Taiwan Semiconductor (TSM)

TSMC is the single un-substitutable toll-booth on AI compute, and Q2 FY2026 is the quarter where that translated into hard numbers: revenue at the top of guidance, a record 67.7% gross margin above the guided range, an operating margin 180bps past the high end, and a full-year outlook raised from 'above 30%' to 'slightly above 40%'. The company is self-funding a US$60-64B capex program out of cash flow while still carrying ~US$83B of net cash.

• Guidance was raised, not merely met - full-year 2026 USD growth went from 'above 30%' (Q1) to 'slightly above 40%' (Q2), and Q3 is guided to US$44.6-45.8B, a further ~12% sequential step. That is acceleration inside an already-record year, and it is management's own filed number, not a sell-side hope.
• Margins are expanding while the mix shifts to the highest-value work: gross margin 67.7% and operating margin 60.3% both cleared the top of guidance, with HPC now 66% of revenue (+20% QoQ) and advanced nodes 77% of wafer revenue. Pricing and mix are outrunning the cost of the ramp.
• The capex raise is a demand signal, not a distress signal - US$60-64B (up from $52-56B) with 70-80% to advanced process, plus an additional US$100B Arizona commitment, is management underwriting multi-year visibility with its own balance sheet. It is funded from a NT$783B quarterly operating cash flow, so FCF stayed positive (NT$287B) even in the heaviest capex quarter.
• Balance sheet removes financing risk from the thesis: NT$3,518B cash & marketable securities against NT$864B long-term interest-bearing debt, equity at 69.1% of assets. TSMC can spend through a demand air-pocket without a capital raise - very few capital-intensity-40% businesses can say that.
• 2nm is only 3% of wafer revenue and already contributing, with the steep ramp still ahead. The node that carries the next ASP step is in front of the numbers, not behind them, and the margin dilution it causes (~3-4pts) is a known, quantified, temporary cost.

The quarter was excellent and the stock still fell from its late-July highs, which is the tell: at ~$2.1T with 18.6x forward earnings, TSM is priced for the AI capex cycle continuing, while the company is simultaneously raising capex intensity toward ~38% of revenue and guiding margins DOWN. The concentration risks - one end-market, two customers, one island - are all getting worse, not better, and none of them are hedgeable.

• Margins have peaked for now by management's own guide: gross margin goes 67.7% -> 65-67% and operating margin 60.3% -> 56-58% in Q3, with ~3-4pts of 2nm dilution in 2H26 and overseas-fab dilution guided to widen from 2-3pts to 3-4pts. The record print is the near-term high-water mark, and the market has already seen it.
• The capex raise cuts both ways: US$60-64B (up 15% from $52-56B) plus US$100B for Arizona is ~38% of revenue in fixed cost that depreciates for a decade regardless of demand. FCF already fell QoQ from NT$348B to NT$287B despite a record-margin quarter - the FCF conversion of this cycle is structurally worse than the P&L suggests.
• Single-end-market concentration is now extreme: HPC is 66% of revenue and grew +20% QoQ while smartphone actually declined -4% QoQ. The non-AI pillar is shrinking, so any hyperscaler digestion hits the mix with nothing to offset it.
• Customer concentration is worsening on the filed data: the largest customer went from 12% (2024) to 19% (2025) of net revenue, the second 12% -> 17%, and the top ten 76% -> 78%. Two customers are now ~36% of the book, and both design AI/mobile silicon into the same cycle.
• The headline earnings quality is weaker than +77.4% suggests - net income includes a one-time gain on the Vanguard stake sale (~NT$63B per press reports against NT$95.8B of derived net non-operating income). Operating income grew +65.4%, and the excess is non-recurring.
• The un-underwritable tail is unchanged: the crown-jewel leading-edge capacity is in Taiwan, and the Arizona build is partly a political cost (dilutive by management's own numbers) rather than an ROI-optimal one. No multiple can price a China-Taiwan contingency, which is exactly why a monopoly with 60% operating margins trades at 18.6x forward.