TSMC (NYSE: TSM) released its third-quarter 2026 financial results on Thursday. The company reported consolidated revenue of NT$1.4942 trillion (US$46.72 billion), representing a 50.9% year-over-year increase and a 17.6% sequential increase, exceeding its own guidance range of US$44.6 billion to US$45.8 billion. Net income attributable to shareholders reached NT$763.06 billion, up 68.7% year-over-year and 8.0% quarter-over-quarter, marking the sixth consecutive quarter of record quarterly performance and the eleventh straight quarter of profit growth. Diluted earnings per share came in at NT$29.43, above the market estimate of approximately NT$28.3.
Gross margin for the quarter stood at 66.4%, down 1.3 percentage points sequentially, primarily due to dilution from the 2nm process ramp-up. By process node, the most advanced 2nm node accounted for 8% of wafer revenue in the third quarter (up from 3% in the prior quarter), while 3nm represented 33%, 5nm accounted for 28%, and 7nm made up 10%. Advanced nodes (7nm and below) collectively accounted for 79% of wafer revenue. The simultaneous increase in both volume and pricing of advanced nodes was the core driver behind the revenue beat.
TSMC guided fourth-quarter revenue to range between US$52.3 billion and US$54.3 billion, representing approximately 14% sequential growth, with gross margin expected between 65% and 67%. For the full year 2026, the company projects approximately 44% growth in US dollar revenue. Regarding capital expenditure, the company reiterated its 2026 capital budget of US$60 billion to US$64 billion.
Management repeatedly emphasized during the earnings call that "demand exceeds supply" and that the situation is "very, very high," with the supply-demand gap remaining considerable. On AI demand, management stated that AI-related demand is supported by robust underlying demand for AI accelerators, while the emergence of agentic AI CPUs and continued growth in networking chip demand are further driving chip demand.
CEO C.C. Wei said during the call that the company had previously projected a compound annual growth rate for AI accelerators in the upper range of the 50% bracket from 2024 to 2029, and current demand growth is likely to exceed that expectation. He stated that TSMC has "no intention of leaving opportunities to others" and aims to capture every piece of business possible. TSMC expects capital expenditure from 2026 to 2028 to be significantly higher than the prior three years to support multi-year customer demand driven by AI. Although TSMC did not provide 2027 capital expenditure guidance, which is what the market was hoping for, the company expressed very strong confidence in the AI supercycle.
Executive Remarks
Wendell Huang — Chief Financial Officer
Good afternoon everyone, and thank you for joining our conference call today. I will begin my presentation with the financial highlights for the third quarter of 2026, followed by our guidance for the fourth quarter of 2026. In the third quarter, our revenue in NT dollar terms grew 17.6% sequentially, supported by strong demand for advanced process technologies. In US dollar terms, revenue grew 16.2% sequentially to reach US$46.7 billion, exceeding the upper end of our third-quarter guidance. Gross margin declined 1.3 percentage points sequentially to 66.4%, primarily due to margin dilution from the 2nm process ramp-up. Total operating expenses accounted for 7.8% of net revenue, with operating margin at 58.7%. Overall, our third-quarter earnings per share reached NT$29.43, with return on equity (ROE) at 45.5%.
Now let us review revenue by technology node. The 2nm process technology contributed 8% of wafer revenue in the third quarter, while 3nm, 5nm, and 7nm accounted for 33%, 28%, and 10% respectively. Advanced technologies, defined as 7nm and below, accounted for 79% of wafer revenue. Next, let us look at revenue contribution by platform. High Performance Computing (HPC) revenue grew 22% sequentially, accounting for 68% of third-quarter revenue. Smartphone revenue grew 10%, representing 21%. IoT revenue grew 5%, accounting for 4%. Automotive revenue grew 23%, representing 4%. Digital Consumer Electronics (DCE) revenue declined 12%, accounting for 1%.
Turning to the balance sheet. As of the end of the third quarter, we held NT$4 trillion in cash and marketable securities, equivalent to US$126 billion. On the liability side, current liabilities increased by NT$110 billion sequentially, primarily due to a NT$46 billion increase in the current portion of bonds payable and bank loans, and a NT$26 billion increase in dividends payable. In terms of financial metrics, accounts receivable turnover days decreased by one day to 28 days. Inventory turnover days decreased by five days to 82 days, mainly due to increased wafer shipments.
Regarding cash flow and capital expenditure, in the third quarter we generated approximately NT$920 billion in cash from operating activities, with capital expenditure of NT$443 billion, and paid NT$156 billion for the fourth-quarter 2025 cash dividend. Overall, our cash balance increased by NT$488 billion during the quarter to reach NT$3.6 trillion. In US dollar terms, total capital expenditure for the third quarter was US$13.8 billion. This concludes my financial summary. Now let us look at our guidance for the current quarter. Based on our current business outlook, we expect fourth-quarter revenue to be between US$52.3 billion and US$54.3 billion. At the midpoint, this implies approximately 14% sequential growth and 58% year-over-year growth. Based on an exchange rate assumption of US$1 to NT$31.8, we expect gross margin to be between 65% and 67%, and operating margin between 56.5% and 58.5%.
Now let me cover our key messages. I will first address profitability for the third and fourth quarters of 2026. Compared with the second quarter, our third-quarter gross margin declined 130 basis points sequentially to 66.4%, primarily due to margin dilution from the initial ramp-up of 2nm technology. The fourth-quarter gross margin guidance we just provided is 65% to 67%, with a midpoint of 66%, representing a 40 basis point sequential decline, mainly due to continued rapid ramp-up of 2nm technology and margin dilution from overseas fabs. These impacts are expected to be partially offset by cost improvement measures, including productivity enhancements and capacity optimization across process nodes. Although costs at overseas fabs remain elevated, thanks to the company's overall scale expansion and cost improvement measures, we now expect the gross margin dilution from overseas fab ramp-ups to be close to 2 percentage points.
To support strong structural customer demand, we are accelerating the construction of additional fabs, particularly in the United States. As a result, we expect the gross margin dilution from overseas fabs to expand to 3% to 4% in the coming years. We will leverage the growing scale of our Arizona fabs and continue to improve operations to optimize our cost structure. We will also continue to work closely with customers and suppliers to manage the related impacts. Overall, with our two fundamental competitive advantages of manufacturing technology leadership and large-scale production bases, we are confident that TSMC will continue to be the most cost-efficient manufacturer in every region where we do business.
Next, let me discuss the 2026 capital budget. At TSMC, a higher level of capital expenditure has always been associated with higher growth opportunities in the coming years. Given continued strong multi-year demand from our customers, including the emerging agentic AI market, we reiterate our full-year 2026 capital budget of US$60 billion to US$64 billion. We continue to expect that capital expenditure investments from 2026 to 2028 will be significantly higher than the past three years, as we continue to invest heavily to support customer growth. Furthermore, despite tight equipment supply, we always work closely with equipment suppliers well in advance to prepare capacity, just as our customers work with us in advance to plan capacity. Even with capital expenditure investments at this level in the coming years to support future growth, we remain committed to delivering profitable growth for our shareholders. We also remain committed to maintaining sustainable and steadily growing cash dividends per share on both an annual and quarterly basis.
C.C. Wei — Chairman and Chief Executive Officer
Good afternoon everyone. Let me begin with the near-term demand outlook. In the third quarter, our revenue reached US$46.7 billion, exceeding our prior guidance in US dollar terms, primarily driven by strong demand for advanced process technologies. Entering the fourth quarter of 2026, we expect our business to continue to be supported by strong demand for advanced process technologies, including the rapid ramp-up of 2nm technology. Looking ahead, we continue to observe that consumer and price-sensitive end markets are facing challenges due to rising component prices and macroeconomic uncertainty. Therefore, as we enter 2027, we remain cautious in our business planning while focusing on business fundamentals to further strengthen our competitive position. That said, supported by strong technology differentiation and a broader customer base, demand for our advanced process technologies remains extremely strong. As a result, we now expect full-year 2026 revenue in US dollar terms to grow approximately 44% year-over-year.
Next, let me discuss the AI-related demand outlook. The AI market remains very strong. Recent developments are very positive for the long-term AI demand outlook. As we shift from chatbots to AI assistants, we observe increasing adoption of agentic AI, which is driving AI demand growth. In addition, we have also observed high adoption rates of open-source AI models. This will improve efficiency, lower the barrier to future AI usage, and further drive demand for advanced chips. Therefore, AI-related demand is being supported by strong underlying demand for AI accelerators. Furthermore, the emergence of agentic AI CPUs and growing networking chip demand are also further driving chip demand. As a result, we believe semiconductors will continue to be an indispensable component of the AI supercycle. In our view, the value of semiconductors in the global supply chain will continue to increase. This will create more value capture opportunities for our customers and also create more value opportunities for TSMC. Regardless of which approach is adopted, the most advanced semiconductor process technologies will be required. With our technology leadership, excellent manufacturing capabilities, global capacity expansion, and customer trust, we are well-positioned to capture all the opportunities presented by AI-related multi-year structural demand. As a result, all AI companies are working with TSMC.
Next, let me discuss TSMC's capacity expansion strategy. As a foundry, our primary responsibility is to provide customers with the best technology and effective capacity support to unlock their innovation capabilities. To address the long-term structural growth in semiconductor market demand, we work closely with customers and our customers' customers to plan capacity together. Furthermore, we have now gone a step further by sending our own teams to check on data center construction progress and power supply conditions in the United States. The feedback so far has been very good, and cloud service providers (CSPs) are well prepared. Internally, TSMC employs a rigorous capacity planning system, assessing market demand from both top-down and bottom-up perspectives. This is an ongoing process. Given the inherent complexity of advanced process technologies, along with long design-in and delivery lead times, we already have a very clear understanding of our customers' multi-year product roadmaps and production plans. Therefore, we remain strongly confident in the AI supercycle. We are proceeding at full speed with capacity expansion to meet our assessed demand and increasing capital expenditure investments to support our customers' future growth.
TSMC is expanding at an unprecedented pace, with approximately 20 fabs currently under construction worldwide. In the coming years, we will build 13 advanced process and advanced packaging fabs in Taiwan, while also building another 5 to 6 fabs overseas. Our new process node N2 has already entered mass production in the fourth quarter of 2025, with good yield performance. N2 is ramping up smoothly in multiple phases at two locations, Hsinchu and Kaohsiung. We expect that by 2027, the annual capacity of the N2 family of technologies will more than double from current levels, supported by strong demand from smartphones and HPC/AI applications. We are also leveraging our excellent manufacturing capabilities to improve N2 wafer output, further enhancing the capacity support we provide to customers. In addition, we have already begun preparing capacity for A14 technology. Risk production in 2027 is progressing well, and mass production will begin in 2028. Therefore, TSMC's semiconductor technology and manufacturing capabilities will continue to play a key role in supporting the global semiconductor industry and unlocking customer innovation.
Now let me discuss the latest developments in TSMC's global N3 capacity expansion. As we have said before, historically, once a process node reaches target capacity, we typically do not add additional capacity. However, based on our assessment, we are smoothly executing our global capacity plan and adding three new fabs to support the strong multi-year demand pipeline for 3nm technology. In Taiwan, our new 3nm fab at the Tainan Science Park is progressing well, with mass production planned to begin in the first half of 2027. In addition, driven by strong customer demand, we have decided to add another 3nm fab in Tainan. This fab is planned to begin mass production in the first half of 2028. In Arizona, we have completed construction of our second fab and are currently moving in and installing equipment. This fab is planned to begin mass production in the third quarter of 2027, slightly ahead of our previous schedule. In Japan, we are on track to adopt 3nm technology at our second fab, with mass production planned for 2028. In addition to all these new fabs, we continue to convert 5nm equipment in Taiwan to support 3nm capacity. We are also leveraging our excellent manufacturing capabilities to improve production efficiency at fabs across all regions. To increase wafer output, we are also focusing on capacity optimization across different process nodes, including providing flexible capacity support among N7, N5, and N3 processes. Therefore, we are using multiple approaches to expand capacity as much as we can, wherever we can, and as quickly as possible to maximize support for our customers. We have no intention of leaving any business opportunity to others. We want to capture every piece of business we can. This concludes my key messages. Thank you.
Q&A Session
Laura Chen — Citi Analyst
Good afternoon. Thank you, C.C., for sharing TSMC's capacity expansion strategy. I would like to follow up further. Given increasingly aggressive AI semiconductor demand, how does TSMC currently view AI revenue CAGR growth? Additionally, how does TSMC reconcile customer capacity demand with potential concerns about data center power supply? Recently we have also seen some financing issues, as well as issues with other key components such as HBM, APS, and advanced packaging. On the networking equipment side, we have also seen tightening laser supply. If these trends lead to a mismatch between semiconductor supply and actual AI development, how will TSMC adjust capacity to minimize the risk of potential overbuilding in the supply chain? That is my first question.
C.C. Wei
That is a very, very long question. Laura, let me share with you. We have a very advanced capacity planning process, and we also adopt a very rigorous assessment approach. We take all possibilities into consideration, including CSPs, capital expenditure, data center construction, financial conditions, power supply, and other component supply, among others. We make our own judgment. Based on our assessment of demand, we still believe that demand exceeds supply. Therefore, we are working very hard to narrow this gap.
Laura Chen
So far, based on our comprehensive research of the supply chain, we remain quite optimistic about long-term demand. Under these circumstances, could you elaborate further? What is the latest change in AI CAGR growth? I recall we previously talked about approximately 50% plus.
C.C. Wei
Laura, I can only tell you now: demand is extremely strong. We have not seen any signs of slowing down. As for whether it is a 50% plus CAGR, that does not matter. It could be higher.
Laura Chen
My second question is also related to very strong demand and the capital expenditure cycle. We see that this round of capital expenditure cycle will clearly be much higher than before. TSMC previously also announced that it would adopt High-NA EUV around 2030. Should we believe that this will further improve production efficiency, thereby more significantly accelerating revenue growth over the longer term? Additionally, as capital expenditure intensity increases, how should we understand the company's pricing power, as well as long-term gross margin and return on capital efficiency?
C.C. Wei
Let me answer that. Yes, in terms of long-term production efficiency, we will continue to improve. High-NA is undoubtedly a very good piece of equipment. At present, I think our plan is to use it in 2030. Again, our capital expenditure will be much higher than today. You are asking whether gross margin or profitability can improve. Let me put it this way: there is no ceiling on TSMC's profitability, but a significant margin improvement will not happen overnight. Our philosophy or strategy has always been to work with our customers. We want to help our customers succeed, and then we can also capture our own value.
Gokul Hariharan — JPMorgan Analyst
My first question is still about capacity expansion. It appears there is a supply shortage in the market. C.C., I would like to hear your view. I think it has been a long time since we have seen TSMC increase capacity at a rate exceeding 10% per year. If you look at the company's annual reports, I believe there has not been capacity growth exceeding 10% for about 10 years. Given the clear acceleration in capacity construction, could we begin to see wafer capacity growth approaching or even exceeding 10% in the coming years? Could you comment on the planned wafer capacity growth over the next few years? That is my first question.
C.C. Wei
As I mentioned earlier, demand is extremely strong and continues to rise. Of course, we have our own judgment and look at it from all angles. As I said, we communicate with customers, with our customers' customers, and then we further check data center construction ourselves. We incorporate all results into our judgment and assessment. The result is: demand is very, very high. The gap between demand and supply remains quite large. That is why we are mobilizing all our forces to increase capacity, as I just said. As for whether the gap will widen or narrow, we are working very hard to address this. I cannot answer right now because our customers are still continuously raising their demand.
Wendell Huang
Okay, Gokul. I think it is still too early to discuss future capacity growth percentages. I would just like to remind everyone that while the annual growth figures may not be large, growth in advanced process nodes is actually the main driver of capacity increases.
Gokul Hariharan
Understood. Thank you. Secondly, I think you mentioned more aggressive overseas expansion plans, especially in the United States. There has also been a lot of discussion in the market recently, such as the possibility of building a new fab cluster in Texas. I think one of your customers also talked about discussions related to TSMC. C.C., could you talk about overseas plans, such as more accelerated capacity construction? Additionally, given that one of your customers mentioned certain cooperation with TSMC last week, is there contact with SpaceX or Tesla? What might the general framework of such cooperation look like? Is it similar to the cooperation you have with some customers in Japan and at ESMC?
C.C. Wei
Okay, Gokul. We are working very hard to expand capacity in the United States. This is necessary because most of our largest customers are in the United States, and they are asking TSMC to increase capacity. The US government is also very supportive. This is where we want to expand. Recently, we have further expanded our capacity assessment scope to study whether there are other expansion opportunities beyond Arizona. Texas is one of them. We have sent teams to Texas to conduct all relevant assessments. We are also in communication with a customer that is trying to build a fab, more specifically, Terafab. We are discussing with them to see whether we can also leverage the actual capacity from that project. Discussions are still ongoing. If we reach any definitive conclusions, we will share them with everyone.
Gokul Hariharan
Okay. But all these fabs will ultimately be operated by TSMC, correct? I do not think you would be part of the Terafab project, or something like that. Could you help us understand what principles you are applying here? Is it similar to the principles applied at JASM or ESMC?
C.C. Wei
We are still discussing. Let me put it this way: both approaches are possible. But the fab must be operated by TSMC.
Krish Sankar — TD Cowen Analyst
I have two questions. First, C.C., some customers have talked about a so-called "secondary wafer market," where one customer transfers its wafer quota to another customer. Is this just a very small part of the business? Is this specific to certain customers, or do you think this is becoming a new trend, where customers with weaker demand transfer part of their quota to customers with stronger demand?
C.C. Wei
If you are asking whether this is a new strategy or new approach, the answer is no. TSMC's principle is: we do not allow customers to sublease or auction capacity. However, in certain cases, customers can cooperate through mask licensing agreements. But this is not our common practice. It is very, very small in scale.
Krish Sankar
Then a quick follow-up. You did mention a High-NA agreement with ASML starting in 2030. I am curious because historically TSMC has been at the forefront of EUV adoption. But it appears that by the time TSMC adopts High-NA EUV in 2030 or 2031, some competitors may already be using this technology by then. Do you think this could create a competitive disadvantage? Or are you not concerned about competitors being ahead in High-NA EUV?
C.C. Wei
We always evaluate new equipment. If we want to use it for mass production, it must have: the best technology; very reasonable cost; and most importantly, the highest level of maturity. We say we will start using it in 2030 because we believe High-NA will be mature by then. We are also working closely with ASML, planning to establish 12-inch mask capability by 2031. By then, we believe the cost of High-NA will be more reasonable and suitable for TSMC to put into mass production. Let me emphasize: this is a very good piece of equipment. At smaller line widths, it does have some advantages. As for whether using it later than competitors would put TSMC at a disadvantage? No. Because we are confident that we remain the best company capable of providing the most advanced technology to our customers.
Charlie Chan — Morgan Stanley Analyst
I am very glad to hear that AI semiconductor demand remains very strong despite the recent stock market correction. My first question is indeed related to the company's unprecedented capacity expansion strategy. As I understand it, one of your IDM customers is actively expanding capacity. There are some concerns in the market that this customer may increase in-house production rather than outsourcing to TSMC. First, I am not sure how you will address this business opportunity. Second, I recall C.C. mentioned last time that A14 could become a larger node than N2. I am wondering whether this assumes that this large IDM company will outsource more CPUs, such as server CPUs, at the A14 node?
C.C. Wei
Charlie, we treat every customer equally. Whether it is an IDM customer or a pure fabless customer, we treat them the same. Of course, we also consider the customer's own internal capacity. Based on our own judgment, we calculate the support we can provide and the scale of capacity construction. Actually, regarding A14, so far we have seen more and more customers engaging in A14 activities, at a scale already higher than 2nm activities. We expect that A14 will become a larger node than the current N2 technology.
Charlie Chan
I see. That is to say, without assuming that this IDM company will also increase outsourcing, we can still see A14 becoming a larger node than N2. Is my understanding correct?
C.C. Wei
Charlie, as I said, frankly, I will not comment on specific customers. But let me assure you: we will take all factors into consideration.
Charlie Chan
My second question is a follow-up on advanced packaging. From what we have observed, reticle sizes are becoming increasingly larger. C.C., you mentioned last time, or about two quarters ago, TSMC's solutions including SoW, 3.5D, etc., so there are many projects currently underway. My question is: do you think these SoW 3.5D and other solutions can really satisfy customers who are rapidly developing larger reticle size chips, such as 14x reticle chips? Secondly, if a customer initially chooses Intel's EMIB-T but for some reason cannot obtain sufficient supply, will TSMC also provide CoWoS solutions or capacity for these customers?
C.C. Wei
Okay, Charlie. Actually, I still believe TSMC's advanced assembly, the so-called advanced packaging technology, is very competitive. But most importantly, we use these technologies to help our customers' products successfully enter the market. That is our first priority. We must provide the best technology and sufficient capacity. At present, I can tell you one actual situation: we have already achieved 5x reticle size in actual production. I do not think any other company besides TSMC can do this currently, and our yield is excellent. Therefore, our customers are very satisfied with the results. We are advancing toward 9x reticle size and continuing to move toward 14x reticle size. Of course, we are also advancing CoWoS to ensure reasonable cost reductions can be achieved in the future. However, at present: CoWoS remains customers' first choice. Over time, yes, our competitors are offering very competitive solutions. But I still want to emphasize: currently, we are the only company that can ensure mass production of 5x reticle size advanced packaging, and we remain very competitive. I want to emphasize this point.
Sunny Lin — UBS Analyst
My first question is, I hope to further understand how we should think about capital expenditure beyond 2026. I recall that a few quarters ago, you mentioned that if you do a good job, you could potentially make capital expenditure growth slightly lower than revenue growth. But now clearly, overall customer activity may be further strengthening, and you are also willing to support customers in more large-scale expansion. Therefore, I would like to know: at present, could capital expenditure potentially grow in line with sales growth, or even slightly faster than sales growth, as in 2026? At the same time, due to very tight supply, equipment prices are also rising. Therefore, entering 2027, will equipment price increases also become an important factor pushing up capital expenditure?
Wendell Huang
Okay, Sunny. It is still too early to discuss capital expenditure figures beyond 2026. As we said in our prepared remarks, we have very strong confidence in the AI supercycle. Therefore, we expect capital expenditure from 2026 to 2028 to be significantly higher than the past three years. This is to increase investment levels to support customer growth. We have no intention of leaving any business opportunity to others. Equipment price inflation is real. Equipment prices, materials, labor costs, and so on are all rising. However, most capital expenditure is always related to long-term demand. The composition of future capital expenditure will also be the same. In terms of capital expenditure intensity, as we have always said: if we do a good job, revenue growth will exceed capital expenditure growth. Based on what we are seeing now, this still holds true.
Sunny Lin
If I may, I would like to quickly follow up on capital expenditure. Earlier, Wendell mentioned that due to accelerated overseas expansion, the gross margin dilution from overseas operations will reach 3% to 4% starting from 2027. Regarding the United States, you are currently still mainly focused on Arizona. Does this mean that the second phase of expansion in Arizona will be very rapid? Or are you actually accelerating the third phase or even other phases of expansion? Therefore, causing further expansion of gross margin dilution next year?
Wendell Huang
As we said, we are doing everything possible to expand capacity wherever we can to meet strong customer demand. We are also accelerating in various regions, including the United States. We previously mentioned that in the latter half of the five-year period, the gross margin dilution from overseas fabs will reach 3 to 4 percentage points. This still holds true. Therefore, in the latter half of the five-year period, the dilution from overseas fabs will reach 3 to 4 percentage points.
Haas Liu — Bank of America Analyst
My first question is about your view on potential structural changes in AI demand. You mentioned in your prepared remarks that AI accelerator chip demand is expanding toward agentic AI demand, including CPUs and networking switch chips. We understand that overall AI investment remains very strong, and trends across various areas are even further strengthening. Could we get further insight into whether you are seeing customer demand shifting toward more efficient computing solutions? Especially as commercialization becomes increasingly important. How should we understand the incremental opportunities this brings to manufacturing and packaging businesses?
C.C. Wei
I can only say: AI demand is extremely strong. As for whether the demand comes from accelerators or more from the CPU side, that is not something I can comment on. However, we do see continuously growing demand for agentic AI CPUs. Yes, this is a current trend. Beyond that, everything is very strong. Let me emphasize again: based on our judgment, we have considered all parameters, and currently demand is actually still far above what we are currently able to provide. We are working very hard to narrow this gap. That is all I can say.
Haas Liu
My second question is about the yield learning curve under the future dual-technology roadmap strategy. Historically, I think you have been improving new process node yields by producing chips with smaller die sizes and simpler designs, thereby improving the learning curve for new processes. But for future process node families that simultaneously support traditional power delivery and Super Power Rail technology: traditional power delivery is mainly used for mobile devices and consumer products; Super Power Rail is used for HPC and AI chips. If both process architectures are run simultaneously, would it make the learning curve more difficult, or increase manufacturing complexity? How should we understand the impact on gross margin, including yield and cost control?
C.C. Wei
Actually, the current situation is somewhat different from the past few years. Yes, we are now facing very strong demand from both smartphones and AI/HPC applications simultaneously. Therefore, small die sizes and large accelerator die sizes are happening at the same time. Let me say: this is favorable for TSMC. Because we are good at this. We are well prepared for this situation, and customers are happy to work with us. Therefore, the learning curve certainly exists. But we are doing quite well. So customers are currently very satisfied.
Arthur Lai — Macquarie Analyst
My question is related to power as a constraint. I would like to understand: in the past, have grid constraints ever caused you to lower customer forecasts? Additionally, as grid constraints become increasingly severe, could this potentially prompt customers to accelerate adoption of Super Power Rail products to achieve better power performance?
C.C. Wei
Undoubtedly, power supply is a consideration we are monitoring. This is also why we further verified the power supply situation in the United States. We have completed the relevant study. So far: the situation is good. As for whether this will change customer behavior, pushing them more toward backside power rail technology? The answer is: no. Whether to use backside power delivery and such metallization schemes depends on the product's own requirements. Some accelerators may use it. Some products, such as phones, will not use it. So it depends on the customer's product.
Arthur Lai
Coming back to the earlier discussion about agentic AI CPU demand. I would like to know: under our previous demand forecast framework, does the 50% plus growth include CPUs? If we see customers significantly raising CPU demand forecasts, when will the AI demand growth data be updated next time?
C.C. Wei
CPUs are not included in our AI demand calculation. The reason is very simple: we cannot determine whether a particular CPU user is a traditional server CPU or an AI server CPU. Additionally, it depends on the customer's application and the customer's own capacity allocation. So, currently, no. As of today, we still adhere to our original very narrow definition of AI demand.
Stifel Financial Analyst
Entering 2027, N3 begins to contribute incremental profit, and that phase is now behind us. N2 is still in the margin-dilutive ramp-up phase, and overseas margin dilution is expanding. On the offsetting side, there have been recent reports that your mature process and advanced process prices have both seen significant increases. Could you provide further details? If prices have indeed changed, do they generally apply to existing orders with signed contracts, or mainly to new capacity beyond what customers have already committed?
C.C. Wei
As we have discussed before: our pricing strategy is strategic. Pricing also depends on: market segmentation; geographic location; technology; and other factors. Once we provide flexibility in physical production location, such as manufacturing in the United States, then the price will differ from products we manufacture in Taiwan. Regardless, we strive to have customers recognize our value and maximize the value our shareholders receive from the value created by our technology.
Sahej Singh
My second question is about a statement you made in your prepared remarks. That statement left a deep impression on me: "We will not leave any business opportunity on the table." About one or two weeks ago, GlobalFoundries announced it would provide silicon interposers to TSMC. I would like to understand: in these strategic decisions, how do you consider the relationship between sourcing from specialty foundries versus providing in-house?
C.C. Wei
We designed this cooperation agreement with GlobalFoundries as part of our overall back-end supply strategy. Actually, our first priority is to help customers succeed in the market. We want customers to be fully supported in all supply areas, including front-end and back-end. Currently, there is indeed a certain degree of supply shortage in the back-end. Therefore, we want to cooperate with our so-called partner GlobalFoundries. In this way, both TSMC and GlobalFoundries can support TSMC customers' products, ensuring sufficient capacity support when products enter the market.
Comments