CICC International: AI demand drives the coordinated upgrade of computing efficiency and internet bandwidth, and material demand in the first half of 2026 is expected to see rapid growth.

CICC International released a research report stating that AI inference demand catalyzes cloud vendors' capital expenditures, and computing efficiency and interconnection bandwidth are synergistically upgraded. AI PCB is a core incremental link in the wave of AI infrastructure upgrades. The three major raw materials for AI PCB, electronic cloth, copper foil, and resin, are the core barriers to building PCB dielectric performance. The bank expects that the upstream supply chain of Rubin servers will start the stockpiling tide in the first half of 2026, by which time the core raw materials of M8.5 and M9 PCB/CCL are expected to reach a crucial "from 0 to 1" juncture. Considering that Rubin servers will start mass production and delivery in the second half of 2026, the bank believes that the upstream supply chain will start the stockpiling tide in the first half of 2026, and the demand for AI materials is also expected to experience rapid growth. CICC International's main points are as follows: Low dielectric performance is a key indicator for AI PCB design. GPU and ASIC manufacturers are actively improving the single-chip computing power efficiency and cabinet-level interconnection bandwidth. In the trend of upgrading AI PCB to higher layers and smaller line widths, it will face problems such as electrical performance loss, decreased heat dissipation performance, and signal interference. PCB made with low dielectric constant (Low-Dk) and low dielectric loss (Low-Df) materials is crucial for reducing channel losses and maintaining signal integrity. The bank believes that regardless of changes in the downstream GPU and ASIC competitive landscape, the trend of AI infrastructure pursuing higher computing efficiency and larger interconnection bandwidth will not change, and the technical exploration of upstream low dielectric performance materials will continue to advance. The three major raw materials of AI PCB, electronic cloth, copper foil, and resin, construct the core barriers of PCB dielectric performance. The strict requirements of AI infrastructure for data transmission loss are driving the upgrade of PCB and CCL to M8/M9. In terms of electronic cloth, quartz fiber is the preferred material for its excellent dielectric loss (Df value at 1MHz is 0.0001) and thermal expansion coefficient (0.54ppm/C). In terms of copper foil, HVLP4/5 is the preferred material for its extremely low surface roughness (Rz1.5m). For resin, PCH resin and PTFE resin are the preferred materials for their excellent dielectric performance. The core raw materials of M8.5 and M9 PCB/CCL are expected to reach a crucial "from 0 to 1" juncture. The bank expects that the Compute Tray/Switch Tray/Midplane/CPX PCB and CCL solutions corresponding to NVIDIA's Rubin servers will respectively upgrade to M8/M8.5/M9/M9 solutions, with the M9 solution potentially using a material combination of high-frequency, high-speed resin + HVLP4/5 copper foil + Q cloth, and the M8.5 solution potentially using a material combination of high-frequency, high-speed resin + HVLP4 copper foil + Low-Dk second-generation cloth. According to reports from Wall Street News and Sina Finance, NVIDIA is expected to mass produce Rubin GPUs in October 2026. The bank believes that the Rubin Ultra server is expected to adopt a solution of M9 resin + high-end HVLP copper foil + Q cloth orthogonal backplane. The bank predicts that the upstream supply chain of Rubin servers will start the stockpiling tide in the first half of 2026, by which time the core raw materials of M8.5 and M9 PCB/CCL are expected to reach a crucial "from 0 to 1" juncture. The market size of AI-related materials is expected to experience rapid growth. Based on calculations, the bank estimates that the global market size of CCL raw materials for HDI boards and high-layer boards of 18 layers and above in 2025/2029 is approximately 30.98/38.91 billion US dollars, with the electronic cloth market size at approximately 7.75/9.73 billion US dollars; the copper foil market size at approximately 12.39/15.56 billion US dollars; and the resin market size at approximately 7.75/9.73 billion US dollars. Considering that NVIDIA's Rubin servers will start mass production and delivery in the second half of 2026, the bank believes that the upstream supply chain will start the stockpiling tide in the first half of 2026, and the demand for AI materials is also expected to experience rapid growth.