Xiangcai Securities released a research report stating that AI giants and global internet companies are investing heavily in capital expenditures to expand data centers. According to Dell’Oro, the compound annual growth rate (CAGR) of global data center capital expenditures will reach 21% by 2029. The massive capital expenditure plans of AI and internet giants will drive continuous expansion of computing power clusters.

The expansion of computing power will simultaneously amplify "network bottlenecks," transforming switches from "supporting components" into "critical capacity" for improving data center performance, thereby accelerating demand for high-speed switches. Additionally, the number of PCB layers in AI servers and high-speed switches continues to increase, with materials also being upgraded, significantly boosting the value of PCBs per unit.

Xiangcai Securities highlights the following key points:

**Rubin Architecture Mass Production Expected to Increase PCB Value per Unit** At the GTC conference in late October, Jensen Huang publicly unveiled the Vera Rubin superchip for the first time, featuring a CPU codenamed Vera and two large Rubin GPUs. The Kyber architecture used in Rubin Ultra NVL576 will adopt orthogonal backplane technology, replacing copper cables with multilayer PCBs in Switchtray, Midplane, and CX9/CPX connections.

To achieve low loss and low latency, the Rubin platform has upgraded manufacturing materials. Switch Tray uses M8U-grade materials (Low-Dk2+HVLP4) and a 24-layer HDI board design, while Midplane and CX9/CPX incorporate M9 materials (Q-glass+HVLP4), with layers reaching up to 104. This increases the PCB value per server by over two times compared to the previous generation.

Furthermore, Rubin’s design logic has become an industry standard, with ASIC AI servers like Google TPU V7 and AWS Trainium3 also adopting high-layer HDI, low-Dk materials, and ultra-low roughness copper foil. NVL144 also replaces copper cables with Midplane in Compute Tray. Xiangcai Securities believes that in next-gen server architectures, PCB’s replacement of copper cables and material upgrades will significantly increase PCB value per server.

**Switch Upgrades Drive Higher PCB Value per Unit** Data center switches are high-performance networking devices designed for large-scale environments, facilitating high-speed data exchange between servers, storage, and other nodes. Currently, 400G switches dominate the market, accounting for 38% of shipments in 2024, while 800G switches will scale up in 2025, and 1.6T switches are expected to enter mass production in 2026.

As switch port speeds evolve, the number of PCB layers increases, and materials upgrade from M6 to M9, driving substantial growth in PCB value per unit.

**Rapid Growth in AI Servers and High-Speed Switches** With AI leaders aggressively expanding data centers—such as OpenAI’s plan to build 250GW by 2033 and Google’s goal to increase computing power 1,000-fold in five years—coupled with sustained high capital expenditures from North American internet giants, Dell’Oro projects a 21% CAGR in global data center capital expenditures by 2029.

Xiangcai Securities argues that massive capital spending by AI and internet giants will fuel ongoing computing power expansion. TrendForce forecasts AI server shipments to grow over 20% by 2026, reflecting strong demand.

Computing power expansion will exacerbate "network bottlenecks," turning switches into "critical capacity" for enhancing data center performance and accelerating demand for high-speed switches. Dell’Oro predicts that 800G and 1.6T switch shipments will surge in the next five years, with nearly 90 million ports deployed in front-end networks and over three times that volume in back-end networks.

**Risk Warnings** Potential risks include slower-than-expected AI application development, delays in large model iterations, and lower capital expenditures from AI and internet giants.