The Guide Leaps From 30% to 40%+ and CapEx to $64B: The Tape Sold the Capital Bill, Not the Business — Maintain Outperform

Results vs. Consensus

Q2 2026 Scorecard

Year-Over-Year Comparisons

Quarter-Over-Quarter Comparisons

Quality of Beat

Revenue: The +0.4% headline beat versus consensus is small by design. TSMC discloses revenue monthly, so by the time the quarter prints, the top line is effectively known. The more informative fact is that $40.20B finished at the absolute top of the $39.0–40.2B guide range, the third consecutive quarter of landing at or above the guide ceiling, and grew 12.0% sequentially in a quarter that carries no A-series ramp. Sequential double-digit growth at a $40B quarterly scale, driven by HPC rather than seasonality, is the signal.

Margins: These are the cleanest numbers in the report. The 67.7% gross margin (up 150bp sequentially) came, per CFO Huang, from cost-improvement efforts and slightly higher utilization, with the record 77% advanced-node mix doing the structural work. The 60.3% operating margin is the standout: crossing 60% for the first time reflects operating leverage on a revenue base that grew 12% sequentially while operating expense grew far slower. Neither figure depends on FX or one-time items; both are operational and, on the Q3 guide, largely sustainable.

EPS: Here the analyst has to separate signal from flatter. Operating income grew 16.3% sequentially, in line with the operating-margin expansion. Net income grew 23.4% sequentially, and EPS the same. The seven-point gap between operating-income growth and net-income growth is almost entirely the non-operating line: NT$95.8B this quarter versus a ~NT$29B norm. That is a real number, TSMC does earn substantial interest on NT$3.4T of cash, but it is lumpy and partly FX-linked, and it is not the same quality of earnings as a wafer sold at 67.7% gross margin. The right way to read the quarter is: operations delivered a ceiling-beating result, and non-operating income turned a strong EPS into a spectacular one. We anchor our forward estimates to the operating result and treat the non-operating surge as non-recurring.

Segment Performance

Platform Revenue Mix — Q2 2026

HPC / AI — 66% of Revenue, a New Record

HPC rose to 66% of revenue from 61% in Q1, growing 20% sequentially in a quarter where total revenue grew 12%. The arithmetic is the same as last quarter, only more extreme: HPC contributed roughly $4.4B of the quarter's incremental revenue while the rest of the portfolio was collectively flat. Two years ago mobile held plurality; today AI accelerators, data-center GPUs, custom silicon, and now data-center CPUs are two-thirds of the company and still the fastest-growing piece. The mix shift is structural and, in the near term, one-directional: incremental advanced-node capacity is being allocated almost entirely to HPC customers who engage two to three years ahead of production.

Assessment: At 66%, HPC is now the business. The relevant question is no longer whether AI is a real demand driver but whether TSMC can add capacity fast enough to serve it, which is exactly the question the capex raise answers. Watch whether HPC holds 65%-plus into Q3 as the smartphone A-series ramp begins to reload the mix.

Automotive — The Correction Finally Turns

Automotive grew 15% sequentially, a genuine inflection after Q1's −7% and roughly two years of inventory digestion. At 4% of revenue it remains immaterial to the model, but the direction matters: it removes a small persistent drag and hints that the broader analog and specialty-node inventory correction, which has weighed on TSMC's automotive-exposed customers, is ending. Some of the strength is AI-adjacent (power-management and sensor content for the edge), which management flagged separately as a mature-node shortage area.

Assessment: A positive at the margin, not a thesis mover. If the automotive and industrial recovery broadens into H2, it is additive upside to a model that assumes essentially nothing from these segments.

Smartphone — Seasonal, Not Structural

Smartphone declined 4% sequentially to 22% of revenue, standard for a June quarter that precedes the flagship A-series ramp. There is no evidence of competitive displacement; Apple's leading-edge silicon remains on TSMC's most advanced nodes, and the Q3 guide implicitly assumes the seasonal reload. Smartphone's shrinking share of the mix is a function of HPC's growth, not smartphone weakness.

Assessment: Unremarkable, which is the point. Smartphone is now a stable, high-margin base on which the HPC growth layer is stacked.

Technology Node Revenue Mix — Q2 2026

The record 77% advanced-node mix is the direct cause of the record gross margin. Two structural shifts happened this quarter: N2 debuted at 3% of wafer revenue, beginning its dilution phase but also validating the fastest node ramp in TSMC's history, and N3 stepped from 25% to 30% of the mix, moving toward its H2 2026 crossover above the corporate-average gross margin. The near-term margin choreography is a tug-of-war: N3 maturation is a tailwind, N2 ramp (guided to a 3–4 point gross-margin dilution over the coming quarters) is a headwind. That the company still guides Q3 gross margin at 65–67% while absorbing steepening N2 dilution is the quiet bullish tell.

Key Topics & Management Commentary

Overall Management Tone: Management was the most expansive it has been in our coverage, but the confidence is now spent on capital rather than caution. Where Q1's defining act was a demand characterization (agentic AI as a token step-up), Q2's is a capital commitment (a ten-point guide raise, a $60–64B budget, and $100B more for Arizona), delivered as routine fact rather than as a pitch. Wei repeatedly declined to quantify multi-year growth or AI CAGR, offering instead that the demand is "stronger than what we said before" and that the industry is "a kind of a new industry," which reads as visibility outrunning the company's willingness to pin a number to it. The only defensive register on the call was around competition and customer concentration, and even there the posture was dismissive rather than hedged.

1. The FY26 Guide Leaps From ">30%" to "Slightly Above 40%"

"We now expect our full year 2026 revenue growth to be slightly above 40% year-over-year." — C.C. Wei, Chairman & CEO

This is the headline of the quarter. At Q1, management guided FY26 to ">30%" USD growth; three months later they raised it to "slightly above 40%." A ten-point mid-year revision to a full-year growth outlook, on a business that will do more than $170B of revenue, is extraordinary, and it is not a mechanical catch-up from a strong H1. With H1 at ~$76.1B (Q1 $35.9B + Q2 $40.2B) and the Q3 guide midpoint at $45.2B, "slightly above 40%" on the ~$122.4B FY25 base implies roughly $172–175B, requiring an H2 that steps up sequentially into the high-$40sB per quarter. That is a demand statement, not a seasonality statement.

Assessment: The guide raise is the number that matters most for near-term earnings, because it re-rates the base on which H2 estimates are built. It also reframes how to read the capex raise below: TSMC is not spending more against a hope, it is spending more against a full-year outlook it just lifted by ten points. We move our FY26 revenue estimate to $172–176B.

2. CapEx Raised to $60–64B, With "Significantly Higher" Ahead

"We have decided to raise our full year 2026 capital budget to be between $60 billion and $64 billion." — Wendell Huang, CFO

The 2026 capital budget rose to $60–64B, roughly $8–10B above the ~$55–56B "toward the high end" framing at Q1 and up ~50% over FY25's ~$40.9B. Wei added that capex over the next three years "will be even more significantly higher than the past three years," a sharpening of the Q1 language that already implied a multi-year elevated regime. Pressed on what changed in three months, Wei gave two reasons: demand "continue to increase, and we feel the pressure from the customer" to add capacity, and tool-price inflation. The first is the important one; the second is a real but secondary cost-push effect.

This is the disclosure the market punished, and it is worth being precise about why the market is wrong to read it as bearish. A capex raise is bearish when it is speculative, front-running demand that may not arrive. TSMC's raise is the opposite: it is demand-pulled, sized against a full-year revenue outlook the company simultaneously lifted, and constrained on the supply side by equipment availability rather than by any doubt about customer orders. The capital-intensity ratio does rise (spending ~$62B to produce ~$174B of revenue is ~36% of sales, up from prior years), and that pressures near-term free cash flow and lifts future depreciation. But it is the signature of a supply-constrained monopoly investing into contracted demand, not of a company chasing a bubble.

Assessment: The capex raise is a bullish demand signal wearing a bearish costume. It raises FY26–FY27 depreciation and compresses near-term FCF, which is a legitimate model input, but it is underwritten by a raised revenue guide and customer pressure to build. We would fade the capex-driven selloff.

3. An Additional $100B for Arizona — Roughly Four More Fabs

"We would like to announce an additional $100 billion investment in Arizona [for] 2 nanometer and below technologies… it probably additional four more fabs will be built." — C.C. Wei, Chairman & CEO

On top of the 2026 budget, TSMC announced a further $100B for Arizona, taking the total announced US commitment to roughly $265B and adding approximately four fabs on 2nm-and-below process technology. Management was explicit that the schedule "will depend on the market situation" and declined a firm timeline, but the direction is a material deepening of the US leading-edge footprint: Arizona moves from an N4/N3 outpost to a 2nm-and-below cluster. The strategic logic is twofold. It scales the geopolitical hedge, putting leading-edge capacity inside the US tariff and policy perimeter, and it responds to the customer and policy pressure to onshore advanced manufacturing.

Assessment: Positive for the multi-year thesis, incrementally dilutive to margin (overseas fabs remain a 2–4 point gross-margin headwind that widens as more ramp), and de-risking on the geopolitical tail. The market lumped this into the "too much capex" reaction; we read the willingness to put 2nm in Arizona as evidence that customer demand for US-made leading-edge is now firm enough to underwrite the premium.

4. Agentic AI, Reframed as a CPU Resurgence

"The emergence of agentic AI is leading to a resurgence in the role of CPUs in AI data centers." — C.C. Wei, Chairman & CEO

Q1 introduced agentic AI as "another step up in the amount of tokens being consumed." Q2 advances the thesis one layer: agentic workloads, which orchestrate multi-step tasks and coordinate sub-agents, are reviving demand for data-center CPUs alongside GPUs and custom accelerators. This matters for TSMC specifically because it widens the silicon footprint of an AI data center beyond the accelerator. The Q1 recap flagged that TSMC excludes data-center CPUs from its AI-accelerator revenue definition (because it cannot cleanly attribute CPUs to data center versus PC); the agentic-CPU dynamic makes that excluded category a growth vector, which means TSMC's reported AI-accelerator revenue understates the total AI-driven silicon demand it is capturing.

Asked to size the GPU-versus-CPU-versus-XPU split, Wei declined a number but noted "all of them are in TSMC" on the same leading-edge nodes, and that the company is allocating wafer supply to balance the mix. Asked directly whether the AI accelerator CAGR is being revised, he would only say it is "stronger and stronger" and "stronger than what we said before," declining to update the 54–56% framework with a new figure.

Assessment: The CPU-resurgence framing is a genuine incremental positive: it broadens AI demand across more of the compute stack, all of which lands on TSMC's leading edge. Wei's refusal to put a number on the AI CAGR ("we don't know how to answer this question" because it keeps rising) is, in context, a stronger signal than a specific figure would be. It says the demand curve is still steepening.

5. Gross Margin 67.7% and the First 60% Operating Margin

"[Gross margin was] 67.7%, up 150 basis points sequentially… primarily due to cost improvement efforts and slightly higher overall capacity utilization." — Wendell Huang, CFO

Gross margin cleared the guidance ceiling for a third straight quarter, and operating margin crossed 60% for the first time in company history. The drivers are the record 77% advanced-node mix, cost-improvement efforts at N3 and N5, and near-full utilization on AI-driven demand. The forward tension is N2: management guided that the steep 2nm ramp will dilute gross margin by 3–4 points in the coming quarters, and the Q3 guide (65–67% gross margin) already reflects the beginning of that. The bullish read is that N3 maturation and utilization are strong enough to hold the guided gross margin in the mid-60s even as N2 dilution steepens.

Assessment: The through-cycle margin framework keeps getting repriced upward by the actuals. TSMC's stated long-term target remains "56% and higher through the cycle," but a business printing 67.7% with a record advanced mix, confirmed pricing power, and N3 about to turn accretive has a through-cycle floor closer to the low-60s than to 56%. N2 dilution is the honest offset that keeps this from being a straight line up.

6. N2 (2nm): The Fastest Ramp in Company History

N2 entered the revenue mix at 3% of wafer revenue this quarter, its first meaningful contribution, and CFO Huang framed the "steep ramp-up" as the reason for the coming 3–4 point gross-margin dilution. Asked about the 2nm-family capacity CAGR (a ~70% figure through 2028 shown at the Technology Symposium), Wei's answer was telling in its brevity: "Okay, now is bigger. That's all I say." N2 carries roughly a 50% ASP premium over N3, so even a low single-digit share of wafer revenue represents a large dollar contribution, and the dilution it causes today is the leading edge of a multi-year accretive tailwind once yields mature (the pattern N3 is completing now).

Assessment: N2 is on its best possible trajectory: a fast ramp, a confirmed premium, and a capacity plan management now says is "bigger" than the symposium roadmap. The dilution is the cost of admission to a node that will be a margin compounder in 2027–2028. The competitive gap remains wide: no rival has a commercially proven 2nm alternative at scale.

7. The Supply-Demand Gap Is "Very Big," and Persists Toward 2029–2030

"I believe from this day on all the way to probably 2029, 2030, the demand is very strong… The trend is so robust that I believe we are witnessing a kind of a new industry." — C.C. Wei, Chairman & CEO

Pressed repeatedly on when supply catches demand, Wei extended the tight-supply window toward the end of the decade and, asked whether the unconstrained demand for 3nm-and-below is 30–50% above capacity or larger, replied only that "the gap is very big." He declined to quantify it, but the refusal, paired with the capex raise and the "new industry" framing, is itself the disclosure: TSMC is telling the market that even after the largest capacity build in its history, leading-edge demand outruns supply for years.

Assessment: A supply gap that persists to 2029–2030 underwrites both pricing power and high utilization for the model's entire forecast horizon. This is the deepest structural support for the margin and revenue trajectory, and it is the reason the capex raise should be read as a revenue-enabling investment rather than a cost.

8. Advanced Packaging: Capacity So Tight It Is "Limiting Customers' Growth"

Advanced packaging drew more Q&A attention than any topic except capex. Wei said bluntly that TSMC's back-end (CoWoS) capacity "is so tight that now it's limiting my customers' growth," and, in an unusual move, welcomed competitors' packaging capacity (Intel's EMIB-T was named) as a release valve: if a customer's chip is packaged elsewhere, TSMC still captures the front-end wafer, which is the larger part of the value. Management also updated the roadmap: the COUPE silicon-photonics platform is entering production and expected to scale as data centers demand lower power and higher interconnect bandwidth, and a glass-substrate pilot line is roughly a year from production maturity as a lower-cost CoWoS alternative.

Assessment: Packaging is now a binding constraint on AI system output, which is bullish for pricing and for the front-end wafer business it gates. Welcoming competitor packaging capacity is a confident, rational stance: TSMC would rather see its wafers shipped than hold them hostage to its own back-end bottleneck. COUPE and glass substrates are the next packaging legs to watch in 2027.

9. Competition: "No Shortcut," and It Takes About Five Years

"Choosing a technology, ramping it up, is not buying a milk from 7-Eleven… It is no shortcut… That's why I say it takes about five years." — C.C. Wei, Chairman & CEO

Asked about Samsung Foundry (flush with memory profits), Intel (with US policy support), the Terafab initiative, and ASML's just-announced 2028 EUV capacity expansion that could let rivals buy more tools, Wei's answer was structural: leading-edge foundry is a five-year build of technology, manufacturing excellence, customer trust, and service, and additional EUV tools do not compress that. He extended the same "no shortcut" logic he used on Terafab at Q1 to the entire competitive set.

Assessment: The competitive moat commentary is unchanged and, if anything, reinforced by rivals having to fight across N2, N3 capacity, and advanced packaging simultaneously. ASML expanding EUV supply is a mild long-term watch item (it lowers the tool-access barrier for competitors) but does not touch the yield, ecosystem, and trust barriers that take years to build. Not a near-term thesis risk.

10. Mature Nodes: Only the AI-Adjacent Corners Are Tight

Management drew a sharp line through the mature-node market. The AI-adjacent pieces (power-management ICs for data centers, sensors) are in genuine shortage; the consumer-driven mature nodes remain soft. This bifurcation matters because it explains the automotive and IoT sequential recovery (AI-adjacent content) without implying a broad consumer rebound. It also flags a second-order AI beneficiary: every AI data center needs large volumes of power-management silicon on trailing-edge nodes, a demand pool the market underweights when it focuses only on leading-edge accelerators.

Assessment: A useful nuance rather than a needle-mover. It confirms that TSMC's mature-node strength is AI-derived, not a sign of a cyclical consumer recovery, and it adds a small, stable demand leg (power management, sensors) tied to the same AI build-out.

Guidance & Outlook

The Q3 guide is the operational headline beneath the full-year raise. A $45.2B midpoint is +12.4% sequential on an already-record Q2, and +37% year-over-year, with CFO Huang citing "the steep ramp-up of our 2-nanometer technology" as a Q3 driver. The gross-margin guide steps down to 65–67% (from 67.7% actual) and operating margin to 56–58% (from 60.3%): part of that is real N2 dilution, and part is simply that Q2's non-operating windfall will not repeat in the operating-margin line. Investors should not read the sequential margin step-down as deterioration; it is the expected cost of the fastest node ramp in company history plus a normalizing below-the-line.

Implied FY26 revenue trajectory: H1 came in at ~$76.1B. "Slightly above 40%" on the ~$122.4B FY25 base implies ~$172–175B, so H2 must deliver ~$96–99B, or roughly $47–49B per quarter after the $45.2B Q3 midpoint, i.e., a Q4 stepping to ~$50B. That cadence is aggressive but consistent with a supply-constrained order book and the smartphone A-series reload in Q3–Q4. We model FY26 at $172–176B (+40–44% USD).

Implied FY26 margins: With H1 gross margin at ~67% and H2 guided to the mid-60s as N2 dilution steepens, a full-year FY26 gross margin of ~66–67% is the base case, above our prior 65–67% framework. Operating margin lands ~58–59% for the year. We treat Q2's 55.6% net margin as an above-trend print and model FY26 net margin closer to ~51–53% once non-operating income normalizes.

CapEx and FCF: At $60–64B, FY26 capex is ~50% above FY25 and lifts capital intensity toward ~36% of revenue. That compresses near-term free cash flow and pushes depreciation higher into FY27, both legitimate model inputs the market is right to weigh even as it is wrong to treat the demand-pulled raise as bearish. FY27 capex is not yet quantified, but "significantly higher than the past three years" points to another elevated year.

Analyst Q&A Highlights

Whether a Multi-Year CapEx or Revenue Number Is Coming Back

The dominant line of questioning on the call pressed management, from several directions, to quantify the multi-year outlook, echoing the three-year capex framework TSMC gave during the COVID super-cycle. Management held firm in declining a number while making the directional message unmistakable: capacity will follow demand without hesitation.

Q: "Back in the COVID super cycle, TSMC did provide a three-year CapEx outlook back then. I wonder at this point, will it be possible for you to share any color, maybe for the coming three years' CapEx?"
— Sunny Lin, UBS

A: "We do not have a number to share with you, but as you know, we invest CapEx this year for the future business opportunity. As long as there are business opportunities, we will not hesitate to invest."
— C.C. Wei, CEO

Assessment: The refusal to give a three-year number, paired with "we will not hesitate to invest," is more bullish than a figure would be. It signals demand visibility that outruns the company's willingness to bound it, and it tells investors the elevated-capex regime is open-ended rather than a one-year peak. For modeling, it argues for holding capex elevated across the forecast rather than reverting toward a historical ratio.

What Actually Pulled the CapEx Budget Up ~$10B in Three Months

A recurring question sought the specific source of the year-to-date capex increase, whether accelerators, memory-adjacent components, or back-end packaging. Management's answer separated a demand cause from a cost cause.

Q: "From year to date, TSMC raised the CapEx guidance by almost $10 billion. Can you give me some color where is the upside… Is that from CPU accelerator or memory components or back-end CoWoS expansion?"
— Felix Pan, KGI

A: "Simply put, the most important reason is because of demand continue to increase, and we feel the pressure from the customer… to cooperate with TSMC for the capacity increase. That's one of the major reason. The second reason is inflation. Now we buy the tools with inflation price."
— C.C. Wei, CEO

Assessment: This is the exchange that most directly rebuts the market's bearish read. The capex raise is customer-pulled ("we feel the pressure from the customer"), not TSMC front-running speculative demand. The inflation component is a real cost-push that pressures capital efficiency, but it is secondary. A demand-driven capex raise on a raised revenue guide is a bullish combination the tape mispriced.

Duration of the Supply Shortage

Analysts repeatedly probed how long leading-edge supply stays behind demand even after the elevated build-out, and how large the unconstrained-demand gap really is. Management extended the window to the end of the decade and declined to quantify the gap beyond "very big."

Q: "I think all of us on this call are assuming that the unconstrained demand for three nanometer and below is sort of 30%-50% above your ability to supply. Is it in fact much larger than 30%-50% above?"
— Robert Sanders, Deutsche Bank

A: "No, we don't have a number to share. Let me say that the gap is very big… a very big gap."
— C.C. Wei, CEO

Assessment: A supply-demand gap characterized as "very big" and persisting toward 2029–2030 is the single most important structural datapoint for the multi-year model. It underwrites utilization and pricing power for the entire forecast horizon and reframes the capex raise as capacity that is pre-sold rather than speculative.

The AI Accelerator CAGR — Is It Being Raised Again?

A recurring line of questioning tried to extract an updated AI-accelerator revenue CAGR (the 54–56% figure carried since Q1), given accelerating token consumption and the agentic-CPU dynamic. Management would only say it is higher, declining a new number.

Q: "What's the update of the AI semi CAGR, and how should we look at the contents of this AI semi related to TSMC's growth?"
— Charlie Chan, Morgan Stanley

A: "If you asking about the AI's CAGR, let me give you not a number, but it's stronger and stronger. We don't give you the number today because it continue to increase, so we don't know how to answer this question. Stronger than what we said before."
— C.C. Wei, CEO

Assessment: "We don't know how to answer this question" because the demand keeps rising is, in a CEO of Wei's discipline, a stronger statement than a revised figure. It says the AI-accelerator demand curve is still steepening, and it implies the 54–56% framework is now a floor. We continue to treat AI accelerators as approaching half of TSMC revenue by the end of the decade.

Customer Concentration as AI Outgrows Everything Else

A pointed question surfaced the risk that TSMC's exposure to its top handful of customers is now larger than at any point in its history, as AI demand dwarfs other end-markets. Management rejected the framing, arguing the customer base is broadening rather than concentrating.

Q: "Could I ask about the risk that you see around customer concentration, as AI demand continues to significantly outgrow other end markets? I think your exposure to your top five customers is becoming meaningfully larger than at any point in your history."
— Jim Fontanelli, Arete

A: "No, that's not our concern… it's not what you said that the bigger customer is growing bigger and bigger. No. There's a lot of new player in the AI industry."
— C.C. Wei, CEO

Assessment: This is the one bear point that got materially louder this quarter, and management's flat dismissal does not fully retire it. The "lot of new players" answer (sovereign AI, AI labs, hyperscaler ASICs, custom-silicon startups) is a real diversification argument, but it does not change the fact that a small number of very large accelerator customers now drive a majority of incremental revenue. We move customer concentration from a dormant to an active-mild watch item.

Back-End Competition and Cannibalization Risk

Questioners pressed on whether rising packaging competition, specifically Intel's EMIB-T, threatens TSMC's foundry value-add across the front-to-back stack. Wei separated the two businesses and turned the competition into a positive for the front end.

Q: "if on the back end competition is rising, especially coming from Intel EMIB-T, are you worried that your value add for your overall foundry business… could actually be cannibalized with growing competition?"
— Haas Liu, Bank of America

A: "The front end's wafer business and the back end's business are two different things… since our capacity in the back end is so in shortage mode… I welcome that the competitor offers some of the flexibility to my customer so that their front-end wafer can be put into the package, and that help TSMC's front-end wafer business."
— C.C. Wei, CEO

Assessment: Confident and correct. With CoWoS capacity a binding constraint, competitor packaging that lets a TSMC wafer ship is accretive, not cannibalistic, because the front-end wafer is the larger and higher-margin part of the value chain. This reframes back-end competition as a release valve rather than a threat, at least while TSMC's own packaging is sold out.

Balancing GPU, CPU, and XPU Wafer Allocation

Building on the agentic-CPU theme, an analyst asked management to size the growth potential across chip types. Wei declined a split but confirmed all of it runs on TSMC's leading edge and that allocation is an active balancing exercise.

Q: "you mentioned about the agentic AI and the CPU growth potential. Can you give us more update among that AI, different kind of chips between GPU, accelerators, or CPU? What you see the growth potential and your visibility?"
— Laura Chen, Citi

A: "all of them are in TSMC, and they're also using the same kind of leading-edge technologies. We working with our customer to allocate the wafer, the supply, to balance the CPU, GPU, XPUs ratio."
— C.C. Wei, CEO

Assessment: The strategic point is that TSMC is agnostic to which AI chip type wins, GPU, CPU, or custom XPU, because it manufactures all of them on the same nodes. Agentic AI reviving CPU demand simply widens the wafer pool without introducing a mix risk. The active-allocation comment also underscores that supply, not demand, is the binding constraint.

What They're NOT Saying

1. A three-year capex or revenue number: Management repeatedly declined to revive the COVID-era multi-year capex framework or to quantify multi-year growth. The silence is bullish in tone ("stronger than before") but it does leave the outer-year capital commitment, and its FCF and depreciation consequences, undefined for investors trying to model 2027–2028.
2. An updated AI-accelerator CAGR: Despite multiple attempts, no revised figure beyond "stronger and stronger." The refusal reads as confidence, but it also means the 54–56% framework is now stale, and the Street is left to guess how much higher.
3. The composition of the non-operating income surge: The NT$95.8B non-operating line, roughly 3.3x normal, was not decomposed in the disclosed condensed statements. How much is durable interest income versus one-time FX or investment gains matters for the quality of the EPS beat, and it was not addressed on the call.
4. A hard Arizona timeline: The additional $100B and ~4 fabs came without a schedule ("depends on the market situation"). The commitment is real; the pace, and therefore the phasing of the overseas margin dilution, is deliberately left open.
5. Pricing specifics: Wei's "the higher the better" on wafer pricing confirms the posture but avoids any node-level ASP detail, leaving the market to infer the pricing power embedded in the margin guide rather than see it.
6. Any acknowledgment of an AI-digestion risk: With the tape selling AI-capex-bubble fears, management offered no scenario in which hyperscaler capex moderates. The confidence is likely warranted given the order book, but a complete absence of downside framing is itself worth noting.

Market Reaction

• Pre-print setup: TSM ADR closed at $419.48 on July 15, entering the print +38.0% YTD, +76.6% over the trailing twelve months, and roughly flat over the prior 30 days (−1.5%). The stock sat near the top of its 52-week closing range of $227.33–$477.57, versus an S&P 500 up 10.6% YTD.
• Reaction (early trading, session still open at publication): The ADR traded down approximately 5% in the pre-market and early session despite the comprehensive beat-and-raise. The regular-session close was not yet settled at the time of writing.
• What moved it: The capex raise, not the earnings. Coverage pinned the selloff on the step-up to $60–64B plus the additional $100B Arizona commitment, read through an "AI-capex-intensity" lens.
• Sector backdrop: The print landed inside a broad, risk-off semiconductor tape, with the Kospi down ~6%, the Nikkei down ~3%, and memory names (SK Hynix, SanDisk) off 7%+ on memory-cost-inflation and AI-bubble narratives.

The pattern is familiar. TSMC has now delivered several beat-and-raise quarters that were met with a modest-to-sharp sell rather than a rally: Q3 2025 fell ~3% on a beat-and-raise and then re-rated ~25% over the following three months; Q1 2026 fell ~2.5–3% on a comprehensive beat. The mechanism each time is positioning into a stock that has run hard, plus a specific worry the market fixates on. This quarter that worry is capital intensity.

The capital-intensity concern deserves a direct answer rather than a dismissal. It is legitimate that a $60–64B budget compresses near-term free cash flow and raises future depreciation, and that a rising capex/revenue ratio is a real change in the financial profile. What the selloff gets wrong is the causality. TSMC did not raise capex against a hope; it raised capex while simultaneously raising the full-year revenue guide by ten points, and management was explicit that the increase is customer-pulled. A monopoly investing into contracted, supply-constrained demand is the opposite of a bubble. Against a hostile semis tape, that distinction was not what the market wanted to hear, but it is the one that matters twelve months out.

Street Perspective

Debate: Is the CapEx Raise a Demand Signal or a Bubble Tell?

Bull view: The $60–64B budget plus $100B more for Arizona is demand-pulled (management said customers are pressuring TSMC to build) and sized against a revenue guide raised the same day. Capacity that is pre-sold is not speculative; the supply gap is "very big" through 2029–2030. Spending more is the correct response to a monopoly's demand outrunning supply.

Bear view: A capex/revenue ratio pushing ~36% is historically high and compresses free cash flow while lifting depreciation for years. If AI demand digests even briefly, the fixed-cost base becomes a margin liability, and the whole AI supply chain (memory down 7% the same day) is flashing bubble anxiety. The raise is what a top-of-cycle looks like.

Our take: Bull. The distinguishing test is whether capex leads or follows demand, and here it demonstrably follows: the revenue guide went up ten points on the same call. The FCF and depreciation costs are real and we model them, but a demand-pulled raise from a supply-constrained monopoly is a strength the tape mispriced.

Debate: Valuation — Too Far, Too Fast?

Bull view: At a post-print ~$398–419 ADR and FY26 EPS of ~$16.50–17.50 (raised on the guide and margin), TSM trades at ~23–25x forward earnings for a business now guided to 40%+ revenue growth with 66%+ gross margins and a monopoly at the leading edge. That is below where NVDA and AVGO trade for slower growth. The ten-point guide raise lifts the E faster than the multiple, so a flat multiple still carries the stock.

Bear view: The stock is +76% over twelve months and near an all-time high. At ~24x for a capital-intensive foundry with rising customer concentration and a capex/revenue ratio at record highs, the multiple has little room to expand, and any AI-demand wobble or macro shock creates outsized downside. The easy money in the re-rating has been made.

Our take: Bull, but with less asymmetry than at any prior point in our coverage. This is now the most valuation-dependent Outperform we have held on TSM. The reason it is still Outperform is that the FY26 earnings base was just revised up materially, and the stock de-rated ~5% on the print; the entry is better after the selloff than before it.

Debate: Can Margins Hold as N2 Dilution Steepens?

Bull view: Q3 gross margin guided to 65–67% already absorbs a steepening 3–4 point N2 dilution, and N3 crosses the corporate average in H2 2026 as an offset. A business printing 67.7% with a record 77% advanced mix has a structurally higher through-cycle floor than the stated 56%.

Bear view: N2's 3–4 point dilution is the leading edge of a multi-quarter drag, overseas fabs widen dilution to 3–4 points as Arizona scales, and Q2's 60% operating margin was partly a non-operating and utilization peak. Mid-60s gross margin is the ceiling, not a way station to higher.

Our take: Split. We agree Q2's operating margin is a near-term high and model a sequential step-down, but the H2 guide holding the mid-60s through the steepest phase of the N2 ramp is a bullish tell. We model FY26 gross margin at ~66–67% and treat the 56% long-term target as a conservative floor management has not yet chosen to revise.

Model Update & Valuation Framework

Valuation framework: At a post-print ~$398–419 ADR and FY26 ADR EPS of $16.50–17.50, TSM trades at roughly 23–25x forward earnings. For a business now guided to 40%-plus revenue growth with mid-60s gross margins, a widening process-technology monopoly, and a supply gap the CEO calls "very big" through 2029–2030, that multiple remains below where the highest-quality AI beneficiaries (NVDA, AVGO at ~30–35x) trade on comparable or slower growth. The persistent discount reflects Taiwan concentration and capital intensity, both of which the Arizona build progressively (and expensively) addresses.

Revised risk-reward: From a post-print ~$405 midpoint, base ($440–470) implies ~+9–16% upside, bull ($500–520) ~+23–28%, bear ($300–320) ~−21–26% downside. The up-to-down ratio has compressed to roughly 1:1 in base-to-bear terms, the least asymmetric setup in our TSM coverage. What keeps it an Outperform rather than a Hold is that the earnings base was just revised up by ten points of growth, and the stock cheapened ~5% on the print, so the multiple is being asked to do less of the work than the raised E.

Thesis Scorecard: Q1 2026 Signposts Revisited

The Q1 2026 recap laid out the downgrade triggers and the bull/bear signposts for Q2. None of the downgrade triggers fired, and every operating signpost tripped bullish.

Scorecard summary: Ten of twelve signposts tripped bullish or better, including the FY26 guide, which beat even our above-consensus expectation (we had modeled a migration to "mid-30s"; management went to 40%+). Two items carry a caveat rather than a bearish verdict: capital intensity (a bullish demand signal that nonetheless raises the FCF and depreciation watch) and customer concentration (newly surfaced by analysts). No downgrade trigger from the Q1 recap fired.

Full Bull / Bear Matrix: Updated Through Q2 2026

Overall matrix assessment: All nine bull pillars are confirmed or strengthened, with agentic AI evolving into a CPU-resurgence vector and the rare-capacity pillar validated by the 2nm Arizona commitment. The bear side is where this quarter genuinely shifted: valuation escalated to active-elevated and is now the primary bear, customer concentration surfaced as a new active-mild watch item, and an AI-capex-intensity concern emerged (the proximate cause of the selloff). The net thesis is stronger on fundamentals and more valuation-dependent on the stock, a combination that keeps the rating at Outperform but tightens the conditions under which it would move to Hold.

Bottom Line: The Sixth Consecutive Outperform

Rating decision: We maintain Outperform on TSM, the sixth consecutive Outperform across our coverage (Q1 2025 initiation, Q2–Q4 2025 recaps, Q1 2026 recap, and this one). The maintenance is active, not inertial: the quarter cleared every operating criterion in the framework.

• Revenue at guide ceiling: $40.20B vs $39.0–40.2B. ✓
• Gross margin above guide: 67.7% vs 65.5–67.5%. ✓
• Operating margin above guide: 60.3% vs 56.5–58.5%, first ever ≥60%. ✓
• FY26 guide raised, not cut: ">30%" to "slightly above 40%." ✓
• N2 ramp on track: 3% of wafer revenue, "steep," capacity "bigger" than the symposium roadmap. ✓
• HPC record with broadening demand: 66% of revenue; management cites "a lot of new players." ✓

Why the selloff does not change the call. The market sold the capital bill: a $60–64B budget and $100B more for Arizona, read through an AI-bubble lens on a hostile semis tape. But the capex raise was accompanied by a ten-point increase to the full-year revenue guide and was explicitly customer-pulled. A demand-pulled capital increase from a supply-constrained monopoly is a bullish signal, not a bearish one. The costs it carries, lower near-term FCF and higher depreciation, are real and in our model; the demand it reflects is more important twelve months out. The reaction rhymes with Q3 2025's −3% on a beat-and-raise, which re-rated ~25% higher over the subsequent quarter.

What is genuinely different this quarter. Two things. First, valuation is now the binding constraint on the rating: at +76% over twelve months and near an all-time high, this is the most valuation-dependent Outperform we have held on TSM, and the base-to-bear asymmetry has compressed to roughly 1:1. Second, customer concentration surfaced as a real question for the first time. Neither breaks the thesis, but both narrow the margin for error, which is why the Hold-downgrade tripwires below are tighter than in prior quarters.

What could change our mind:

• Downgrade to Hold: H2 2026 gross margin sliding below 64% on N2 dilution running ahead of the 3–4 point guide; the FY26 growth guide trimmed on the Q3 call; a hyperscaler capex-digestion signal (2027 AI capex guided to sub-15% growth); or a further multiple expansion that pushes TSM above ~28x forward without a commensurate estimate revision.
• Downgrade to Underperform: A Taiwan-Strait escalation with operational impact to the fabs; gross margin sustained below 57% for two consecutive quarters; or evidence of an N2 yield problem delaying the 2nm ramp.

For now, the AI foundry is compounding faster than the multiple, the supply gap is "very big" into the end of the decade, and the stock just cheapened ~5% on a beat-and-raise. We stay Outperform and treat the capex-driven weakness as an entry.