Below is Dolphin Research's compiled Trans of $Taiwan Semiconductor(TSM.US) 26Q3 earnings call

Earnings review: TSMC: Taking orders without overpromising — can it lead the pack?

I. Core takeaways from the print

1. Shareholder returns: Reiterated a sustainable and steadily rising per-share cash dividend, for both annual and quarterly cadence. For the third quarter, declared NT$156 bn in cash dividends for 4Q25.

2. Outlook and capex

a. 4Q: revenue of $52.3–54.3 bn, mid-point +14% QoQ and +58% YoY. GPM 65%–67%, OPM 56.5%–58.5%, with FX at 1 USD to NT$31.8. b. Full year: 2026 USD-reported revenue growth expected at approx. 44% (vs. 'slightly above 40%' last quarter). c. Capex: reiterated $60–64 bn in 2026 (unchanged). Capex for 2026–2028 will be meaningfully higher than the past three years, with tight tool supply and early coordination with equipment vendors to secure capacity.

3. GPM drivers and overseas dilution

a. 3Q GPM at 66.4%, down 130bps QoQ, mainly on initial 2nm ramp dilution. b. 4Q mid-point at 66%, a further 40bps drop, reflecting a steep 2nm ramp and overseas fab dilution, partly offset by productivity gains and cross-node capacity optimization. c. Overseas fabs: helped by scale and cost-down, full-year 2026 dilution to narrow to about 200bps. As more US fabs ramp, dilution will widen to 300–400bps in coming years, with cost structure to improve as Arizona scales.

4. Key financials this quarter

a. Totals: revenue +17.6% QoQ in NT$ / +16.2% in USD to $46.7 bn, above the high end of guidance. Opex ratio 7.8%, OPM 58.7%, EPS NT$29.43, ROE 45.5%. b. Node mix: 2/3/5/7nm accounted for 8%/33%/28%/10% of wafer revenue, with 7nm and below at 79% in aggregate. c. Platforms: HPC +22% QoQ to 68% mix; smartphone +10% to 21%; IoT +5% to 4%; auto +23% to 4%; DCE −12% to 1%. d. Cash flow: operating cash flow about NT$920 bn, capex NT$443 bn (~$13.8 bn). Ending cash balance NT$3.6 tn, cash and marketable securities NT$4.0 tn (~$126 bn). e. Ops: inventory days fell 5 days to 82, driven by higher wafer shipments. DSO fell by 1 day to 28.

II. Earnings call details

2.1 Management commentary

1. Demand outlook

a. 4Q supported by strong demand in advanced nodes, including a steep 2nm ramp. b. Consumer and price-sensitive end markets face component inflation and macro uncertainty, so TSMC remains prudent in its 2027 planning. That said, advanced-node demand stays exceptionally strong with a broader customer base.

2. AI demand

a. Agentic AI is accelerating (moving from chatbots to AI assistants), with open-source models improving efficiency and lowering barriers to use, sustaining advanced-node silicon demand. b. Beyond AI accelerators, agentic AI CPUs are rising alongside continued networking chip demand, and semis will keep gaining value share across the global supply chain. c. Regardless of the technical path, the most advanced nodes are required, and all AI players are working with TSMC.

3. Capacity planning approach

a. In addition to talks with customers and their customers, TSMC dispatched teams on-site in the US to check data center build-outs and power availability. Feedback is positive, and CSPs appear well prepared. b. Using a combined top-down and bottom-up system, TSMC continually assesses demand; advanced-node design lead times are long, and it has clear visibility into customers' multi-year roadmaps and production plans.

4. Capacity expansion

a. Nearly 20 fabs globally are progressing in parallel: over the next few years, 13 advanced-node and advanced packaging fabs will be built in Taiwan, plus 5–6 overseas. b. N2: entering HVM in 4Q25 with good yields, ramping across multiple phases in Hsinchu and Kaohsiung; the N2 family capacity in 2027 will more than double, driven by both smartphones and HPC/AI. c. A14: capacity preparations underway, with risk production in 2027 and HVM in 2028, on schedule.

5. N3 global capacity add (breaking the old norm of no further adds post-node qualification)

a. Taiwan: a new 3nm fab in the Tainan Science Park will begin HVM in 1H27. Due to customer demand, an additional 3nm fab in Tainan will start HVM in 1H28. b. US: Arizona Fab 2 is completed and moving tools in for installation, targeting HVM in 3Q27, slightly ahead of plan. c. Japan: the second fab is planned to bring in 3nm, with HVM in 2028. d. Other levers: continuing to convert 5nm tools to 3nm in Taiwan; flexibly allocating capacity among N7/N5/N3. Management said it does not intend to leave any business on the table and aims to capture every possible opportunity.

2.2 Q&A

Q: With constraints across power, financing, HBM, ABF, advanced packaging, lasers and more, how is TSMC aligning customer demand and avoiding overbuild in the supply chain?

A: All constraints are embedded in our assessment, and we still conclude demand exceeds supply, so we are working to narrow the gap. CSP capex, data center construction, financing, power and other component supplies are all included in TSMC's capacity planning process. Based on its own judgment, it reached the above conclusion.

Q: Any update to the 2024–2029 AI accelerator revenue CAGR (mid-to-high 50s%)?

A: No formal update, but demand is extremely strong with no slowdown observed, and growth 'could be higher'. Management noted whether it is high-50s is not the point, and that the CAGR refers narrowly to AI accelerators.

Q: This capex cycle is higher, and High-NA EUV will be introduced in 2030. With rising capital intensity, how do you view pricing power, long-term margins and returns?

A: There is no ceiling to TSMC's profitability, but step-ups will not be instantaneous. High-NA is an excellent tool, planned for use in 2030, when capex will be far above today. TSMC will keep driving productivity, and its strategy remains to help customers succeed first, then capture its fair value.

Q: Will the supply-demand gap be tighter in 2027 than in 2026? Can capacity growth approach or exceed 10% in the coming years?

A: It is hard to say whether the gap will widen or narrow next year because customers are still raising demand; it is too early to discuss capacity growth rates. Demand is 'extremely strong and rising', and after triangulating with customers and their customers plus on-site DC checks, the gap remains sizable. TSMC is expanding as fast as it can.

Historically, disclosed annual capacity growth has often been mid-to-high single digits or lower. Management reminded that even if the total growth figure is not large, advanced nodes are the primary driver of capacity growth.

Q: How is US expansion progressing? There are rumors of a new Texas site, and could cooperation with SpaceX/Tesla follow models like JASM or ESMC?

A: Texas is under evaluation, and discussions continue with a customer on Terafab, but any fab must be operated by TSMC. Most large customers are in the US and request TSMC to expand there, with strong Gov. support, so considerations have extended beyond Arizona to other regions including Texas. Teams have been dispatched for on-site assessments.

For the customer building its own fab (Terafab), the parties are discussing whether TSMC can utilize that capacity. Both ownership structures are possible, and TSMC will disclose once decisions are final.

Q: Some customers mentioned a 'secondary market' for wafer capacity reallocations. Is this becoming a trend?

A: This is not new: in principle customers are not allowed to sublet or auction capacity, and such cases are minimal. In certain cases, customers may cooperate via mask licensing agreements, but this is not standard practice.

Q: TSMC plans to introduce High-NA EUV around 2030–2031, later than some peers. Is this a disadvantage?

A: We do not see it as a disadvantage; the 2030 timing reflects a judgment on when tools will be mature and costs reasonable. New tools must enter HVM only when the technology is optimal and costs are sensible, with maturity the priority. TSMC works closely with ASML and plans to establish 12-inch masks by 2031.

Management acknowledged High-NA advantages at smaller geometries, but remains confident it will continue to be the best provider of leading-edge technology to customers.

Q: One IDM customer (Intel) is sharply lifting capex and may insource more. Did your view that A14 will be larger than N2 assume outsourcing of server CPUs?

A: No comment on specific customers, but all factors are incorporated, and A14 is expected to be a larger node than N2. TSMC treats IDM and fabless customers alike while considering customers' internal capacity, and plans capacity based on its own judgment. Customer engagement on A14 already exceeds that of 2nm at the same stage.

Q: As customer chips evolve toward 14x reticle sizes, can TSMC's advanced packaging roadmap keep pace? How does it stack up vs. Intel's EMIB-T?

A: TSMC is currently the only foundry in HVM at 5x reticle size, and CoWoS remains customers' first choice. 5x reticle-size production is live with strong yields, with 9x advancing and 14x on the horizon, while CoWoS cost-down is also progressing.

Management acknowledged competitors are 'very competitive', but stressed the first priority is to ensure customers' products launch smoothly by providing the best technology and sufficient capacity.

Q: If customers choosing EMIB-T face capacity shortfalls, would TSMC serve as a backup?

A: The reply avoided a direct answer, but said 'yes' and noted sufficient hints had been given. Management reiterated that getting customers' products to market smoothly is the top priority.

Q: Will capex growth in 2027–2028 catch up with or even exceed revenue growth? Will tool price hikes significantly lift 2027 capex?

A: No numbers beyond 2026 for now, but revenue growth is still expected to outpace capex growth. Capex in 2026–2028 will be markedly above the past three years to support customers' growth. Equipment, materials and labor inflation is real, but long-term demand remains the primary driver of capex.

Q: Overseas GPM dilution expanding to 300–400bps suggests faster follow-on phases in Arizona?

A: Expansion is accelerating everywhere, but the 300–400bps figure remains the same guidance for the back half of the five-year cycle. TSMC is expanding wherever it can, including in the US.

Q: Is AI demand shifting from accelerators to more efficient compute like CPUs and switching silicon?

A: No comment on whether demand favors accelerators or CPUs, but agentic AI CPU demand is indeed rising. Overall AI demand is exceptionally strong, and across all metrics demand is still well above current supply capability.

Q: Future nodes will support both traditional front-side power delivery for phones and Super Power Rail for HPC/AI. Will dual-track ramps lengthen yield-learning curves and impact margins?

A: It is different from the past, but TSMC is well prepared with strong execution, and it did not quantify the margin impact. Historically, new nodes ramped with small, simple chips; today, small phone chips and large accelerator dies are both introduced. Management sees this as a TSMC advantage, with a learning curve that the company is handling well.

Q: Has grid constraint ever caused customers to lower forecasts? Will power tightness push customers toward Super Power Rail?

A: Power is a consideration, but checks so far look good, and customers will not adopt back-side power just because of power constraints. TSMC did not directly answer whether this ever led to lowered forecasts. Whether to adopt back-side power depends on product needs: some accelerators may use it, while phones and others will not.

Q: Does the AI accelerator CAGR definition include CPUs? If CPU demand surges, when will you update AI demand growth?

A: CPUs are excluded, and TSMC sticks to the original narrow definition with no update timing given. The reason is that it is hard to separate CPUs for traditional servers vs. AI servers, and it depends on customer use and deployment. CPUs are outside that metric, though obviously part of TSMC's business.

Q: Reports suggest significant price hikes for both mature and advanced nodes. Do increases cover existing contracts or only incremental capacity?

A: No comment on whether existing contracts are covered; pricing is highly strategic. Pricing depends on segment, location and technology, with US production providing geographic flexibility and different pricing vs. Taiwan. TSMC seeks recognition of its value and to maximize technology value for shareholders.

Q: GlobalFoundries supplies silicon interposers to TSMC. How do you balance sourcing from specialty foundries vs. in-house?

A: This cooperation is part of a broader back-end supply strategy due to slight tightness in back-end capacity. The top priority is to fully secure both front-end and back-end supply for customers, with both parties jointly providing the capacity needed for customers' product launches.

<End of text>

Risk disclosure and statements for this article:Dolphin Research disclaimer and general disclosure