At a glance. Q2 landed at the high end of dollar guidance, full-year growth was lifted to slightly above 40% in US dollars, and CapEx went to US$60–64bn. The CoWoS sentence that matters: the majority is still CoWoS; a glass alternative needs about another one year to be production-ready with a customer; and packaging capacity is so tight it now limits customers' growth. The call never said 800 V.
What changed this quarter
The dollar print and the guidance both moved up. Q2 revenue US$40.2bn, +12% sequentially, high end of the range. Full-year 2026 growth is now slightly above 40% in US dollars. Q3 is guided US$44.6–45.8bn.
CapEx was raised again. January was US$52–56bn; April leaned to US$56bn; July is US$60–64bn. About 70–80% advanced process, 10% specialty, 10–20% advanced packaging, testing, mask-making and others. Wendell Huang would not split packaging out of that last bucket.
Arizona and Taiwan were both stepped up. An additional US$100bn in Arizona (cumulative US$265bn) for more 2 nm-and-below logic fabs and advanced-packaging fabs — "additional four or more," frontend and backend. In Taiwan, 13 leading-edge and advanced-packaging fabs over the next several years.
N2 started to dilute the margin on purpose. Q2 gross margin 67.7%. Q3 is guided 65–67% because the N2 ramp is expected to take 3–4 percentage points off gross margin in the second half, partly offset by utilisation and cost.
The numbers
| Q2 2026 | Value | Change |
|---|---|---|
| Revenue | NT$1,270.38bn / US$40.20bn | +12.0% QoQ, +36.0% YoY (NT$); +33.7% YoY (US$) |
| Gross profit | NT$860.31bn (~US$27.3bn), margin 67.7% | +1.5pp QoQ, +9.1pp YoY |
| Operating profit | NT$766.60bn (~US$24.3bn), margin 60.3% | +2.2pp QoQ |
| Non-operating | NT$95.83bn gain | includes NT$63.20bn VIS share disposal / mark-to-market (NT$2.24 of EPS) |
| Net income | NT$706.56bn (~US$22.4bn), margin 55.6% | +23.4% QoQ, +77.4% YoY |
| Diluted EPS | NT$27.25 (US$4.31 per ADR) | +23.4% QoQ, +77.4% YoY |
| ROE | 45.9% | — |
Cash and marketable securities NT$3.5tn / US$110bn. Q2 operating cash NT$783bn; CapEx NT$496bn / US$15.7bn. Cash dividend 2026 NT$24 per share, +33% versus 2025's NT$18.
Platform mix: HPC 66% (+20% QoQ), smartphone 22%, IoT 5%, automotive 4%, DCE 1%. Wafer-revenue nodes: N2 3%, N3 30%, N5 33%, N7 11%; 7 nm and below 77%.
NVIDIA 800V read-through
This is a CoWoS call, not an 800 V call. NVIDIA's 800 VDC list does not name TSMC. The GPU that list is written to feed is a CoWoS part. Treating Q2 2026 foundry growth as 800V evidence would be a misread.
What the call actually adds to the CoWoS file:
- Majority still CoWoS. Glass-core / glass-substrate / glass-carrier work is an alternative "to lower down the cost," in a pilot announced a few quarters earlier, about one more year from being mature enough to put into production with a customer. TSMC is working with a substrate vendor so the customer can still get a product to market.
- 14× reticle is the size roadmap already shown at the Symposium, restated here. Not a 2026 volume mix.
- Backend is the bottleneck. "Our packaging capacity is so tight that now it limits my customers' growth." C.C. Wei welcomed Intel EMIB-T as extra flexibility so frontend wafers can be packaged — and said frontend and backend are different businesses: if they were the same, "you can expect ASE become the frontend competitor also."
- Packaging CapEx stays in a 10–20% bucket with test and mask-making because bottlenecks move. Testers are also short. He would not break CoWoS out as its own CapEx line.
- COUPE (silicon photonics / CPO) has started production and should become "fairly important" over the next few years as data centers cut power and raise interconnect bandwidth.
N2, A14 and Arizona logic are frontend facts. They sit next to CoWoS on the same call; they are not CoWoS revenue.
Segments that moved
HPC at 66% is the AI print: accelerators, plus CPUs that agentic AI is pulling back into the data center. C.C. Wei would not split GPU / XPU / CPU, only that "all of them are in TSMC" on leading-edge nodes and wafers are being allocated to balance the three.
N2 is 3% of wafer revenue and the named gross-margin dilute. A14 is on track: risk production 2027, volume 2028, with A13 / A12 volume 2029. Device performance and 256 Mb SRAM yield were each described as close to 90% on an internal vehicle.
Mature nodes are not a broad shortage. Power-management IC (0.18 μm, 90 nm class) and sensors tied to AI data centers are tight; consumer commodity mature is not.
Guidance
- Q3 2026: revenue US$44.6–45.8bn (+12% QoQ / +37% YoY at the midpoint); gross margin 65–67%; operating margin 56–58%; FX assumption NT$32 per dollar.
- Full year 2026: revenue growth slightly above 40% in US dollars. CapEx US$60–64bn.
- Gross margin: N2 dilution 3–4pp in the second half; overseas-fab dilution 2–3% early, 3–4% later.
- Next three years' CapEx: "even more significantly higher" than the past three. No three-year dollar figure. AI CAGR previously mid- to high-50s: now "stronger than what we said before," no new number.
- Supply gap: "very big." Demand described as strong from now through 2029–2030, with no guarantee there is no dip in between. TSMC checks data-center build progress so chips are not built into inventory.
Management Q&A
Q: Can TSMC give a three-year CapEx outlook for 2026–28, as it did in the COVID cycle?
A: No number. CapEx this year is for future opportunity, and the company will not hesitate to invest while the AI megatrend holds. Last time it said the next three years would be significantly higher than the past three; now that next three years will be even more significantly higher.
Q: With Arizona now at a cumulative US$265bn, what is the schedule for the extra US$100bn?
A: The schedule depends on the market. Today's demand is why the extra US$100bn was announced. How many fabs? Many — probably four or more additional, combining frontend and backend.
Q: How does TSMC answer Samsung Foundry (memory profits) and Intel (US policy), and ASML expanding EUV slots for 2028?
A: C.C. Wei said he is jealous of the Korean competitor's profits, and that TSMC also has government support even if it does not announce it. The fundamentals have not changed in 30–40 years: technology, manufacturing, and customer trust. Choosing a technology partner is "not buying milk from 7-Eleven"; it takes about five years to understand the node, run test chips, prepare capacity and ramp.
Q: Is the five-year AI-semiconductor CAGR still mid- to high-50s, and does agentic-AI CPU count?
A: Not a number today, because it "continues to increase." The qualitative update is stronger than what we said before. CapEx is being raised for that reason.
Q: EMIB-T is gaining traction — how does TSMC react?
A: Packaging capacity is so tight it limits customers' growth, so extra flexibility in the market is welcome and helps TSMC's frontend wafer business, which is the majority of the company. The technology "looks good, according to the newspaper"; TSMC hopes it succeeds and shares some of the loading. TSMC is working to shorten its own demand-capacity gap.
Q: If that new technology hits problems, will TSMC support the customer?
A: Number one is customer success. "Whatever that we can help our customers, the business, we will do it."
Q: How does TSMC decide capacity — including competitors building, and data-center / power delays that could leave chips in inventory?
A: Competition is always in the calculus, plus top-down and bottom-up demand. Every customer "tell me the truth," but "you put all the truth together, it's not a truth," because CEOs are aggressive — so TSMC marks the numbers down and still raised CapEx from US$52–56bn to US$60–64bn. It also checks AI data-center buildings, locations, demand and racks so TSMC chips are not built into inventory. From today through probably 2029–2030 demand is very strong; whether there is a dip in between, "I'm not very sure."
Q: Should leading-edge foundry profitability be higher than memory, and will TSMC take pricing up?
A: Higher is better, but TSMC is a partner and will not "suddenly increase our price" by 4× or 5×. It earns enough gross margin to sustain expansion for customers and itself. C.C. Wei is jealous of memory gross margins of 86% — "68, I'll be happy about that."
Q: Is customer concentration a risk as AI outgrows other end markets?
A: Not a concern. Large customers getting larger is welcome, and there are also a lot of new players in AI.
Q: Will TSMC finance or invest in customers' customers, as some direct customers are doing?
A: No. Every company has a different strategy. TSMC's current model with current customers is working smoothly.
Q: What is the timeframe for the extra US$100bn in the United States?
A: There is a plan, not a firm public schedule. Progress depends on the market and customer demand. The company is trying to move as fast as possible in the US, in Taiwan, and in Japan, because the demand-supply gap is so big.
Q: When does the COUPE platform become a material top-line contributor?
A: Production has started now and will ramp. AI data centers need lower power and more interconnect bandwidth, so COUPE demand should keep rising and become fairly important in the next few years.
Q: Within AI, how do GPUs, accelerators and CPUs compare?
A: No specific split. All of them are at TSMC on the same class of leading-edge technology. The company is allocating wafers to balance CPU, GPU and XPU.
Q: What is the progress on glass-core, glass-substrate or glass-carrier CoWoS versus the 14×-reticle CoWoS roadmap?
A: Today the majority is still CoWoS. An alternative is being developed to lower cost, working with a substrate vendor so the customer can still get product to market. A pilot line announced a few quarters ago needs about another one year to be mature enough for production with the customer.
Q: What is the revenue-growth outlook for the next several years, and is the CapEx raise cloud AI, edge AI, or tool-price inflation?
A: "Next few years is going to be a very good business for TSMC. That's all I can say." The driver is "all AI related, everything."
Q: If EMIB-T gains backend share, does that become a gateway that eats TSMC's frontend logic wafers too?
A: Frontend and backend are two different things. If they were the same, ASE would already be a frontend competitor. Backend is still in shortage, "the gap is bigger," so competitor flexibility that lets customers package TSMC frontend wafers helps TSMC's frontend business.
Q: Does high-NA's smaller field size and die-stitching slow adoption even if the tool gets better?
A: High-NA is a very good, high-performance tool; today's field size is one-half, and that goes into manufacturing-cost math. TSMC works with ASML on cost and maturity. Use-or-not is always technology, maturity, and cost.
Q: Is unconstrained demand for 3 nm and below 30–50% above supply, or larger?
A: No number. "The gap is very big. Bigger than…" — he stopped short of a memory comparison. "A very big gap."
Q: Are the Symposium CAGRs still right — N2 family ~70% from 2026–28, N3+N5 ~25% from 2022–27?
A: Those were on a chart at the Symposium. "Now is bigger. That's all I say."
Q: How much of the 10–20% packaging / test / mask CapEx is advanced packaging alone, and will TSMC break it out?
A: The company keeps flexibility between frontend and backend because the bottleneck moves. Ballpark backend remains 10–20%, "a big range." As time goes on, some products need more testers — testers are in shortage — so CapEx shifts among tester, packaging and other areas. It will not be more specific.
Q: Of the almost US$10bn CapEx raise this year, how much is CPU / agentic AI, memory companion chips, or backend CoWoS?
A: The most important reason is demand continues to increase, and customers are pressing TSMC to add capacity. The second reason is inflation on tools.
Q: Are mature nodes in shortage, and how does AI crowding-out versus weak consumer demand split?
A: Mature covers many segments. Only AI-related mature is in shortage: number one is power-management IC (0.18 μm, 90 nm class) for AI data centers, and sensors. Other consumer commodity mature is not in a lot of shortage.
Disclaimer
English notes on TSMC's second-quarter 2026 earnings conference of 16 July 2026, from the company's English materials and the edited call transcript. Figures and statements are as presented by management, are unaudited, and have not been independently verified. USD figures that TSMC reported in dollars are used as reported; other USD equivalents are approximate at about NT$31.5 to the dollar. For reference only; not investment advice.