[Text/Column Author from Guancha Observation Network]

In the eye of the hurricane in Sino-US tech competition, the global chip hegemon TSMC is being pushed toward an unprecedented strategic choice. The blade of Trump's tariff war has reached the throat of the global semiconductor supply chain - as the "silicon shield" accounting for 60% of the global foundry market share, TSMC's revenue from processes below 7nm has already accounted for 90%, leaving its competitors far behind, but this silicon shield must make enough gestures to ease America's anxiety over the return of manufacturing, submitting to Washington's "chip iron curtain."

TSMC's 2024 annual report and the Q1 earnings conference call provided a small but significant annotation for this "crossroads" moment.

Arizona plant, a thunderstorm

TSMC's latest financial report shows that the new factory in Arizona incurred a loss of approximately 3.21 billion RMB, making it the most loss-making overseas facility.

TSMC President Zijia Wei recently announced under Trump's pressure that the company would invest an additional $10 billion in Arizona, bringing the total investment in the U.S. to $16.5 billion. This initial investment may lower the gross margin of the site by five percentage points within five years.

Beyond that, it should be noted that TSMC claims that the first factory has achieved N4 mass production; the second factory is currently in operation, using N3 technology, which has been completed and will accelerate the production process. TSMC will also build two advanced packaging facilities and one research and development center, and plans for more than 30% of the capacity above 2nm to be located in Arizona.

In the Q&A session of the earnings conference, TSMC exposed the "fox tail" of the Arizona research and development center, stating that it will have more than 1,000 engineers with the aim of optimizing overseas factory technology to reduce operating costs, meaning this center serves the production process rather than researching cutting-edge technologies. These 1,000 engineers are likely mostly quality process administrators (QE/PE) or customer service CSE categories.

TSMC Arizona Plant

The more advanced the process nodes, the higher the upfront investment, and it takes time for capacity to ramp up. Given that TSMC's Japan Kumamoto plant only began contributing capacity at the end of last year and suffered severe losses, it is normal for the Arizona plant to incur consecutive deficits. However, since 2021, the annual losses of the Arizona plant have increased nearly by 400 million NT dollars each year, with a loss of 1.43 billion NT dollars last year, accumulating to nearly 4 billion NT dollars, which has caused varying degrees of concern among TSMC's small and medium shareholders.

However, analyst Lin Junji from the island calculated for TSMC.

He calculated that if the Arizona plant produces 10,000 wafers per month in 2025, the annual output will reach 120,000 wafers. With Taiwan's factory pricing for 4nm at around $15,000-$16,000 per wafer, conservatively assuming a 20% markup for the Arizona plant, the revenue would be $2.4 billion. Therefore, the revenue of the factory in 2025 would be between $2.16 billion and $2.3 billion, which is sufficient to cover the costs from 2021 to 2025.

Based on this calculation model, the capacity planning of the Arizona plant just needs to ensure a monthly output of 10,000 to amortize the initial R&D costs. However, given that TSMC's U.S. plant capacity planning is significantly higher than this figure, as long as production scales up, turning losses into profits in the near future is not difficult.

Regarding this issue, Mindset Observatorium contacted renowned analyst Zihao Wu, who has long been following TSMC. He pointed out that utilization rate and equipment depreciation are important indicators for measuring the financial performance of TSMC as a pure foundry, and cannot be ignored.

He stated that the Arizona plant focuses on the 4nm process node. Assuming a monthly production capacity of 30,000 wafers, the initial capital investment would be $13.5 billion, divided over five years of depreciation, approximately $2.7 billion annually. A revenue of around $3.1 billion for 200,000 wafers produced annually might be the break-even point for the Arizona plant.

As for utilization rate, Wu Zihao pointed out: "There is a simple logic in the semiconductor foundry industry: a factory undergoing depreciation must maintain an 80% utilization rate to be profitable. For a 30,000-wafer factory, 80% means 24,000 wafers per month, and 70% could mean 21,000 wafers."

Although TSMC is confident about the utilization rate of the Arizona plant, demand determines the market, especially under the "tariff chaos." The profitability model of TSMC's U.S. plant has become somewhat "mystical."

The Hidden Concerns of High-Performance AI Chips

Let’s look at the regional breakdown of TSMC's current business segments. In 2024, North American revenue accounted for 77%, increasing by 9 percentage points year-over-year, while China accounted for 7%, and Japan accounted for 3%.

We can compare this with four years ago, in 2021, when TSMC's regional revenue breakdown was as follows:

American companies have grown significantly, seemingly taking away 10 percentage points of the share from mainland China. However, there are several issues that need clarification here.

Firstly, this calculation method basically follows the headquarters location of the client companies. Apple, Nvidia, AMD, Qualcomm, Broadcom, and other major clients are all based in the U.S., so these few can almost jointly contribute nearly 80% of TSMC's revenue. Thus, this regional breakdown does not truly reflect the final destination of chip production capacity and market demand.

Secondly, this percentage only reflects the "price," but cannot intuitively show the "quantity." It is necessary to consider the fact that in 2024, for the first time in the past four years, the growth rate of mainland China's chip design companies fell below the global average. The reason lies in the absolute leadership position of several U.S. companies in HPC/AI accelerator chips. These high-value-added chips are TSMC's "profit treasure" and overall blood supply source. If AI chips are excluded, the growth rate of mainland China's Fabless companies is still leading the global average.

This can also be verified from the application scenarios. Four years ago, in 2021, approximately half of TSMC's foundry revenue came from mobile phone chips (main customers were Apple, MediaTek, and Qualcomm), while HPC chips accounted for about 30%. In just four years, the proportion of these two types of chip foundries has nearly reversed.

Currently, TSMC's HPC chip share is approaching 60%. Besides Nvidia's continuous generations of flagship AI accelerators, the CoWoS technology, which dominates advanced packaging, mainly targets Nvidia.

TSMC and Nvidia have virtually formed an unbreakable "fate alliance."

How long can Nvidia and Apple sustain their relationship with TSMC?

Invisible knowledge puzzles and fragmented appearances often mask the resonance of underlying logic. Behind the fragmented narrative of the world, there are usually cross-dimensional order laws. Since 2025, leading mainland AI chip companies have been collectively making efforts, forming breakthrough matrices. A non-American supply chain is accelerating to surface under the "tariff chaos."

Recently, Huawei Cloud's newly launched AI computing cluster solution, CloudMatrix 384, has surpassed Nvidia's flagship product GB200 NVL72 in multiple key metrics due to its disruptive system architecture design and full-stack technological innovation. Even SemiAnalysis, a semiconductor analysis firm that once underestimated DeepSeek, had to admit that Huawei CloudMatrix 384, built on 384 Ascend chips, achieves efficient chip-to-chip collaboration through a fully interconnected topology, providing up to 300 PFLOPs of dense BF16 computing power, close to twice that of Nvidia's GB200 NVL72 system.

Another high-performance computing chip company, Hygon, showed a sudden increase of 3.3 billion yuan in contract liabilities in its financial statements, indicating that the company's "Deep 3" is no longer hidden and has received large orders. Hygon inherited AMD's hardware and software architecture and has its own switch card interconnection technology. According to multiple securities research reports, "this may be one of the few truly usable technologies comparable to NVLink currently available domestically."

The "chip first stock" Cambricon's financial report is even more shocking. First-quarter revenue was nearly 1.2 billion yuan, almost equal to the total for the entire 2024. With 2.7 billion yuan in inventory and 900 million yuan in advance payments, it indicates that Cambricon is accelerating tape-out responses to domestic customer demands, expecting a big breakout in the second quarter of this year.

Nvidia's marginal effect on TSMC is weakening, and another major client, Apple, is also having a tough time. Apple's fiscal first quarter 2025 financial report showed a significant decline of 11% in revenue from the Chinese market, reaching $18.513 billion. Moreover, the packaging phase of Apple's chips processed by TSMC still needs to be entirely completed in Taiwan, which may be interrupted under geopolitical tensions, leading to reduced supply chain efficiency.

Conclusion: The "Bottomless Pit" After 3nm

Just as this article went to press, TSMC released its 1.4-nanometer semiconductor engineering technology A14 at the North American Technology Symposium in the U.S., promising significant improvements in performance, power consumption, and transistor density compared to its N2 (2-nanometer class) process.

However, when the daily wear cost of a single EUV lithography machine reaches $300,000, a fluctuation of 1% in yield can consume 5% of quarterly profit (calculated based on Apple A18 chip orders). More fatally, the top three customers contribute 62% of revenue; if a major customer shifts orders to Samsung or builds its own fabs due to technical flaws, TSMC faces the risk of billions of dollars in idle capacity. The ultimate outcome of this gamble may be: the more one pursues the pinnacle of manufacturing processes, the deeper one falls into the financial curse of "high input-low elasticity-strong dependency."

Perhaps, Zijia Wei, bowing humbly before Trump while incessantly rolling his eyes, hides countless self-denials beneath his outward composure—his body language may have already written the script for TSMC's future fate.

This article is an exclusive piece from the Observer Network. The content purely reflects the author's personal views and does not represent the platform's stance. Unauthorized reproduction is prohibited, and legal action will be taken for violations. Follow the Observer Network WeChat account guanchacn for daily interesting articles.

Original Source: https://www.toutiao.com/article/7498152524780798504/

Disclaimer: This article solely represents the author's viewpoint, and you are welcome to express your opinion by clicking the "Top/Downvote" button below.