Sinolink Securities released a research report stating that the PCB industry is policy, capital, and technology-intensive, with extremely high entry barriers. The Ministry of Industry and Information Technology has set high investment and production capacity standards, requiring substantial equipment and R&D investment in the PCB sector. Manufacturing processes are complex, demanding interdisciplinary expertise; environmental standards are stringent, leading to high export compliance costs; and certification cycles for key clients are long with strong customer stickiness. Multiple barriers are driving industry consolidation towards leading players, continuously optimizing the competitive landscape. The main viewpoints of Sinolink Securities are as follows:

AI inference bottleneck iteration and architectural evolution are driving a leap in PCB value positioning. Under the Transformer architecture, large model inference is divided into two stages: Prefill and Decode. The former is compute-intensive, while the latter is memory bandwidth-intensive, resulting in an extreme mismatch between compute utilization and bandwidth occupancy. NVIDIA has introduced a "decoupled inference" architecture, splitting Prefill and Decode across different hardware, imposing higher demands on PCBs for denser HBM packaging substrates, higher-speed inter-chip interconnects, and higher power density for power supply and thermal management. Simultaneously, in the scaling evolution from chip to rack, the introduction of HBM4 requires interposers to support thousands of I/Os; the evolution from CoWoS-L to CoWoP enables PCBs to undertake substrate-like functions for the first time; the layer count for GB300 server PCBs has jumped from 10 layers to over 20 layers, with some high-end models reaching 34 to 64 layers; the Rubin Ultra NVL576 further replaces copper cables with a 78-layer M9-grade orthogonal backplane, handling full interconnect communication for GPUs within the rack. The core of industry competition is shifting from "single-card computing power" to "full-system interconnect bandwidth," making PCBs a critical bottleneck determining the efficiency of AI system computing power release, with technical thresholds and certification cycles comparable to semiconductor packaging.

Rubin ushers in the era of hardware density, with orthogonal backplanes driving a leap in PCB semiconductor-like value. NVIDIA's GTC 2025 unveiled the Rubin series roadmap, marking the beginning of a new era in AI hardware density: mass production of the Vera Rubin NVL 144 platform in the second half of 2026, delivering 3.6 EFLOPS FP4 inference computing power; mass production of the Rubin Ultra NVL576 platform in the second half of 2027, delivering 15 EFLOPS FP4 inference computing power. Jensen Huang proposed a new calculation rule where "GPU count is measured by the number of chips in the package," signaling the arrival of an era where packaging density becomes the core metric for hardware density. From an industry chain perspective, the Rubin series drives both price and volume growth for PCBs: in terms of volume, the Rubin Ultra rack doubles GPU packaging count, leading to a corresponding doubling in PCB usage; in terms of price, the platform adopts M8U/M9-grade high-end materials and ultra-high multilayer designs, increasing the PCB value per server by over twofold compared to the previous generation. The orthogonal backplane, as a landmark engineering innovation, uses a 78-layer PCB to interconnect GPUs and NVSwitch, replacing tens of thousands of copper cables. Goldman Sachs predicts that AI server demand will increase approximately 4.3 times from 2025 to 2030, with the supply-demand imbalance for high-end PCBs persisting until 2027. The proportion of PCBs in the AI system Bill of Materials is approaching that of semiconductor-grade components, completing a value transition from a "carrier platform" to a "core interconnect medium."

The combined empowerment of CoWoP and the M9 system propels AI PCB processes towards semiconductor-grade breakthroughs. The CoWoP solution eliminates the ABF packaging substrate and BGA solder balls, directly mounting the silicon interposer and GPU/HBM assembly onto a reinforced PCB. The PCB assumes all the functions originally performed by the packaging substrate, marking the disappearance of the boundary between PCBs and packaging substrates. This solution offers multiple advantages in signal integrity, power integrity, thermal management, substrate deformation control, and long-term reliability. The PCB value per GPU is as high as $600, three times that of the current GB 200 platform. It is estimated that this will create a market space exceeding $600 million by 2027, surging to over $2 billion by 2028. The concurrently evolving M9-grade copper-clad laminate system employs advanced materials such as third-generation Low DK quartz fabric and HVLP4/5 ultra-low profile copper foil, with filler usage doubling compared to the previous generation. The generational leap in materials leads to an exponential increase in processing difficulty. Coupled with upstream Nitto Denko's production capacity nearing its limit and tight HVLP copper foil supply, these three factors systematically elevate the value center. The combination of CoWoP and M9 brings process precision comprehensively close to semiconductor-grade levels.

Related overseas computing power targets: Shenghong Technology, Avary Holding, Wus Printed Circuit, Gowin Technology, Shengyi Technology, Kingboard Laminates Holdings, Dongshan Precision, Shihua Circuits. Other overseas computing power targets: Dongshan Precision, Foxconn Industrial Internet, Zhongji Innolight, TFC Optical Communication, China Tungsten High-tech, Tianyue Advanced, New Epoch, GigaDevice, Wus Printed Circuit, Dapu Storage, Yuanjie Semiconductor, Ouke Yitai, Envicool, Weike Technology, Lingyi Intelligent Manufacturing, etc.; Intel, SK Hynix, Lumentum, SanDisk, Qualcomm, Broadcom, Marvell, Kioxia, Micron, Advanced Micro-Fabrication Equipment, Naura Technology, Topanga Technologies, Changchuan Technology.

Risk warnings include potential underperformance in AI server shipments and PCB upgrades; risks associated with slower-than-expected commercialization of new processes like CoWoP and orthogonal backplanes; risks from tight raw material supply and price fluctuations; risks of intensified competition and price wars due to overly rapid industry capacity expansion; and risks related to order volatility from major clients and high customer concentration.