The AI trend is still in its early stages! TSM.US (TSM.US) Goldman Sachs conference released the latest judgment: smart devices will expand computing power requirements to CPUs, and production capacity for N3 and CoWoS is still very tight

The Zhitong Finance App learned that recently, TSM.US (TSM.US) Chief Financial Officer Wendell Huang (Wendell Huang) appeared at the Goldman Sachs Communacopia+ Technology Communications Technology Conference to share information on the prospects of the artificial intelligence (AI) industry and the company's internal operations.

Goldman Sachs believes that the core message of this conference is three points: TSMC believes that AI is still in the very early stages of a multi-year trend, and AI agents will expand demand from accelerators to CPUs; the Arizona factory project is progressing smoothly, and the first phase yield is comparable to its factory area in Taiwan Province of China and has achieved profit; although the N2 process's steep rise in mass production and overseas expansion will dilute gross profit margin in the short term, management is still confident that the long-term gross margin target of 56% or more will be achieved.

After the meeting, Goldman Sachs continued to maintain TSMC's “buy” rating. The 12-month target price for Taiwanese stocks was NT$3,100, and the target price for ADR was $620.

AI is still in the early stages of a major trend for many years, and AI agents are pushing computing power requirements to CPUs

TSMC management continues to view AI as a multi-year trend and emphasizes that it is still in its very early stages. Management pointed out that AI is shifting from generative AI to AI agents, which will further expand computing requirements from GPUs and custom AI accelerators to CPUs.

TSMC not only directly connects with downstream customers, but also communicates with its downstream customers. Based on these exchanges, the company is confident in the sustainability of the AI industry cycle. Management also revealed that recent field exchanges with US cloud service providers (CSPs) have further strengthened this confidence: cloud service providers are fully prepared in terms of power packages and overall infrastructure deployment.

Furthermore, TSMC management believes that the continuous improvement of the open source model's superposition calculation efficiency will reduce token (token) costs, promote the expansion of the scale and popularity of AI applications, and thus support the incremental demand for advanced chips. TSMC has now incorporated this into its production capacity plan and continues to actively expand production.

Although management acknowledges that the semiconductor industry is still cyclical, it is more critical to identify underlying long-term trends than anticipating each round of short-term fluctuations.

Arizona is not testing the waters: the first phase is already profitable, 30% of N2+ production capacity is overseas in five years

TSMC reaffirms the two main starting points for overseas factory construction: customer demand for geographical diversification of the supply chain and receiving appropriate government support.

According to reports, the first phase of the Arizona plant uses N4 technology and is now in operation. The yield is comparable to that of the plant in Taiwan Province, the quality and reliability standards have been met, and profits have been achieved. Management sees this as an important milestone, proving that TSMC can produce wafers of the same quality in the US as Taiwan Province.

TSMC said that the Phase II plant will use N3 technology, and equipment moving in is expected to begin soon; Phase III construction has also begun; Phase IV and the first advanced packaging facility in the US are progressing simultaneously.

In addition to existing plans, TSMC has also purchased new land to accommodate 5 to 6 additional fabs to support its recently announced additional $100 billion US investment.

According to the current plan, within five years, about 30% of N2 and more advanced process production capacity will be located outside of Taiwan Province, and the vast majority will be located in Arizona. However, the most advanced technology will still climb the slope first in Taiwan Province, and overseas production lines generally have to wait until the process is mature and stable before following up implementation.

Management said that initially the customer only requested small-scale geographically distributed supply requirements, but with the successful verification of the Arizona project, customer related demands continued to rise. TSMC believes that diversification of regional supply chains has added value, and plans to reflect this value in product pricing, and customers are increasingly willing to pay premiums.

N3 is still in short supply. Pricing depends on value, gross margin target is above 56%

TSMC revealed that demand for the N3 process is particularly strong. Unlike in the past, where demand usually began to decline two or three years after the launch of the node, TSMC is still increasing N3 production capacity, and the supply is expected to remain insufficient. Advanced packaging production capacity is also tight. CoWoS is still tight, and TSMC has begun to outsource part of the packaging process to OSAT partners.

Looking at the entire AI supply chain, TSMC believes that power supply and memory will become potential bottlenecks; however, it is very difficult to expand foundry production capacity itself: it takes 2 to 3 years to build a factory, and 1 to 2 years for subsequent mass production to climb.

Management emphasized that the main purpose of developing advanced packaging is to secure and drive wafer demand, rather than making packaging an independent profitable sector. In terms of pricing, TSMC does not plan to increase prices due to high short-term utilization rates, nor to reduce prices due to temporary weakening of the operating rate. The pricing principle is to achieve a sustainable and stable quotation based on the value provided by itself.

At the gross margin level, mass production of N2 is expected to dilute the gross margin by about 2 to 3 percentage points in 2026. Overseas expansion is expected to dilute 2 to 3 percentage points in the early stages of the current five-year period, and then expand to 3 percentage points as more overseas fabs start production. However, management is still very convinced that disciplined capacity planning, healthy operating rates, cost reduction, productivity improvement, and pricing strategies reflecting TSMC's value will support the company to achieve the long-term gross margin target of “56% or more” and move closer to a higher range.

Finally, TSMC's management reiterated that the AI revenue definition disclosed in the financial report is narrow. The main statistics are GPUs, custom AI accelerators, and memory base chips; CPUs and network chips are not included in the statistics because it is impossible to accurately determine whether the chip will eventually be used in AI systems. As AI agents drive incremental demand for x86, Arm, RISC-V based CPUs, networks, and other supporting chips, TSMC's actual AI-related business scale significantly exceeds the AI accelerator revenue disclosed in the report. And until it is possible to more accurately determine the use of chip terminals, the company will maintain the current statistical caliber.