In 2025, why did the 2nm chips all miss their release schedules?
In 2025, no flagship mobile phones are equipped with 2nm chips.
The A19 and A19 Pro series chips of the iPhone 17 adopt TSMC's N3P process. The upcoming MediaTek Dimensity 9500 and the fifth - generation Qualcomm Snapdragon 8 Extreme Edition will also use this process.
When there was no unexpected situation, MediaTek created an unexpected event.
Recently, MediaTek officially announced that the design and tape - out of its 2nm chip (Dimensity 9600) were completed, making it one of the first companies to adopt this technology. Mass production is expected by the end of next year. It's worth noting that they announced this progress a full year before mass production, which is truly extraordinary.
Based on the new product release schedules of major mobile phone manufacturers, by the end of 2026, in addition to the 2nm Dimensity 9600, Apple's A20 series, the sixth - generation Qualcomm Snapdragon 8 Extreme Edition, and Samsung's Exynos 2600 will all adopt the 2nm process.
It's certain that the "fierce battle" for 2nm chips will intensify in 2026. However, TSMC and Samsung have been hyping up 2nm for years. Why wasn't the A19 chip of the Apple iPhone 17 equipped with 2nm technology? And why didn't the "2nm war" break out this year?
01 "Never dreamed that the demand for 2nm would be higher than that for 3nm"
At the earnings conference on October 17, 2024, TSMC President Wei Zhejia described the demand for 2nm chips in two sentences: "Extremely high" and "Never dreamed that the demand would be higher than that for 3nm."
Here's a question: TSMC only started accepting 2nm orders on April 1 this year and began mass production in the second half of the year. Why could Wei Zhejia predict the demand for 2nm chips in October 2024?
"TSMC has a very top - notch market research team that can integrate the demands from various industries worldwide, including those from NVIDIA, Tesla, AMD, etc.," Wu Zihao, a former TSMC factory - building engineer, said. "Building a foundry takes about four years, which involves capacity construction planning. Fabless (chip design companies without their own foundries) like Apple and NVIDIA must submit order forecasts."
Wu Zihao also revealed that from the perspective of fabless companies, in addition to tape - out in advance, R & D also needs to be aligned with the platforms and technologies provided by wafer fabs, which are also sources of demand information.
In addition, the capacity forecast clause in the wafer foundry agreement requires customers to provide reasonable order forecasts to the wafer fab to facilitate the foundry's reasonable capacity scheduling, which can also partially reflect the demand status of fabless companies.
According to TrendForce data, Apple, AMD, NVIDIA, MediaTek, etc. have all reserved TSMC's 2nm production capacity. A significant portion of them are among TSMC's top ten customers. In 2024, Apple contributed 25.18% of TSMC's revenue, becoming TSMC's largest customer.
Among the above customers, MediaTek has announced its mass - production time. According to the press conference schedules of mobile phone manufacturers, it's basically confirmed that Apple will be the first to obtain TSMC's 2nm production capacity. AMD announced in April when TSMC just released its capacity that it would adopt the 2nm process in its next - generation EPYC data - center processor codenamed "Venice."
For NVIDIA, Rubin already uses the 3nm process, and Rubin Ultra combines four GPU Dies (four dies integrated into the same package). Since the package size cannot be increased further, it will also adopt the 2nm process.
An industry insider revealed that Bitmain is also a customer of TSMC's 2nm process. It may even be the first fabless company globally to use TSMC's 2nm technology. "ASICs for mining machines are relatively easy to manufacture. Adopting the new node first can also be a practice run. Bitmain may be able to start shipping in the second half of the year."
Compared with TSMC, there is less information about Samsung's 2nm customers. In addition to its own Exynos 2600, which will claim the label of "the world's first 2nm chip," industry rumors suggest that Qualcomm may return to Samsung's embrace at the 2nm node.
The unprecedented high demand from customers is essentially due to the performance improvement brought about by the transition from 3nm to 2nm.
TSMC previously disclosed the parameters of the N2 node - compared with the first - generation N3E, the transistor density increases by 15%, the performance improves by 10% - 15% under the same power consumption, and the power consumption decreases by 25 - 30% under the same performance.
MediaTek's press release about "taking the lead" in the 2nm field also basically confirms the rationality of the above data. MediaTek said that compared with the existing N3E process, TSMC's enhanced 2nm process technology increases the logic density by 1.2 times, improves the performance by up to 18% under the same power consumption, and reduces the power consumption by about 36% at the same speed.
In short, the performance improvement brought about by 2nm has made major fabless companies eager to adopt this technology. However, the mass - production times of major manufacturers are mostly in 2026.
02 TSMC's "hiccup"
The reason why flagship mobile phone chips in 2025 can't use 2nm technology is still due to TSMC's "hiccup."
According to TSMC's plan, the 2nm production capacity was originally scheduled to start in mid - 2025, and the current progress is in line with expectations. However, if mobile phone customers want to mass - produce 2nm chips in 2025, the available time window is too short.
"It takes several months from tape - out to getting the chips back. After getting the chips back, several more months are usually needed for function and performance debugging," said a chip design professional.
That is to say, even if a major customer like Apple completed the tape - out and testing of the A20 chip at the end of 2024, it would have to wait until June this year for mass production, missing the inventory - building schedule for the iPhone 17. After all, Foxconn's assembly lines also need time to get up and running.
Yield rate is another factor why mobile phone manufacturers aren't chasing 2nm technology this year, but its impact is less significant than the mass - production schedule. Different fabless companies also have different sensitivities to it.
At the 3nm node, the early yield rate was only about 60%. Later, the yield rates of N3E and N3P gradually climbed above 80%. The same process will occur at the 2nm node.
"The yield rate for product introduction (of 2nm) may have exceeded 70% and will gradually climb to 80% next year," estimated the aforementioned industry insider.
The early low yield rate also means relatively high prices. Customers sensitive to price will plan their mass production after the yield rate climbs and adopt the "wafer buy" model. Otherwise, the more they produce, the more they will lose. However, price is not an absolute obstacle.
Take Apple as an example. It has a "finished goods buy" agreement with TSMC, paying only for good - quality chips. As long as the yield rate is not extremely low, price won't be a decisive factor. However, TF Securities analyst Guo Mingji has a different view. He believes that although Apple purchases finished chips, the purchase cost actually includes the cost of defective chips.
"The best evidence is that the cost of new processors used in new iPhones has increased significantly every year, and this year's A17 is no exception," said Guo Mingji.
03 The wafer foundry war
Overseas wafer fabs are currently working on the mass production of 2nm chips. There are slight differences in node naming, including N2, 20A, SF2, 2nm, etc. However, they've "tacitly" adopted the new GAA transistor architecture and have all planned to use back - side power supply technology in subsequent iterations.
In addition, there's an important point: Back - side power supply can separate the power and signal connections and move them to the back of the integrated circuit, reducing resistance, increasing transistor density, and improving performance.
In terms of mass - production schedules, Samsung, TSMC, etc. are generally progressing according to the timelines. On the contrary, Intel, which was originally the most aggressive, planned to open its 2nm production capacity at the end of 2024. However, due to multiple factors such as technical challenges and management changes, it ultimately cancelled the 2nm (20A) process and suspended the development of 18A (1.8nm), focusing all its efforts on the 14A (1.4nm) process.
Regarding specific production capacity, according to TrendForce, next year, TSMC is expected to have four 2nm wafer fabs operating at full capacity, with a total monthly production capacity of 60,000 wafers.
The aforementioned industry insider said, "The monthly production capacity of Fab 20 in the Hsinchu Science Park is at least 60,000 wafers, and the monthly production capacity of Fab 22 in Kaohsiung is expected to be 30,000 wafers. Next year, the monthly production capacity of 2nm will be at least 90,000 - 120,000 wafers."
Regarding Samsung's production capacity, in April, TrendForce cited data from the Seoul Economic Daily (SEDaily), stating that its monthly production capacity for 2nm was 7,000 wafers.
2025 is a crucial point for the launch of 2nm production capacity, but this wafer foundry war can be traced back several years.
In October 2021, Samsung announced the start of 2nm R & D at its annual foundry conference and released the relevant timeline and technical roadmap. TSMC was even earlier, officially announcing its entry into the R & D stage in June 2019. It revealed at the Global Technology Forum that it planned to establish a new 2nm R & D line and invest more than 8,000 engineers.
Overall, the R & D time for major wafer fabs at the 2nm node is between 4 - 6 years. During this period, wafer fabs' annual capital expenditure on R & D generally exceeds $1 billion, and TSMC even reached $3.6 billion in 2022.
The huge R & D investment is not only reflected in technical solutions but also in the competition for R & D equipment. The most typical example is the competition for ASML's High NA EUV lithography machines.
In 2022, Samsung's Lee Jae - yong visited ASML in an attempt to secure advanced lithography equipment. However, Intel won the world's first High NA EUV lithography machine, which costs nearly $400 million, at the end of 2023, and received another one of the same model in 2024.
Compared with Intel and Samsung, TSMC has been relatively conservative in the competition for the most advanced equipment. It was once unmoved when ASML executives visited TSMC to discuss advanced lithography machines. However, against the backdrop of its competitors' aggressive purchases, Wei Zhejia also visited ASML in 2024. Rumors suggest that TSMC received an "bundled discount package" from ASML - a price discount for the High NA EUV lithography machine but with the requirement to purchase some older models.
Zhang Xiaoqiang, TSMC's deputy head of R & D, has repeatedly stated that the High NA EUV lithography machine is too expensive. "As long as we continue to find alternative solutions, there's no need to use this expensive equipment."
04 Is Moore's Law dead?
In the past few years, the industry has been discussing whether Moore's Law is dead.
Moore's Law originated from Gordon Moore, the late founder of Intel. In 1965, Gordon Moore published an article in the magazine "Electronics," predicting that the number of transistors and resistors integrated on semiconductor chips would double every year. Ten years later, in 1975, Moore revised his view, changing "doubling every year" to "doubling every two years."
For the next half - century, the number of transistors followed this law until the 7nm node, when the time interval extended from 24 months to 30 months.
The mass - production times of 7nm, 5nm, 3nm, and 2nm were 2018, 2020, 2023, and 2025 respectively, with an average interval of 30 - 36 months. The industry generally believes that for nodes beyond 1nm in the future, the interval will likely extend to over 40 months.
The extended rhythm will directly cause multiple generations of products from fabless companies to remain at one major node.
Take Apple's A - series chips as an example. In the past, there was usually a node change every two years. For example, both the A14B and A15B were based on the 5nm process, but the A17 Pro, A18, and A19 have remained at the 3nm node for three years.
So, how long will the chips stay at the 2nm process in the future?
According to TSMC's plan, the 2nm node has four iterations: N2, N2P, N2X, and the upgraded A16 (1.6nm). Including the A14 (1.4nm) process with the second - generation GAA architecture, they correspond to Apple's A20, A21, A22, and A23 chips. Then, in 2030, the 1nm process will be introduced, and the A24 series chips will enter mass production.
This means that from now on, it will take at least another five years to transition from the 2nm era to the 1nm era.
Of course, node upgrades mainly refer to the upgrade of names and line widths, which doesn't mean that the number of transistors won't increase. For example, there have been multiple iterations of the 3nm process, from N3, N3E, N3P, etc., and the number of transistors has still increased significantly in each generation.
TechNews previously conducted statistics. Compared with N3, N3E reduces power consumption by 32% under the same performance and improves performance by 15% under the same power consumption. Compared with N3E, N3P reduces power consumption by 5% - 10% under the same performance and improves performance by 5% under the same power consumption. Roughly estimated, compared with N3, N3P reduces power consumption by about 20% - 27% under the same performance and improves performance by