October 15, 2026 14 min read

TSMC (TSM) — Q3 FY2026 Analyst Q&A Breakdown

Source: TSMC Q3 2026 earnings call transcript (transcripts/TSM/Q3_2026.txt), October 15, 2026 · Fiscal period: Q3 FY2026 · Report date: 2026-10-15

Eight analysts, 17 question units. Every question below is accounted for in the count check at the end.

Laura Chen, Citi

The Question: As AI semiconductor demand turns aggressive, how does TSMC size AI revenue growth, and how does it reconcile the chip-demand surge with power availability, funding, and chokepoints (HBM, advanced packaging, networking)? What would adjust the capacity plan to minimize a supply-chain overbuild?
Management’s Response: C.C. Wei anchored to process, not numbers: the “very sophisticated capacity planning process” weighs CSP capex, data-center build-out, financing, power supply and component supply. Genuinely new is the validation discipline behind it — TSMC now sends its own teams to verify U.S. data-center build-out and power supply rather than relying on customer forecasts. The bottom line: “the demand is higher than the supply, and we are working very hard to narrow the gap.” No updated AI sizing.
The “Analyst’s Take”: Evasive in a familiar way — process credibility substitutes for quantification, and no scenario table on overbuild was offered. The self-verification step is the one piece of new substance, and it raises the floor on demand-data quality.

Laura Chen, Citi (turn 2)

The Question: Is the previously communicated AI accelerator CAGR — the company’s own “mid to high 50s” percent for 2024–2029, re-read back by the analyst — still the operative figure?
Management’s Response: No revision. C.C. Wei: “Whether it is a high 50 CAGR, it does not matter. It probably will be higher.” The only affirmation: no sign of demand slowdown.
The “Analyst’s Take”: A deliberate non-revision that functions as a soft raise — direction stated, magnitude withheld. The burden now shifts to results to prove the direction.

Laura Chen, Citi (turn 3, two-part)

The Question: (a) Does High-NA EUV adoption around 2030 accelerate productivity and further revenue growth? (b) With capex intensity much higher, how should pricing power, long-term gross margin and capital efficiency be thought about?
Management’s Response: (a) High-NA remains timed for 2030, with productivity improving every year. (b) The margin philosophy of the call: “there is no ceiling for TSMC’s profitability, but a dramatic increase will not happen overnight either” — customers succeed first, TSMC captures value after. No mechanism, margin target or capital-efficiency metric offered.
The “Analyst’s Take”: Part (a) answered flatly; part (b) produced the most quotable margin framing of the Q&A, notable precisely because it pairs a no-ceiling claim with a pace warning. Evidence-free but directional.

Gokul Hariharan, JPMorgan

The Question: Will wafer-supply tightness be greater in 2027 than 2026, and what is planned wafer-capacity growth? The analyst’s framing — roughly a decade since TSMC added capacity at more than 10% a year — was an analyst-provided premise management disputed in the relay (annual adds have run mid-to-high single digit or lower).
Management’s Response: C.C. Wei: demand “very, very high,” the demand–supply gap “still quite big” — but whether the gap widens or narrows next year: “I cannot answer … because [customers] continue to increase their demand.” Wendell Huang added the only reusable frame: capacity percentages are “a bit too early” to discuss, and the growth that matters is in advanced nodes rather than the total.
The “Analyst’s Take”: Premise corrected, forecast withheld. The advanced-node composition argument is a real point — but nothing here lets an analyst model 2027 supply.

Gokul Hariharan, JPMorgan (turn 2)

The Question: On the more aggressive U.S. expansion — reported Texas exploration, and reported engagement with a customer building a fab (the analyst’s framing, with the SpaceX/Tesla reference analyst-provided and not confirmed). What are the contours — TSMC-run like JASM or ESMC?
Management’s Response: Substantively new. Confirmed teams are in Texas evaluating expansion outside Arizona. Confirmed discussions are underway with the customer planning a “Terafab” about TSMC utilizing that capacity — “still in discussion,” details to come only with “any solid conclusion.” The one fixed principle so far: “the fab has to be run by TSMC.”
The “Analyst’s Take”: More operational detail than most unaudited expansion stories get. Operation-control requirement effectively rules out minority-venture structures and keeps JASM/ESMC as the working template.

Krish Sankar, TD Cowen

The Question: Some customers discuss a secondary wafer market — allocation passed between customers. Is it small, customer-specific, or an emerging trend TSMC permits?
Management’s Response: A hard rule, not a trend: TSMC does not allow customers “to sublet or auction off the capacity.” The only sanctioned pass-through is a mask-licensing agreement — “this is not our common practice. The amount is very, very small.”
The “Analyst’s Take”: The rare fully closed answer — prohibition plus a minuscule exception, both quantified as negligible. Directly deflates secondary-market speculation.

Krish Sankar, TD Cowen (turn 2)

The Question: TSMC adopting High-NA only around 2030–31 while competitors adopt earlier — a technological disadvantage?
Management’s Response: Adoption is a gate, not a race: a tool enters mass production only with the best technology, reasonable cost and — “most important” — the best maturity; 2030 is when High-NA matures. A new, dateable commitment: TSMC is working with ASML on 12-inch photomask supply “by 2031.” On the disadvantage question: “No, because we have confidence that we are still the best company to offer the most advanced technology.”
The “Analyst’s Take”: Direct and historically consistent — TSMC also judged EUV later and won. The 2031 photomask timeline is the falsifiable nugget, implying High-NA weaves into the A14/A16 generation rather than following it.

Charlie Chan, Morgan Stanley

The Question: (a) One IDM customer is expanding capacity aggressively — concern that it may insource more than outsource to TSMC; how does TSMC handle such a customer? (b) Does “A14 could be a bigger node than N2” already assume this IDM outsources server CPUs at A14?
Management’s Response: (a) A principle and no news: “we treat every customer fairly” whether fabless or IDM, and customers’ internal capacity is explicitly counted in TSMC’s capacity math (“we take everything into consideration”). (b) A fresh, checkable datapoint: A14 customer engagement is running above N2’s at the same lifecycle stage so far — and “we expect that the A14 will be a bigger node than our current N2 technologies.”
The “Analyst’s Take”: (a) got no signal; (b) got signal. The A14-above-N2 engagement claim is new evidence that the next node’s demand pre-load is real, independent of any one IDM’s choices.

Charlie Chan, Morgan Stanley (turn 2)

The Question: With customers moving to very large reticle dies (14x), can TSMC’s system-on-wafer/3.5D packaging roadmap hold those customers — and would TSMC support a customer that initially chose a competitor’s EMIB-T but cannot get sufficient volume?
Management’s Response: The packaging answer of the call: “We are in the real-world production of 5x reticle size. I don’t believe that anybody can do that today except at TSMC, and the yield is excellent.” Development runs 9x toward 14x; CoWoS keeps a cost-reduction roadmap and remains “the number one choice.” On being the rescue path for stranded customers: “I try to avoid to answer it directly, but let me say that, yes, we will be.”
The “Analyst’s Take”: Two falsifiable new facts — live 5x reticle production with claimed-exclusive availability, and an explicit willingness to absorb competitor fallout. The dodge-and-confirm structure is itself the message.

Sunny Lin, UBS

The Question: (a) For 2027–2028, does capex now grow in line with sales or faster (the analyst recalled management’s earlier promise it could grow a bit less than top line); (b) does tool-price inflation become a meaningful 2027 capex variable?
Management’s Response: Wendell Huang: numbers for beyond-2026 are premature; the 2026–2028 capex pledge is the same step-up as prepared remarks (reiterated, not newly framed). The intensity promise was reaffirmed in place: “if [we] do our job right, the increase in revenue will outpace the increase in CapEx, and that is still the case.” On inflation, the candid concession: “The tool price inflation is real. The tool price, material, labor cost, et cetera” — with the counter that the majority of capex is tied to long-term demand rather than current prices.
The “Analyst’s Take”: Direct where possible, silent where not: inflation acknowledged but unsized, intensity reaffirmed, 2027 unquantified. The composition argument concedes the direction while denying the magnitude.

Sunny Lin, UBS (turn 2)

The Question: Does the widened overseas dilution — which the analyst read as raised to 3–4% starting from 2027, an analyst’s interpretation rather than management’s words — imply accelerating Arizona phase-2 or phase-3 buildouts?
Management’s Response: Wendell Huang confirmed acceleration is happening “everywhere, including in the U.S.” — and re-timed the dilution figure to match prepared remarks: “the latter part of the five-year period, the margin dilution from overseas fab will be 3 percentage point to 4 percentage point, and that still remain the case.” Jeff Su then closed the turn — “That was two questions, Sunny” — cutting off her attempted follow-up before it was heard.
The “Analyst’s Take”: A timing-clarification, not a raise: the from-2027 reading was quietly walked back to the endpoint management had already scheduled. The interruption cost the room the Arizona-phasing detail.

Haas Liu, Bank of America

The Question: Is customer demand shifting toward more efficient compute — CPUs, networking/switches — as monetization pressure increases, and what is the incremental manufacturing and packaging opportunity?
Management’s Response: Essentially the scripted AI framing, with one new admission: demand from agentic AI CPUs “continue[s] to increase” — a stated trend. Attribution refused: whether accelerators lose share is “not for me to make a comment.” The unchanged center: total demand “far above our capability today to supply.”
The “Analyst’s Take”: The refusal is informative — management welcomes CPU growth without conceding any rotation story, and offered no incremental packaging quantification.

Haas Liu, Bank of America (turn 2)

The Question: Ramping a node while simultaneously carrying conventional power delivery and Super Power Rail architectures — does the dual process architecture complicate the yield learning curve and, by extension, gross margin?
Management’s Response: The premise was flipped: smartphone small-die and HPC/AI big-die now arrive on the same node at the same time — “this is to TSMC’s advantage,” because the die-size mix feeds yield learning. “The learning curve, yes, but we are doing a reasonable job, and the customer are very happy today.”
The “Analyst’s Take”: A complexity risk reconstructed as a mix benefit — with customer satisfaction as the only offered evidence and no yield data. Note the asymmetry: no numbers where numbers were requested.

Arthur Lai, Macquarie

The Question: Is power-grid constraint a gating factor that has forced reductions in customer forecasts, and could power tightness accelerate adoption of Super Power Rail?
Management’s Response: Power is “one of the consideration factor[s]” — and the self-verification teams that checked U.S. build-out and power returned “so far so good.” On SPR acceleration: “Answer is no” — backside power is a product-need decision (some accelerators, few phones), not a grid response.
The “Analyst’s Take”: Both halves closed cleanly. The SPR “no” is load-bearing for anyone connecting grid headlines to node migration; management de-coupled them explicitly.

Arthur Lai, Macquarie (turn 2)

The Question: Does the mid-to-high-50s AI demand framework include agentic-AI CPUs — and when will TSMC refresh its AI demand numbers as customers raise CPU forecasts?
Management’s Response: CPUs are not in the calculation — because TSMC “cannot identify which CPU” usage is conventional-server versus AI-server, and customer allocation blurs the line; TSMC “stick[s] to our initial very narrow definition.” Jeff Su enforced the frame: narrow, “but it is certainly part of our business.” No refresh date given.
The “Analyst’s Take”: Definitional honesty over marketing: the headline CAGR deliberately understates one business TSMC elsewhere calls a growing tailwind. The absence of a refresh date keeps the headline number safe for another quarter.

Sahej Singh, Stifel

The Question: On reported significant price increases across mature and advanced nodes (an analyst-reported figure; management neither confirmed nor disputed the reports): do increases apply across the existing contracted book, or mainly to incremental capacity?
Management’s Response: Strategy altitude only: pricing is “very strategic,” varying “by the market segment, … by the location, and … by the technology,” with U.S.-fab output priced “a little bit different from the one we produce in Taiwan.” The goal: “we try to get our value to be recognized” and deliver shareholders the maximum value of TSMC’s technology. The contracted-book questions were never answered.
The “Analyst’s Take”: Non-answer on mechanics; the one new element is explicit location-differentiated pricing in the company’s own words — consistent with a Taiwan-versus-U.S. cost story, but silent on the existing order book.

Sahej Singh, Stifel (turn 2)

The Question: On the reported GlobalFoundries arrangement to supply silicon interposers to TSMC (analyst-reported; management confirmed the deal): how does TSMC think about procuring from specialty foundries versus in-house?
Management’s Response: Designed choice, framed as strategy: the deal was explicitly made as “part of our overall strategy in the back-end supply,” with the priority being customers’ product launches getting enough front-end and back-end support. The admission underneath: “Definitely, we have a little bit shortage on the back end.”
The “Analyst’s Take”: Confirmation plus admission — external sourcing for interposers is a bridge over a self-declared bottleneck, not an arbitrage of cost. Watch whether in-house advanced-packaging capacity still outgrows the workaround.

Synthesis

Clusters

Concern	Analysts (Q#s)	Uniform management answer	Signal
Can supply ever catch the AI demand curve	Citi (Q1), JPMorgan (Q4), BofA (Q12)	“Demand is higher than the supply,” supported by a sophisticated planning process and TSMC’s own data-center/power field checks	The same conviction delivered to every analyst, with no scenario analysis — capacity confidence is programmatic, not improvised
Quantifying the upside	Citi (Q2), Macquarie (Q15), JPMorgan (Q4)	No numbers: “probably will be higher”; no capacity-growth rate; no refresh date	Management is booking unprecedented capex on conviction while withholding the scoreboard
U.S./overseas build-out cost	UBS (Q11), JPMorgan (Q5)	Faster U.S. expansion under TSMC operational control (“the fab has to be run by TSMC”), dilution endpoint pre-scheduled	Acceleration is real; the margin cost is staged, not hidden
Advanced packaging depth	Morgan Stanley (Q9), Stifel (Q17)	Exclusive-formula leadership at large reticle sizes plus partners covering a back-end shortage	Packaging functions as both the moat and the admitted weak link
Long-run pricing power	Citi (Q3b), Stifel (Q16)	“No ceiling for TSMC’s profitability”; strategic, location-differentiated pricing	Value-capture language grows louder exactly as dilution widens

Deflections

Topic (Q#)	What was asked	Deflection technique
Wafer-capacity growth rate (Q4)	Can capacity growth reach about 10% in coming years	Premise correction, then “too early,” then recasting the metric to advanced-node composition
2027 tightness (Q4)	Will supply be tighter next year	Unfalsifiable input — customers “continue to increase their demand”
AI accelerator CAGR (Q2)	Updated mid-to-high-50s figure	Direction-only upgrade: “probably will be higher”
AI-demand refresh schedule (Q15)	When will AI demand numbers be revised up	Definition-stickiness — the narrow accelerator definition holds, no date
IDM insourcing (Q8a)	Will the IDM competitor shift mix toward internal fabs	Customer-confidentiality refusal, pivoted to A14 engagement strength
Pricing mechanics (Q16)	Contracted book vs incremental capacity application	Strategy abstraction (segment/location/technology); the mechanics question left standing
Demand-mix attribution (Q12)	Is demand rotating from accelerators to CPUs/networking	“Not for me to make a comment” — confirms CPU growth without conceding rotation

Contradictions

Contradiction	Prepared-remarks claim (source)	Q&A walk-back
Overseas-dilution timing	Prepared remarks: dilution widens to 3–4 points “in the next few years” as more U.S. fabs ramp	UBS (Q11): CFO re-timed the endpoint to “the latter part of the five-year period … that still remain the case,” invalidating the from-2027 read
“Full speed ahead” expansion breadth	Prepared remarks: “moving full speed ahead … at unprecedented pace”	JPMorgan (Q4): CFO — “it is a bit too early” to name any capacity-growth percentage; the pace remains deliberately uncountable

Unanswered

Gokul [JPM] (Q4): no figure for planned wafer-capacity growth, and no view on whether 2027 tightness exceeds 2026.
Laura [Citi] (Q2): no updated AI accelerator CAGR — only a direction.
Arthur [Macquarie] (Q15): no date for refreshing AI demand growth numbers.
Haas [BofA] (Q12): no attribution of demand between accelerators and CPUs/networking.
Charlie [MS] (Q8a): no position taken on whether the IDM competitor shifts mix toward insourcing.
Sunny [UBS] (Q11): her attempted second follow-up was cut off before the question was stated (“That was two questions, Sunny”).

Count check: 17 entries above vs 17 question units in the transcript — Citi 3 units (its third a single two-part turn), JPMorgan 2, TD Cowen 2, Morgan Stanley 2 (first a two-part turn), UBS 2 (first a two-part turn), BofA 2, Macquarie 2, Stifel 2.