The 1.4nm battle begins!

In semiconductor manufacturing, 2nm can be regarded as an important milestone that encompasses advanced processes, EUV clusters, GAA transistors, advanced packaging, supply chains, and geopolitical factors, viewed by various countries as a threshold for computing power sovereignty in the AI era.

Currently, a global competition centered around "building 2nm fabs" is unfolding, involving capital and national strategies: TSMC is ramping up its 2nm fab layout in Taiwan while advancing overseas projects in Arizona, USA, Kumamoto, Japan, and Dresden, Germany; Intel is attempting to reshape the foundry landscape with its 18A process and backing from a "national team of shareholders"; Samsung continues to catch up in 2nm yield and customer structure; and Japan's Rapidus, under policy support, is betting on single-wafer processes, providing a "restart" opportunity for Japan's semiconductor industry, which has been in decline for years.

TSMC Plans to Build 10 2nm Fabs

According to the latest reports from Taiwanese media, TSMC's 2nm layout in Taiwan has upgraded from the concept of "seven fabs" to "ten fabs"—two in Hsinchu Baoshan, five in Kaohsiung Nanzi, plus three planned in the Southern Taiwan Science Park, totaling 10 2nm fabs. It is estimated that the cost of a single 2nm fab is approximately NT$300 billion (US$8-10 billion). Therefore, the additional three fabs would require about NT$900 billion.

Recent public information also indirectly supports this speculation:

• TrendForce cited reports from Taiwanese media such as the Liberty Times, stating that due to the surge in AI chip orders, TSMC's existing seven advanced process fabs in Hsinchu and Nanzi are struggling to meet demand, and it is evaluating the construction of up to 12 new advanced process and packaging fabs in Taiwan, with a focus on 2nm and 1.4nm.
• According to reports from the Central News Agency of Taiwan, the Central Taiwan Science Park has confirmed that TSMC's 1.4nm (A14) fab in the Taichung A14 area has obtained construction permits and is set to begin construction by the end of 2025, aiming for mass production by 2028. This means that Taiwan will seamlessly transition to the next node after 2nm.
• Previously, TSMC had planned to build five new fabs in the Kaohsiung Nanzi new park for 2nm, A16, and more advanced processes, gradually transforming southern Taiwan into another hub for advanced processes, in conjunction with the supply chain base in the Pingtung Science Park.

In addition to continuing to add 2nm and 1.4nm in Taiwan, TSMC is also actively expanding overseas: In March of this year, TSMC announced plans to increase its investment in three fabs and two advanced packaging facilities, along with a large R&D center in Arizona, USA, to a total of US$165 billion. When announcing the increased investment in the U.S., CEO C.C. Wei emphasized, "All AI customers are reaching out, and Taiwan's capacity is still severely insufficient," hoping the government will continue to assist in finding land, electricity, and water locally.

From a business logic perspective, TSMC's basic judgment on 2nm is very clear: First, advanced processes can only expand production for key customers. The 2nm family will serve AI GPUs/accelerators (NVIDIA, AMD, self-developed ASICs), high-end CPUs/GPUs/APs, and some high-end mobile SoCs for a long time. Even with macro demand fluctuations, these customers have enough bargaining power to secure long-term capacity.

Secondly, the most advanced nodes must be held in Taiwan. Overseas manufacturers mainly "catch up" politically and in customer relationships: obtaining subsidies in the U.S., Japan, and Europe, and locking in local automotive/industrial customers, but the actual technological focus remains on Taiwan's 2nm/1.4nm clusters. This is why the mass production pace of 2nm is prioritized in Baoshan, Nanzi, and Nankang, rather than in Arizona.

From an industrial competition perspective, TSMC's aggressive factory construction is both a passive response to the unexpected demand for AI and an active defense against a round of "policy + capital attacks" from Samsung, Intel, Rapidus, etc.: when you firmly nail down 2nm capacity deep in the moat, even if competitors receive subsidies and large customers, it is difficult to shake the ecological stickiness in the short term.

Intel: A Do-or-Die Battle with 18A + National Team Capital

In the narrative of 2nm, Intel's 18A is technically comparable to 2nm. 18A is Intel's new generation GAA (RibbonFET) node following 20A, adding backside power delivery PowerVia. From the public roadmap and third-party breakdowns, 18A has reached a position where it can compete head-on with TSMC's N2 family in terms of transistor density, power consumption, and performance.

The biggest market doubt over the past year has been: can the 18A process smoothly ramp up to stable mass production? Recently, there have been relatively positive signals:

According to TechPowerUp, Intel Vice President John Pitzer stated that over the past seven to eight months, the yield curve for the company's 18A chips has steadily risen. Intel's technical experts also indicated at the previous Panther Lake conference that as of October this year, Intel's 18A process has officially started production. The current yield level of 18A is not lower than that of any previous core node, and yield improvement has entered a more predictable trajectory. It is expected to reach the yield threshold required for mass production by the fourth quarter of 2025, fully entering the high-capacity production phase.

Over the past year, Intel's 18A defect density has shown a stable downward trend, with overall performance improving.

A month ago, when Intel announced the mass production of Panther Lake processors, the company publicly acknowledged that all wafers required for the first batch of CPUs would be produced on the pilot production line in Oregon, with plans to gradually switch to the high-capacity Fab 52 in Arizona starting in 2026, leveraging scale effects to dilute costs and further stabilize yields If 18A is the "revenge round" on a technical level, then the changes in capital structure have given Intel a more obvious "national team" color:

• In August this year, the U.S. government converted the $11.1 billion subsidy originally from the CHIPS Act and the Secure Enclave program into equity, directly acquiring approximately 9.9% of Intel's shares (along with an additional 5% warrant), becoming the company's largest single shareholder.
• In September, NVIDIA announced a $5 billion subscription for Intel's common stock, with both companies collaborating on PC SoC and data center platforms. Although NVIDIA has not yet entrusted core GPU orders to Intel for manufacturing in the short term, this "mutual binding" action is essentially a strategic endorsement of Intel's manufacturing capabilities.

It can be anticipated that at the 2nm node, Intel's success or failure will not only depend on how good 18A itself is but also on whether it can "run out a truly foundry company"—to achieve a collaborative experience for designers in process, IP, packaging, and testing that is at least close to Taiwan Semiconductor.

Samsung: 2nm Yield Exceeds 60%, Accelerating Dual Base in Korea and the U.S.

In the past two years, Samsung has almost bet all its chips on the 3nm/2nm GAA nodes— the former is used to capture some high-end mobile and HPC customers, while the latter attempts to "position" itself in AI, automotive, and mining chip markets.

Samsung's latest announcement shows that its 2nm process (SF2) yield has climbed to the 55–60% range, significantly improving from the previous expectation of about 30%. Market research firm Counterpoint Research predicted on the 20th of this month that Samsung Electronics' 2nm capacity will increase by 163%, from 8,000 wafers per month in 2024 to 21,000 wafers per month by the end of next year. This capacity expansion follows the stabilization of Samsung's 2nm process yield.

Samsung's 2nm wafer fabs are mainly located in two places:

• One is in Hwaseong, South Korea, which has a well-established EUV foundation, with a highly concentrated R&D team and mass production engineers. Samsung is currently introducing the 2nm (SF2) production line into the Hwaseong S3 line. The reported 2nm yield of 55–60% is believed to mainly come from the data of the Hwaseong line;
• The other is in Taylor, Texas, USA. According to information disclosed by the U.S. Department of Commerce and Samsung, the total investment scale of this project is about $37–40 billion, regarded as one of the largest greenfield foreign investment projects in U.S. history. The Taylor plant was originally planned to complete major construction between 2024 and 2025, but due to fluctuations in global semiconductor demand and customer uncertainties, the project was reported to be delayed until 2026 for formal production;

The real turning point was securing the Tesla AI6 chip order. In July 2025, Samsung signed an $16.5 billion, 8-year foundry agreement for the AI6 chip with Tesla. The AI6 plan will use Samsung's 2nm node, produced at the Taylor plant, for Tesla's Robotaxi, Optimus robot, and other next-generation platforms The industry generally believes that this long-term large order is key to saving the Taylor project and enhancing Samsung's bargaining power in U.S. foundry services.

The Exynos 2600 is widely regarded as the first SoC to be mass-produced on the SF2 process, which will supply the Galaxy S26 series in 2026, with a yield rate of approximately over 50%. Among external vendor orders, in addition to Tesla, Samsung has also secured 2nm mining chip foundry orders from two Chinese cryptocurrency mining machine manufacturers (MicroBT and Canaan), with expected annual revenue reaching hundreds of millions of dollars.

From a financial and strategic perspective, Samsung's 2nm strategy has several characteristics:

Accumulating experience with "hard-to-please customers." Whether it's their own Exynos or clients like Tesla and mining machine manufacturers, these are customers that are extremely sensitive to power consumption, frequency, and yield, but are relatively willing to accept early risks. This is somewhat similar to how TSMC built its 7nm/5nm experience through the iPhone A series chips.

Exchanging profit pressure for capacity ramp-up. Before the yield is fully matured, 2nm orders are unlikely to bring high profits immediately, but they can pave the way for broader high-end customers after 2027-2028. Samsung has set a goal for its foundry business to "return to profitability within two years and increase market share to 20%," essentially betting that 2nm can outperform the industry average during this time window.

The problem is that the 2nm market itself is highly sticky: for major customers that are already deeply bound to TSMC's ecosystem, the cost of migrating the entire suite from PDK, IP combinations, to packaging and testing to Samsung is extremely high. Samsung's opportunities come more from three directions: first, emerging customers like cryptocurrency mining machines and specific AI ASICs that are willing to take risks; second, U.S. customers (such as Tesla) that are highly sensitive to geopolitical issues and have a "second source" demand; third, certain mainland customers that are extremely sensitive to cost-performance ratio and can accept early risks.

Rapidus: One 2nm and one 1.4nm and below

Compared to TSMC, Intel, and Samsung, Japan's Rapidus is much smaller in scale, but the government has high expectations for it in "building domestic 2nm capacity."

Source: Rapidus

In July of this year, Rapidus announced the start of trial production of the 2nm GAA test line at its IIM-1 factory in Hokkaido, with the first batch of test wafers meeting the expected electrical performance indicators.

According to public materials, at its first factory located in Chitose, Rapidus aims to begin mass production of 2nm chips in the second half of the 2027 fiscal year. Even though mass production of 2nm chips is not yet fully matured, Rapidus plans to quickly advance the construction of its second factory. Starting from the 2026 fiscal year, Rapidus will continue to collaborate with IBM (which provides technology related to 2nm chips) while launching comprehensive research and development for 1.4nm products Rapidus plans to start construction of its second factory in Hokkaido in the fiscal year 2027. This factory will not only produce 1.4nm products but may also manufacture 1nm chips, with production of 1.4nm chips expected to begin as early as 2029. The project is expected to cost trillions of yen, with funding for the second factory primarily coming from government support. The Japanese government will invest hundreds of billions of yen in the company, part of which will be used for research and development. The remaining funds will be raised through loans from major Japanese banks and investments from private enterprises. The relevant loans will be guaranteed by the government. The total investment for the second factory is expected to exceed 2 trillion yen.

Global chip manufacturers are competing to reduce chip line widths. Taiwan Semiconductor plans to mass-produce 2nm chips this year and 1.4nm chips in 2028. South Korea's Samsung Electronics plans to mass-produce 1.4nm chips in 2027. It is evident that the competition around advanced process nodes continues.

Rapidus has chosen a completely different technological path from traditional IDMs—its front-end process fully adopts single-wafer processing: each wafer is independently processed and measured at each key step, with intensive data feedback combined with AI to optimize process parameters, allowing for finer control over yield and defects. The trade-off is higher capital expenditure and lower capacity efficiency. In terms of business model, it will initially focus on attracting AI data center custom chip design companies (fabless), and then expand to edge device customers in sectors such as automotive and robotics.

For the Japanese government, the significance of Rapidus goes far beyond just one company; it represents a systematic project of "IBM technology transfer + EDA/IP ecosystem alliance + domestic complete machine factory order introduction": while Taiwan Semiconductor's Kumamoto factory focuses on 12–28nm automotive/image sensors, Rapidus takes on the symbolic task of "rebuilding 2nm-level manufacturing capability in Japan."

Why is everyone rushing to build 2nm factories?

Now let's analyze why the world is rushing to build 2nm factories. This can actually be viewed from three layers of logic:

First, in terms of technology and economic logic: 2nm is the "energy infrastructure" of the AI era. After the GB200 and Vera Rubin generations of AI accelerators, the next generation of training/inference chips is almost destined to be realized at nodes beyond 3nm (N2, N2P, 18A, SF2, etc.). 2nm can support further improvements in computing power per watt with higher transistor density and lower power consumption—in a context where power supply, heat dissipation, and data center load are constraints, leading by one node essentially means having an advantage in the unit CAPEX of AI infrastructure.

Furthermore, in terms of capital and industrial chain logic: Huge capital expenditures can only be bundled with "policy + leading customers." A 2nm factory typically costs around $8–10 billion, plus EUV machines, advanced packaging plants, and supporting water and electricity networks, making it difficult for a single enterprise to absorb it solely through free cash flow TSMC relies on Apple, NVIDIA, AMD, large-scale cloud vendors + subsidies from the US, Japan, and Germany; Intel has nearly 10% equity from the US government + a $5 billion strategic investment from NVIDIA; Samsung feeds its 2nm production line with "high-margin orders" through internal group orders + high ASP orders from Tesla and mining machine customers, using time to create space; Rapidus is entirely "policy-driven + ecological partners on stage," with IBM, Cadence, and others forming its technology and tool barriers. In this model, building a 2nm factory is no longer just a corporate decision but an execution tool for national industrial policy.

Finally, 2nm carries a strong geopolitical color. The US government directly invests in Intel, turning CHIPS subsidies into state-owned shares; Japan's Ministry of Economy, Trade and Industry views Rapidus as key to "next-generation information infrastructure"; Europe introduces TSMC, Intel, and GlobalFoundries through the European Chips Act; Taiwan, China, further elevates its "irreplaceability" in the global supply chain through intensive local factory plans. In other words, whoever masters 2nm capacity will hold the discourse power in the AI computing game over the next decade.

Of course, all these seemingly bustling factory plans are not without concerns:

First, will demand always be this "outrageous"? Wei Zhejia has publicly acknowledged that the demand for advanced processes driven by AI "far exceeds expectations," roughly three times the current capacity—yet whether this over-expectation is a short-term supply gap or a sustainable structural trend over 5–10 years remains uncertain.

Once the world simultaneously invests in dozens of factories for 2nm/1.4nm, the price wars, impairments, and financial pressures during the next economic downturn could be more severe than the memory price collapse in 2018.

Second, the issue of geopolitical concentration. If TSMC indeed builds 10 2nm factories in Baoshan, Nanzi, and Nanke, along with the A14 factory in Taichung, three factories in Arizona, Japan's Kumamoto, and the planned factory in Germany, global 2nm+ capacity will still be highly concentrated in Taiwan, China, and a few allied countries—this poses a double-edged sword for supply chain resilience and geopolitical risk management.

Third, whether talent and supply chains "can keep up with the factories." 2nm is not just about building a factory; it also requires quickly assembling thousands of engineers and technicians with EUV experience, along with sufficient chemical supplies, photomasks, and maintenance teams for front-end/back-end equipment. This is why TSMC is simultaneously hiring aggressively in Taiwan while being recognized as a "key force driving the growth of the Taiwanese workforce in the US."

Who will benefit?

The construction of 2nm factories directly benefits semiconductor equipment manufacturers (one of the biggest winners). A 2nm factory is essentially "built from equipment." EUV/immersion lithography, deposition, etching, cleaning, measurement/testing, CMP, and packaging-related equipment will receive orders and delivery schedules during the construction period. Additionally, upstream materials and consumables supply chains, such as wafers, photoresists/chemicals, specialty gases, target materials, polishing materials, photomasks/masks, and filtration systems, will continue to expand alongside production line growth Once the yield rate reaches mass production, the advanced packaging and testing chain will be beneficial in the long term. AI chips no longer rely solely on more advanced processes to win, but rather on a system combination of process + CoWoS/2.5D/Chiplet + HBM. The expansion of 2nm capacity often synchronously drives demand for advanced packaging, testing, substrates, and more.

For major clients such as NVIDIA, Apple, AMD, Qualcomm, Tesla, and cloud vendors, the entry of multiple 2nm players provides these large clients with more bargaining power and second source options.

However, the biggest uncertainty lies in the fact that if AI demand declines or yield rates do not meet standards, the party that "invested heavily in capex but cannot fill the production line" will be the most pressured.

Summary

If we take a longer view, the 2nm process and the wafer fabs built around it resemble the "final exam" of Moore's Law over the past 40 years: for TSMC, this is an asset reallocation concerning "domestic vs overseas, moat vs risk mitigation"; for Intel, this is a do-or-die battle to secure "the lowest fidelity in advanced processes" using 18A and national team capital; for Samsung, this is a gamble to forcibly elevate itself from "capable challenger" to "an unignorable second choice" through 2nm; for Rapidus and Japan, this is a window to reclaim a voice in advanced manufacturing amid the reshaping of global division of labor.

The physical significance of the 2nm process itself may become increasingly blurred, but the technological, capital, and political struggles surrounding "building 2nm fabs" clearly delineate a new industrial watershed: who can endure this high-capital, high-risk juncture, and more importantly, who can still stand firm after AI computing power demand returns to normal, will determine who is qualified to continue writing the rules in the next 1.x nm cycle.

Risk Warning and Disclaimer

The market has risks, and investment requires caution. This article does not constitute personal investment advice and does not take into account the specific investment goals, financial conditions, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article align with their specific circumstances. Investment based on this is at one's own risk