There are too many G-spots for chrysanthemums; several of my articles have been deleted by the system due to mentioning related situations. I do not keep drafts, so teachers, please do not ask what was written; if it’s gone, it’s gone. If you find the article useful, I suggest taking screenshots after reading.
Due to the ongoing positive adjustments in overseas computing power, rumors emerged yesterday that the orthogonal backplane would be canceled, continuing the use of copper cables. At that time, Wol was also pulling a board. Regardless of whether the cancellation occurs, even without the orthogonal backplane, there is already a gap of about 12 billion in AIPCB. Once the orthogonal backplane is used, combined with the midplane in NV’s midplane and the two new PCB boards from CPX, the PCB market will add over 20 billion. Even with expansion, according to Xingzheng, the industry still has a gap of about 20 billion. Although Shenzhen has been hit by a typhoon these days, I went outside to listen to expert meetings because my home is too high up. After several sessions, there was no mention of canceling the orthogonal backplane; instead, the industry is basically certain that the second round of test results will be released in October. Below are updates from the PCB expert meetings I attended these days, for teachers’ reference.
The global demand for AIPCB is expected to reach 90 billion yuan by 2026, and by 2030, with the continuous penetration of AI large model training and inference, as well as autonomous driving scenarios, the AIPCB market scale is expected to exceed 300 billion, with a compound annual growth rate of over 30% over ten years. PCB is a capital-intensive industry, with a new capacity construction cycle of 12-18 months, and overseas production progress is lagging. The capacity gap is expected to reach over 20% in the coming years; the sector’s PE in 2026 is about 20 times. The current sector is at a turning point of valuation recovery brought by simultaneous volume and price increases, with both growth certainty and elasticity.
Let me show the specifications of AIPCB to the teachers. Without discussing NV and domestic products, North American ASIC (custom chips) AIPCB basically starts at 20 layers; once the volume increases, the gap mentioned above will be hard to fill.
Amazon’s AWS TR3: 26-layer through-hole for the substrate, 4-stage 22-layer for the computing board.
Google’s TPU V7: 16+2+16 structure
Microsoft: 16+34 structure, forming a 50-layer N+M structure through lamination.
Meta: 22+8 N+M structure, forming a 30-layer board after lamination.
AMD: 24+36 complex design, with 3-4 stages of HDI in both 24 and 36.
Of course, when talking about AI, one cannot overlook NVIDIA. NV is concerned about the uncertainty of tariffs and requires suppliers to transfer 50% of their capacity overseas. Last week, the NV team inspected six factories in Thailand: Pengding, Shenghong, TTM, Huidian, Dingying, and Fangzheng. NVIDIA has completed all testing work for the 7-stage test boards in Thailand and has also proposed a demand for 8-stage new technology to Shenghong.
The market is most concerned about the Rubin orthogonal backplane. Currently, there are four suppliers involved, and only Shenghong reached the semi-qualification standard in the first round of testing. NVIDIA hopes Pengding will continue to advance the PTFE (polytetrafluoroethylene) solution. PTFE differs from M9 material: M9 material uses Q cloth to optimize signal transmission performance. PTFE material has superior signal transmission characteristics, but it is soft and weak, making processing prone to deformation; currently, domestic manufacturers have basically given up (though Shengyi is still persisting).
The next generation of PCB technology is COWOP, which involves placing chips directly on the PCB. The main participants in NVIDIA’s COWOP-related prototyping include Pengding, Xinxing, Shenghong, and Jingwang, each with its strengths and weaknesses: Pengding and Xinxing have advantages in technology maturity and capacity, while Shenghong excels in HDI manufacturing and has partially introduced mSAP processes in its new factory. Jingwang is an alternative supplier.
In other information, one NVL72 requires 36 computing boards, with a single board value of 500 USD. The area of the Rubin mainboard has increased by nearly 29% compared to the GB300, with a value increase of at least 15%.
Rubin updates:
1. Using a midplane to replace copper cables, the midplane is a 44-layer high-multilayer board structure, similar to the orthogonal backplane. Shenghong is also working on this, but specific details are unclear.
2. The computing board may use seven-stage HDI, 7+12+7 structure, 26-layer board.
3. Capex: Four chips are installed on two sides, each on a PCB board.
Company updates:
Pengding: After NVIDIA’s factory inspection in Thailand, they will produce samples of the GB300 compute tray, a 5-stage 22-layer board (5+12+5), with the company’s goal to deliver samples by the end of the month.
The Kaohsiung factory is mainly used for the production of chip test boards (ATE) and can support the manufacturing of high-complexity products with 80 to over 100 layers. After expansion, both soft and hard boards will be supported for production.
The Huai’an RPCB factory mainly serves four core customers: AWS, NV, Microsoft, and Google. The Huai’an substrate factory is currently facing a peak order period due to Apple’s new product launch, making it difficult for existing capacity to support other project expansions.
The first version of the Pengding COWOP project samples has been delivered to Daye Light and Silicones for testing. Pengding used M8 to make large panels, with outer layer BGA and lines, and inner layer without wires. The flatness is good, but the overall yield is not satisfactory. In early September, the COWOP project saw improvements, following the normal HDI process instead of the mSAP process, using Doosan’s M8 material + HVLP4 copper foil.
Shenghong will exceed the revenue target set at the beginning of the year, mainly due to a significant increase in NVIDIA orders in the first half of the year. However, due to the poor connection between GB200 and 300 (new orders for 300 may be placed in October), Q3 performance may be lower than Q2. But there is no need to worry too much, as Shenghong holds an absolute leading position in the HDI field, and its market share is hard to shake. Currently, AIPCB is divided into two tracks: Fangzheng, Shennan, etc., mainly serve domestic customers like Huawei; while European and American customers choose Shenghong, Huidian, and Jingwang.
Shenghong’s factory in Thailand originally mainly produced through-hole boards, but HDI capacity is currently expanding. Future Rubin series will be prioritized for production in the Thailand factory.
It is confirmed that there will be sample orders for the orthogonal backplane, but the specific share cannot be predicted. It is likely to be a high-multilayer through-hole process, 72 layers + M9 material (though some experts say the orthogonal backplane is a 26*3=78 layer structure, costing 150,000 to 200,000 yuan).
The Rubin computing board and switching board are still in the prototyping stage, and specifications have not been fully determined. The computing board sent by Shenghong is a 24-layer 6-stage structure. Whether Rubin can maintain a 60%~70% market share depends on the progress of competitors. Jingwang’s HDI capacity and expansion speed are lower than Shenghong’s, so the short-term impact of Jingwang is limited.
NVIDIA’s PCB design trend is to develop higher-performance substrates (COWOP) to meet future more complex demands, which may involve mSAP processes. The mSAP process is a new process, and the company is currently installing equipment in the factory, but complete verification and certification will still take some time. Pengding and Xinxing have significant advantages in mSAP processes, with maturity expected around 2027. Currently, the Rubin architecture is still based on a 2+1 configuration, meaning two GPUs and one CPU are placed on a single PCB.
Cooperation with Google has only been a few months, with related samples delivered for testing in mid-August, and feedback expected in three months. Meta is still in the early communication stage, and has not yet received formal test board requirements.
Collaboration with Cisco only involves through-hole technology for server products; AWS currently has no AI collaboration with Open AI, but maintains a long-term partnership with AWS, and is currently conducting preliminary work on system-level and document certification.
Here, I would like to recommend the entire PCB chain again, besides PCB manufacturers, consumables such as drill bits, chemicals, materials, copper foil, and electronic cloth copper powder. For example, high-end HDI boards require drill bits that meet higher winding ratio requirements due to the increased number of layers, and multiple drilling is needed on both sides, significantly increasing the usage of drill bits. Currently, Ding Tai’s drill bits have good quality and are cooperating with the development of new products, while material aspects like HVLP copper foil benefit more than equipment aspects.
Finally, here’s a joke about Ding Tai’s drill bits:
Ding Tai High-Tech: Tenfold demand leads to price increases and profit growth📈
Logic 1️⃣: Tenfold volume increase❗ This year, only domestic PCB companies have expanded production by over 10 billion, stimulating upstream demand for CCL, equipment, drill bits, etc. Next year, CCL materials will accelerate from M7 to M9, and the lifespan of traditional drill bits will plummet from 3000 holes per bit to 500 holes per bit, while M9 # will be below 200 holes. With the increase in demand for drill bits next year due to Rubin and other ASICs, # demand for drill bits will increase tenfold➕, and currently, high-end drill bits are already in short supply.
Logic 2️⃣: Doubling price increase❗ Unlike traditional drill bits, AI PCB boards use drill bits with a diameter of less than 0.2mm, and electronic cloth, especially quartz cloth, is harder. To ensure smooth drilling, drill bits need to be coated (e.g., diamond). Fine diameter + coating + supply shortage have already led to two price increases this year, and a # 20% price increase is brewing this year.
Logic 3️⃣: Volume and price lead to profit growth❗ As a global leader in drill bits (30%), Ding Tai continues to expand production domestically and internationally to support Shenghong, Jingwang, and others in going overseas. To ensure supply, customers like Shenghong # sign supply agreements to lock in capacity in advance. The company’s capacity is expected to start at 120 million bits/month in October, and over 150 million next year. The average price of AI drill bits was 1.2 yuan/bit at the beginning of the year, and after the price increase, it is 1.4 (M9 drill bits 2 yuan+). The simultaneous increase in volume and price leads to a gross profit increase, with a gross profit margin of # 40% in Q2.
Logic 4️⃣: Screw rod & equipment options❗ Ding Tai’s drill bit equipment is self-developed, with a maximum length-to-diameter ratio of over 50, and a minimum diameter of 0.01mm. The grinding machine and technology used for screw rods belong to # dimensionality reduction strikes, and the grinding machine equipment has already shipped millions. In addition, there are also contributions from grinding materials and high-end film businesses.
🔥 Volume increase, price rise, and profit growth driven by robots and options are expected to keep the company’s performance in a high growth trend over the next three years, continuously exceeding market expectations. I recommend leaders to pay long-term attention to the strongest # consumables segment in this round of AIPCB wave!
At this moment, the wind is howling outside, Shenzhen has canceled five stops, and tomorrow will be another day for the working class! Good night~ Also, a side advertisement.
