Old panel factories don’t have to worry about selling, why are chip giants lining up to take over?
Semiconductor industry vertical and horizontal
2026.11.21 11:44
· From Beijing
Full text 4584 words
Text | Semiconductor industry vertical and horizontal
In the past two years, the busiest thing in the display industry has not been to build factories, but to sell factories, and the faces that appear most frequently on the buyer list are wafer foundries and storage plants. Those 5th generation lines, 5.5th generation lines, and 7.5th generation lines that were labeled as "backward production capacity" suddenly became the shortest path in the semiconductor production expansion chain, and were highly favored and even robbed.
01 The buyer list is getting longer and longer
As early as August 15, 2024, TSMC purchased Innolux Nanke's fourth factory, a discontinued 5.5-generation LCD factory, for approximately 3.47 billion yuan. Innolux recognized a disposal benefit of approximately 2.98 billion yuan. Less than a year later, this factory became TSMC's AP8 advanced packaging base. It is understood that the machine was put into operation on April 2, 2025. The clean room area is close to 100,000 square meters, which is about four times that of the previous AP5 factory. The first phase of renovation only took 8 months. It is worth mentioning that the original buyer of this plant was Micron, and TSMC grabbed the deal.
More than two years later, the price has more than doubled. On October 29, 2026, AUO announced that it would sell the entire Taichung Phase II plant to TSMC. The target includes the L5C 5th generation line, L6A 6th generation line and L7A The Gen 7.5 line has a total of nine buildings, clean rooms and factory facilities, with a licensed area of approximately 429,000 square meters and a usable area of approximately 166,000 square meters. The transaction amount is approximately 7.85 billion yuan. The expected disposal proceeds are approximately 3.74 billion yuan, which must be recognized after the delivery conditions are met. The valuations given by the two valuation agencies were around 7.8 billion yuan, and the transaction price was slightly higher than the valuation. Including the previous Innolux Nanke Factory, TSMC paid a total of approximately 11.32 billion yuan for these two second-hand factories.
Of course, TSMC is not the only buyer. Storage plant Micron will take over the C4A, C5D, and C6C color filter factories in AUO's Tainan Science and Technology Industrial Park in August 2024, with a cost of about 1.5 billion yuan. Including AUO Crystal Materials' west side facilities in Houli, about 140 million yuan. yuan, totaling about 1.64 billion yuan, for front-end wafer testing and supporting Taichung's DRAM production capacity nearby; in February 2025, the Houli east side plant was sold for about 620 million yuan, and the east and west sides of Houli totaled about 760 million yuan. Another wafer foundry company, UMC, was reported in June 2025 to be negotiating with Caijing to purchase the factory building opposite its Nanke Fab12A for the development of advanced packaging. UMC did not comment on the rumors and only stated that it would not rule out expanding the factory in Taiwan. This negotiation went nowhere. In April 2026, this factory was purchased by ASE Silicon Products for about 870 million yuan.
Another circle outside is an extension of the same set of logic. In April 2026, Innolux sold Nanke Fab5 to ASE Semiconductor for approximately 3.01 billion yuan, with a building area of approximately 184,000 square meters, and a disposal profit of approximately 2.69 billion yuan; on March 10, the board of directors approved the transfer of Nanke Module Factory to ASE Semiconductor for approximately 3.01 billion yuan. About 180 million yuan was sold to Nanmao, with a building of 34,300 square meters, and a disposal profit of about 130 million yuan; Jingjin sold the Nanke factory to Silicon Products for about 1.31 billion yuan. When combined with the Caijing deal, Silicon Products spent a total of about 2.19 billion yuan. On AUO's side, Kaohsiung Luzhu's fifth-generation color filter plant was sold to ASE for about 1.28 billion yuan, and the Singapore AFPD plant was sold to Western Digital for about 1.86 billion yuan. Strictly speaking, ASE, Silicon Products, and Nanmao are packaging and testing plants, Micron, and Western Digital are storage manufacturing plants. The only real wafer foundries are TSMC and UMC. But this just shows that the competition is not just for OEM production capacity, but the entire chip manufacturing chain is competing for high-standard clean space.
02 What you buy is a factory, and what you save is time.
The underlying background supporting this round of factory building rush is the shift in the focus of the bottleneck in the AI computing power competition. In the past three years, the constraints of AI chips have been constantly switching. Advanced processes and HBM have become industry stuck points. By 2026, the contradiction between advanced packaging production capacity and supply will come to the forefront.
According to agency estimates, TSMC's CoWoS monthly production capacity will increase from approximately 12,000 to 15,000 pieces in 2022 to nearly 70,000 pieces in 2025, and is expected to reach 120,000 to 140,000 pieces by the end of 2026. Adding in the 50,000 to 60,000 pieces outsourced by packaging and testing plants such as ASE and Amkor, the industry-wide monthly production capacity is close to 200,000 pieces. Even so, the gap between supply and demand in 2026 will still be around 20%.
The supply chain estimates that Nvidia has locked in TSMC's CoWoS production capacity of about 800,000 to 850,000 pieces in 2026, accounting for more than half of TSMC's annual CoWoS production capacity that year. Customers such as Broadcom, AMD, Google TPU, Amazon Trainium, etc. need to queue up in the remaining capacity. Supply chain sources say that the order delivery cycle can reach 52 to 78 weeks, and it takes about a year and a half to be delivered after placing the order.
The scarcity of CoWoS is directly reflected in the price. According to calculations by Northeast Securities, the ASP of a single CoWoS wafer is about US$10,000 to US$10,500, and the market value of unit production capacity is about US$12.8 billion per 10,000 pieces per month, which is less than 10% different from US$14 billion for the 7nm advanced process. The gross profit margin of advanced packaging is 48% to 55%, which is not lower than that of advanced processes. In October 2026, it was reported in the supply chain that TSMC was considering raising its CoWoS quotation by 10% to 20%, leaving customers with little room for negotiation.
The evolution of technical routes has further amplified the demand for ultra-large clean spaces. TSMC's CoWoS with a 5x mask size has already been mass-produced in 2026, and the 14x mask size is planned for mass production in 2028. CoPoS panel-level packaging is expected to be trial-produced in small batches in 2027 and mass-produced from the end of 2028 to 2029. The packaging size continues to expand, and the requirements for the clear height of the factory building, floor load-bearing, and large-size substrate transportation channels are also rising. The traditional packaging factory adapted to 12-inch wafers is difficult to support the next generation of large-size packaging solutions. This is the core logic why chip manufacturers are willing to spend heavily to acquire old panel factories. The large clean space brought by the factories just matches the hardware conditions of new architectures such as CoPoS.
Therefore, the extreme market demand requires TSMC to accelerate the pace of production expansion. However, according to TSMC's own construction pace, nine factories will be built or renovated every year from 2025 to 2026. Previously, there were only four factories per year on average. Advanced packaging and advanced processes are competing for construction time at the same time. To build a new semiconductor factory, the process from site selection, environmental assessment, civil engineering, mechanical and electrical installation to installation verification usually takes years; while buying a ready-made factory only requires in-plant modification, equipment installation and trial production. When Innolux Nanke's fourth factory was transformed into AP8, the supply chain's statement was very straightforward. The deal eliminated the annual environmental impact assessment phase. As long as the factory was modified, it could be put into production in less than a year after the machines were installed. Therefore, if you only look at the land and construction, it is not cheap to buy an old factory building for about 7.85 billion yuan, which is even higher than the 6.07 billion yuan that the market had earlier rumored. TSMC is willing to pay a premium because it is not buying assets, but time.
The transaction of AUO's Taichung plant further illustrates the value of geographical location. AUO's Zhongke plant is adjacent to TSMC's Zhongke Factory 15. TSMC is planning four A14 or 1.4 nanometer wafer fabs in Zhongke Fab25. The first one will be completed in April 2027 and will be trial-produced in the third quarter of the same year. According to official statements, mass production will have to wait until 2028. All four fabs are expected to be completed in the second quarter of 2028. Plants, land, power supply, pure water, wastewater permits, talent and supply chains have all been prepared in a science park that has been operating for two decades. Buying an old factory is equivalent to buying out the most difficult prerequisites at once.
03 Why is the panel workshop born like a wafer factory?
The array segment of TFT-LCD is originally a series of semiconductor processes: PVD and CVD are used for film formation, exposure and development are used for patterning, and there are many etching, stripping, cleaning, and detection. However, the substrate is glass instead of silicon wafer, and the line width is at the micron level instead of the nanometer level. In order to maintain the yield rate, this production line is also equipped with clean rooms from Class 100 to Class 1,000, vertical laminar flow full of FFU, raised floors, constant temperature and humidity, ultrapure water, special gases, chemical supply and waste liquid and waste gas treatment systems. It also needs to prevent micro-vibration, anti-static and particle pollution. The clean room area of a 5.5-generation factory can easily be tens of thousands of square meters. The span, floor height, and logistics channels of a 7.5-generation factory are designed for large-size glass handling.
These conditions are just the ticket for advanced packaging and panel-level packaging. What the packaging line needs is not nano-level cleanliness, but a large-area, high-grade, flexibly partitionable clean space, as well as overhead cranes and automated logistics that can handle large-size substrates. The same goes for the storage test line. Micron converted the AUO factory into a front-end wafer test base because of its large space and ready-made factory systems.
However, in the wafer front end, especially in advanced logic, the clean room must be ISO 1 to 3. The control of molecular-level pollutants, micro-vibration, and temperature and humidity fluctuations is much stricter than the conventional ISO 5 to 6 in panel factories. The load-bearing, floor height, and chemical compatibility of the factory building may not be sufficient. Therefore, the best places to transform old panel factories are advanced packaging, panel-level packaging, wafer testing, mature processes and power devices, rather than advanced logic. While TSMC is buying old factories for packaging, it is still building a new A14 factory in Zhongke, walking on two legs.
04 The seller is calculating another account
Panel manufacturers are willing to sell because the debt left in their hands can no longer be settled.
The small generation line has long been uncompetitive in terms of cost. For the same glass substrate, the cutting efficiency, unit depreciation and material utilization of the large-generation lines are completely superior, while the 5th-generation and 5.5th-generation lines can only retreat to niche markets such as industrial control, automotive, and medical. According to Luotu Technology's calculations, mainland Chinese manufacturers' combined share of global LCD TV panel shipments will reach 72.1% in 2025, an increase of 5.6 percentage points from 2024. TrendForce predicts that mainland manufacturers' LCD production capacity share will increase from 68% in 2024 to 71% in 2025. During the same period, the global LCD production capacity shares of Innolux and AUO will drop to 9.1% and 9.2% respectively. Staying where you are and fighting for price will only continue to consume cash on depreciation.
Selling the factory turns the sunk assets into cash at once. Innolux's net profit after tax in the third quarter of 2026 was approximately 1.56 billion yuan, an annual increase of more than 114 times. The cumulative profit in the first three quarters was approximately 2.80 billion yuan, an annual increase of approximately 50 times. However, the operating profit during the same period was only approximately 200 million yuan, with an operating profit rate of 1.70%. The industry's contribution to net profit was approximately 13%, and most of the rest came from asset disposals. This structure is very straightforward: selling factories has supported the financial statements, and the industry is still floating within a narrow range near the profit and loss line.
AUO Chairman Peng Shuanglang said after the announcement that since the beginning of this year, the company has successively promoted the disposal of factory buildings, and the remaining factory space has been reasonably allocated. The asset activation has come to an end, and the recovered funds will be invested in strategic and future projects. Innolux invested money in panel-level fan-out packaging, X-ray sensor integration with automotive systems, and retained the glass substrate trial production line; Caijing set its goals in automotive, industrial control, wearables and drones, with automotive revenue accounting for more than 30% in 2027. These directions have one thing in common: they all require cash and do not require so many old production lines.
There are costs as well. The disposal of profits is a one-time thing. After the sale, the production capacity will be hollowed out, the revenue scale will go down, and employee placement is also a real problem. According to Taiwanese media reports, Innolux Nanke’s fifth factory was given an excellent separation plan when it closed down. Employees with 20 years of experience can receive at least 9 months’ salary, which will be divided into three stages starting in June and completed in August. Caijing's gross profit margin in the first half of 2026 is still negative 3.24%, and its operating loss is about 190 million yuan. It has been losing money for four consecutive years. For it, selling the factory is more like stopping the bleeding, and it is not yet a calm transformation.
05 Not every old factory can be transformed
There are also failure cases on this track, which just shows where the threshold is.
AUO originally sold the Huaya plant to Quanta for about 3.99 billion yuan in July 2026 for server production capacity. However, it was terminated three months later because the two parties could not agree on the factory usage specifications and delivery schedule, and the estimated disposal benefits of about 2.71 billion yuan were lost. Factory buildings are not standard parts. If the buyer wants different floor height, load-bearing capacity, cleanliness level, power redundancy, gas pipeline direction, and equipment entrance size, the account will be recalculated.
There are four real thresholds for transformation. The first is structure and anti-microvibration. The panel factory is designed based on large glass handling, which does not mean designing based on the vibration spectrum of the lithography machine. The second is cleanliness and molecular pollution control. Advanced packaging has much stricter requirements for AMC than panels. The third is power and factory redundancy. Dual power supplies, UPS, ultrapure water and special gas systems must be expanded. The fourth is environmental protection permits and emission indicators. What can be changed at the factory site is subject to the list of local authorities. This also explains why the most sought-after ones are large factories older than 5 generations, rather than older, smaller and cheaper lines.
In contrast, there is another path. In March 2026, Micron completed the acquisition of the Power Semiconductor Tongluo P5 factory and directly obtained a 300mm clean room of about 300,000 square feet. After the transformation, it plans to start shipping in fiscal year 2028, and has planned to build another clean room of about 270,000 square feet at the same factory site. The price of buying a functioning wafer factory is obviously more expensive than buying a suspended panel factory, but it illustrates the same thing: in the competition for AI production expansion, ready-made clean space itself has become a strategic asset.
Back to the original question. Why are there so many cases of panel manufacturers selling factories and chip factories buying factories recently?
Because three forces met at the same window. AI has turned packaging and storage into bottlenecks in computing infrastructure. Semiconductor factories need to get clean rooms as quickly as possible; LCD has entered a mature stage, and old production lines in Taiwan, Japan and South Korea have lost their cost competitiveness, and panel factories need cash to complete the transformation; constraints on land, electricity, and environmental impact assessment are getting harder and harder, and the time cost of building a new factory is unacceptably high. The three forces are superimposed, and the old panel factory has changed from "burden" to "hard currency".
From 2026 to 2028, it is the critical three years for CoWoS to expand production, HBM to expand production, glass substrates to move from verification to small batches, and CoPoS to move from trial production to mass production. In these three years, old factories that can be renovated will be gradually digested. After 2028, the new production capacity of semiconductor factories will be gradually released, the scarcity of old factories will decrease, and prices will return to rationality.
Respect for the original work, if you are qualified, you should be appreciated.
According to the real-name requirement of the Cybersecurity Law, please bind your mobile phone number before posting a comment.
Express
11:44
Bitcoin falls below $57,000, hitting lowest level since September 2024
11:36
The Mongolian section of Hanoi has entered the cold flood season from 2026 to 2027
11:11
SoftBank’s Masayoshi Son discussed acquiring SK Hynix for US$1 trillion
11:05
U.S. cargo Dragon spacecraft breaks away from the space station and sets off for return to Earth
10:31
U.S. presidential envoy discusses U.S.-Iran mediation work with senior Qatari officials
10:30
Quantum computer "high-speed memory" is realized on a real machine for the first time in the world
10:27
Agency: Demand has cooled, and prices for TVs, monitors and laptop panels may remain flat across the board in November.
10:25
Ming-Chi Kuo: Google is expected to adjust the shipment proportion of TPU 8t and 8i series in 2027
10:22
The phased adjustment has not changed the background, and many institutions are optimistic about the valuation and resilience of A-shares.
10:21
It is reported on the Internet that the trunk of Leipao A10 was crushed by several boxes of beer. Leapao responded.
10:19
International gold prices continue to decline
10:16
Federal Reserve official Paulson: Supports prudent interest rate hikes and can continue to raise interest rates if necessary
10:15
If the sale to SoftBank cannot be completed, ABB is reported to be considering spinning off its robotics business.
10:12
Anthropic IPO received approximately US$115 billion to US$120 billion in subscription demand from retail investors
10:10
The price of gold jewelry continues to fall, with Lao Fengxiang falling by 34 yuan in a single day
Food delivery company Wonder is considering selling Grubhub
10:09
Databricks is reported to be negotiating a new round of financing, with a valuation expected to reach US$225 billion
10:08
CrowdStrike reportedly in talks to acquire cybersecurity company Zscaler
10:06
Apollo and Goldman Sachs complete pricing of $44 billion SpaceX GPU financing
‘Big Short’ investor Michael Burry reduces holdings of Nvidia put options, locking in some gains
Scan to download App

