline. Quarterly platform mix, the share of wafer revenue from ≤7nm, and CoWoS packaging commentary will tell you what's happening to margin three to four quarters before P&L confirmation. Q1 2026's 66.2% gross margin was visible in advance from 3nm ramp commentary in Q3 2025.

Read TSMC's capex like an option premium, not an expense. US$52–56B of FY2026 capex pre-loads depreciation but also pre-loads three years of capacity at the price-setting node. The market periodically prices the capex as a cost; through the cycle it has been the cheapest way to buy the next leg of the monopoly. The real risk to watch is not capex level — it's capex productivity (Rev/Net PP&E), which would fall first if capacity were being built ahead of demand.

Calibrate against Samsung 2nm and Intel 18A yield commentary. TSMC's structural advantage is not "Taiwan" or "scale" — it's a multi-quarter yield lead on whichever node Apple, Nvidia, and AMD designed for next. The moment Samsung Foundry or Intel Foundry can hold competitive yield at 2nm or A16, the customer concentration that makes TSMC's pricing power durable starts to fragment. Today neither rival is close — Samsung Foundry's foundry revenue actually shrank 3.9% in FY2025 and Intel Foundry posted a US$13B operating loss — but this is the single technical variable that would force a real rerating.

Don't confuse "deeply cyclical" with "cyclical." The 2023 mini-cycle showed TSMC can grow earnings and hold 42.6% operating margin even as UMC, GFS, and SMIC suffered double-digit revenue declines. That's the AI-mix moat in action. The mistake on the bear side is assuming the 2018–19 or 2008–09 cycle pattern still applies one-for-one; the mistake on the bull side is forgetting that 36% of revenue still cycles like a normal foundry.

Don't underwrite this stock on tactical FX, monthly revenue prints, or quarterly margin oscillations. They move the tape but not the value. The four things that genuinely change the thesis are: Samsung 2nm yield, hyperscaler AI capex trajectory, Taiwan geopolitical state, and management's pricing posture on the next two nodes. If those four are unchanged, the underwriting case is unchanged.