Rapidus Weighs Delaying 2nm Mass Production to Fiscal 2028

Japan's Rapidus is weighing a delay of its 2-nanometer (nm, one-billionth of a meter) semiconductor mass production from the second half of fiscal 2027 to fiscal 2028, Japan's Jiji Press reported on the 11th. The potential slip threatens the commercial case for a venture that has already drawn JPY 2.354 trillion in cumulative Japanese government support. It also lands as TSMC, Samsung Electronics and Intel accelerate advanced-process output and customer acquisition, squeezing the window for a latecomer to enter the 2nm foundry market.

What has gone wrong at Rapidus?

Rapidus is a semiconductor company established in 2022 with investments from major Japanese companies including Toyota, Sony, NTT and SoftBank. It has pursued commercialization of a cutting-edge 2nm process through technical cooperation with IBM of the United States. In July last year, it announced it had successfully fabricated prototype transistors, the core devices of the 2nm process, at its plant in Chitose, Hokkaido.

The progress since then has faltered. Rapidus has encountered setbacks in developing technology to connect individual devices and integrate them with other circuit components, according to Jiji Press.

Bloomberg reported on the 10th that Rapidus had failed to achieve normal operation of static random-access memory (SRAM) cells, which temporarily store data inside a processor. Delays in developing key technologies originally due for completion by this summer have put the overall mass-production preparation schedule at least several months behind, Bloomberg said.

Yield remains unresolved. Rapidus is facing persistent difficulties in raising yields — the percentage of usable chips per wafer — and stabilizing production processes. The company has reportedly yet to resolve issues with dust and fine-particle control in its cleanrooms, the facilities where semiconductors are manufactured.

How far ahead is TSMC?

Taiwan's TSMC, widely regarded as the market leader, began mass production on its 2nm (N2) process in the fourth quarter of last year. The process employs nanosheet transistors with a next-generation gate-all-around (GAA) architecture. This year TSMC has ramped output rapidly, primarily at its Hsinchu and Kaohsiung facilities.

The numbers are concrete. In the third quarter of this year, the 2nm process accounted for 8% of TSMC's total wafer revenue, according to the company. To meet growing demand for smartphone, artificial intelligence (AI) and high-performance computing (HPC) chips in 2027, TSMC plans to more than double annual production capacity for its 2nm-family processes compared with this year.

Follow-on processes are already in motion:

• N2P, offering a performance improvement of approximately 5% over N2, and A16, which incorporates backside power delivery technology, are scheduled for mass production in the second half of this year.
• A16 is designed to deliver 8–10% higher computing speeds than N2P, or reduce power consumption by 15–20% at the same performance level.
• The 1.4nm-class A14 process is set for trial production in 2027 and full-scale mass production in 2028, with preparations to secure capacity underway.

Can Samsung close the gap?

Samsung Electronics is also accelerating mass production as yields improve. According to a September analysis by Taiwanese market research firm TrendForce, the yield on Samsung's 2nm GAA process, which stood below 50% at the beginning of this year, has recently risen to around 70%.

Building on that progress, Samsung plans mass production of mobile chips using its second-generation 2nm process (SF2P) in the second half of this year. During last month's third-quarter earnings announcement, the company offered an optimistic outlook: 2nm orders were increasing, particularly from AI and HPC customers including major hyperscalers (large cloud operators), and the number of projects secured for its 2nm process this year would more than double from last year.

Capacity expansion is following the orders. Samsung is preparing to bring its Taylor plant in Texas into operation, anticipating that mass production of major customers' new 2nm products will begin in earnest in the second half of next year. The Taylor plant has begun operating its cleanrooms and bringing in equipment, and wafer starts to establish production lines are underway. Samsung is also reportedly in discussions to extend its 2nm process, previously focused on mobile chips, into AI and HPC applications, while preparing to develop experimental 1.4nm chips with multiple customers.

Where does Intel fit in the 2nm race?

Intel views 2027–2028 as a critical turning point for its advanced foundry business. After beginning production on its 1.8nm-class (18A) process last year, Intel started trial production of its performance-enhanced 18A-P process in June this year. Its next-generation 1.4nm-class 14A process is scheduled for trial production in the second half of 2027, followed by full-scale mass production in 2028.

Through these efforts, Intel aims to bring its foundry business to break-even by the end of 2027. In August, Intel Chief Financial Officer David Zinsner cautioned that the break-even date could be pushed back to 2028 if securing customers creates additional investment requirements.

The competitive dynamics are already producing customer commitments further downstream. South Korean AI chip company DeepX completed the design of its next-generation AI chip, the DX-M2, on Samsung's second-generation 2nm process and completed tape-out (the final design handoff to manufacturing) early this month. The DX-M2 is designed to run generative AI in robots and industrial equipment using approximately 5 watts of power. DeepX plans to obtain first prototypes in March 2027, begin supplying evaluation samples in the second quarter of that year, start initial mass production in the fourth quarter, and move into full mass production in 2028.

Is Japan's semiconductor comeback blueprint fading?

The stakes for Japan are fiscal as much as industrial. In April, Japan's Ministry of Economy, Trade and Industry decided to provide Rapidus with an additional JPY 631.5 billion in research and development funding, bringing cumulative support to JPY 2.354 trillion. Last month, a proposal was put forward to invest an additional total of JPY 400 billion over fiscal 2027–2029 to strengthen Rapidus's capital base.

Returns on that money remain uncertain. The timetable for mass production is in doubt, and efforts to secure customers remain essentially at an early stage. On the 7th of last month, Reuters reported that Rapidus was pursuing cooperation with 17 companies, including Synopsys and Infosys, to support semiconductor design, but that it remained uncertain whether this would lead to actual large-scale contract manufacturing agreements.

Some observers warn the outcome could echo Elpida Memory. Elpida was established in 1999 through the merger of NEC's and Hitachi's DRAM (memory chip) businesses and grew into a major specialist DRAM manufacturer after acquiring Mitsubishi Electric's DRAM business in 2003. After the global financial crisis in the late 2000s, its financial difficulties deepened amid a sharp decline in semiconductor prices, intensifying competition with South Korean companies such as Samsung Electronics, and weaker export competitiveness due to a stronger yen. In 2009, the Japanese government invested JPY 30 billion in Elpida through the Development Bank of Japan and supported bank lending. Elpida nonetheless filed for court protection in February 2012 with approximately JPY 448 billion in debt, and was acquired by U.S.-based Micron in 2013.

Market observers say the barriers to entry for latecomers will inevitably grow over time as the incumbents lock in customers and capacity. As one assessment cited in the report put it, "Competition is already fierce." If the delay materializes, Rapidus's test will be whether roughly JPY 2.7 trillion in pledged public money — and counting — can buy back a seat at the 2nm table before the market consolidates around TSMC, Samsung and Intel in 2027 and 2028.