The AI mega trend is still at a very early stage! TSMC (TSM.US) shares the latest outlook at the Goldman Sachs conference: Agents will expand computing power demand to CPUs, and N3 and CoWoS production capacity remain extremely tight.

According to Zhihui Finance APP, recently, TSMC (TSM.US) Chief Financial Officer Wendell Huang appeared at the Goldman Sachs Communacopia + Technology Conference to share his insights on the prospects of the artificial intelligence (AI) industry as well as the company's internal operations.

Goldman Sachs believes there are three key takeaways from this meeting: TSMC believes AI is still in the very early stages of a multi-year megatrend, and AI agents will expand demand from accelerators to CPUs; the Arizona plant project is progressing smoothly, with the first phase yield already comparable to its Taiwan facilities and operating profitably; though N2 process mass production ramp-up and overseas expansion may dilute gross margin in the short term, management remains confident in achieving its long-term gross margin target of over 56%.

After the meeting, Goldman Sachs maintained its "Buy" rating for TSMC, with a 12-month Taiwan equity target price of NT$3,100 and an ADR target price of $620.

AI Still in Early Stages of Multi-Year Trend, AI Agents Driving Computing Demand Toward CPUs

TSMC management continues to view AI as a multi-year megatrend, emphasizing it is still at an extremely early stage. Management pointed out that AI is shifting from generative AI to AI agents, and this will cause computing demand to further expand from GPUs and custom AI accelerators to CPUs.

TSMC not only deals directly with downstream customers but also communicates with their downstream customers, and based on these interactions, the company is confident about the sustainability of the AI industry cycle. Management also revealed that recent on-site communications with U.S. cloud service providers (CSPs) further strengthened this confidence: cloud providers are well prepared in terms of power supply and overall infrastructure deployment.

Additionally, TSMC management believes that open-source models combined with continuous improvements in computational efficiency will lower token costs, driving the expansion and adoption of AI applications, thereby supporting incremental demand for advanced chips. TSMC has incorporated this into its capacity planning and continues to actively expand its capacity.

Although the management acknowledges the semiconductor industry remains cyclical, they emphasize that it is more important to identify underlying long-term trends than to predict each round of short-term fluctuations.

Arizona Plant Is Not a Test: First Phase Profitable, One Third N2+ Capacity Abroad in Five Years

TSMC reiterated the two main motivations for building overseas: customers' demand for supply chain diversification and the opportunity to obtain moderate government support.

It was disclosed that the first phase of the Arizona plant, which uses N4 technology, is now operational with a yield rate comparable to Taiwan’s facilities, meeting quality and reliability standards, and is already profitable. Management sees this as an important milestone, proving TSMC can manufacture wafers in the U.S. with the same quality as in Taiwan.

TSMC stated that the second phase will use N3 technology, equipment move-in is expected to start very soon; the third phase construction has already begun; the fourth phase and the first advanced packaging facility in the U.S. are also advancing simultaneously.

Beyond current plans, TSMC has also acquired additional land to accommodate five to six more wafer fabs, supporting the recently announced extra US$100 billion investment in the U.S.

Under current plans, about 30% of N2 and more advanced process capacity will be located outside Taiwan within five years, with the bulk in Arizona. However, the most advanced technologies will still ramp up first in Taiwan, and overseas production lines will only be introduced once the processes are mature and stable.

Management explained that customers initially only requested small-scale regional supply diversification, but as the Arizona project has proven successful, customer demands keep increasing. TSMC believes that diversified regional supply chains represent additional value, which they plan to reflect in product pricing, and customers are increasingly willing to pay a premium for it.

N3 Still in Short Supply, Value-Based Pricing, Gross Margin Target Over 56%

TSMC revealed that demand for the N3 process is particularly strong. Unlike previous nodes where demand typically began falling two to three years after launch, TSMC is still increasing N3 capacity and expects supply to remain insufficient. Advanced packaging capacity is also tight, with CoWoS capacity still under strain, and TSMC has started outsourcing some packaging processes to OSAT partners.

Looking at the entire AI supply chain, TSMC believes power supply and memory may become potential bottlenecks, while capacity expansion for foundry itself is highly challenging: a single factory takes 2-3 years to build, and another 1-2 years for mass production ramp-up.

Management emphasized that the primary goal of developing advanced packaging is to ensure and drive wafer demand, not to make packaging an independent profit center. In terms of pricing, TSMC does not intend to raise prices just because utilization is high in the short term, nor lower them if utilization temporarily weakens. The pricing principle is to achieve sustainable and stable quotations based on the value TSMC provides.

Regarding gross margins, N2 mass production ramp-up is expected to dilute 2026 gross margin by about 2-3 percentage points, overseas expansion in the current five-year period will dilute another 2-3 points early on, later expanding to 3 points as more overseas fabs come online. However, management remains highly confident that disciplined capacity planning, healthy utilization rates, cost reduction, productivity improvements, and a value-reflecting pricing strategy will support the company’s long-term gross margin goal of “56% and above,” with the aim of reaching higher ranges.

Finally, TSMC management reiterated that AI revenue as disclosed in its financial statements is defined narrowly, mainly including GPUs, custom AI accelerators, and memory-based bare dies; CPUs and networking chips are not included as it’s not possible to accurately determine if these chips are ultimately used in AI systems. As AI agents drive incremental demand for x86, Arm, RISC-V architecture CPUs, networking, and other supporting chips, TSMC’s actual AI-related business scale is significantly larger than the AI accelerator revenue disclosed in reports. Until the company can more precisely determine chip end-use, it will maintain its current statistical criteria unchanged.