Xiangcai Securities: Major AI manufacturers expand data center switches upgrade to increase the value of PCB stand-alone machines

The Zhitong Finance App learned that Xiangcai Securities released a research report saying that major AI companies and major global Internet companies have invested huge capital expenses to expand data centers. According to the Dell'Oro report, the compound growth rate of global data center capital expenditure will reach 21% by 2029, and the huge capital expenditure plans of AI companies and major Internet companies will drive the continuous expansion of computing power clusters. Computing power expansion will simultaneously amplify “network bottlenecks”, turn switches from “accessories” into “key production capacity” to improve data center performance, and drive rapid growth in demand for high-speed switches. At the same time, the number of PCB layers in AI servers and high-speed switches continues to increase, and the materials used in PCBs are constantly being upgraded, driving a significant increase in the value of stand-alone PCBs.

The main views of Xiangcai Securities are as follows:

Mass production of Rubin architectures is expected to increase the value of stand-alone PCBs

At the GTC conference at the end of October, Hwang In-hoon publicly displayed the Vera Rubin superchip for the first time, equipped with a CPU codenamed Vera and two huge Rubin GPUs. In the Kyber architecture used by RubinUltra NVL576, orthogonal backplane technology will be applied. In Switchtray, Midplane, and CX9/CPX, multi-layer PCB boards will be directly connected instead of copper cables.

At the same time, the Rubin platform achieves low loss and low latency, and fully upgraded manufacturing materials. The Switch Tray uses M8U grade materials (low-DK2+HVLP4) and a 24-layer HDI board design, while Midplane and CX9/CPX use M9 materials (Q-Glass+HVLP4), with a maximum number of layers of 104 layers. This increases the PCB value of a single server by more than two times compared to the previous generation. Furthermore, Rubin's design logic has become the common language of the industry. ASIC AI servers, including Google TPU V7 and AWS Trainium3, also use high-grade HDI, low Dk materials, and extremely low roughness copper foil. Additionally, the NVL144 also uses Midplane in the Compute Tray instead of copper cables.

Xiangcai Securities believes that in the next-generation server structure, the replacement of copper cables by PCBs and the upgrading of PCB materials will greatly increase the value of individual server PCBs.

Switch upgrade boosts PCB stand-alone value

A data center switch is a high-performance networking device designed for large-scale data center environments for high-speed data exchange between servers, storage devices, and other network nodes. Currently, 400G switches are mainstream, accounting for 38% of shipments in 2024. 800G switches will be released in 2025, and mass production of 1.6T switches is expected to begin in 2026. As switch port rates are iterated, the number of PCB layers used by switches continues to increase, and the material grade used has been upgraded from M6 to M9, driving the value of its stand-alone PCB to continue to increase dramatically.

Demand for AI servers and high-speed switches will grow rapidly

As leading AI vendors aggressively expand data centers, such as OpenAI plans to build 250GW by 2033, Google plans to increase computing power by 1,000 times within 5 years, compounding the capital expenses of major North American internet companies to maintain a high boom and increase capital expenditure plans. According to the Dell'Oro report, the compound growth rate of global data center capital expenditure will reach 21% by 2029. Xiangcai Securities believes that the huge capital expenditure plans of major AI companies and major Internet companies will drive the continuous expansion of computing power clusters. TrendForce predicts that AI server shipments will increase by more than 20% in 2026, reflecting high demand.

Computing power expansion will simultaneously amplify “network bottlenecks”, turn switches from “accessories” into “key production capacity” to improve data center performance, and drive rapid growth in demand for high-speed switches. According to Dell'Oro's forecast, 800G+1.6T switch shipments will increase dramatically in the next 5 years. The front-end network is expected to deploy close to 90 million 800G and 1.6T switch ports, and the back-end network's shipment volume will exceed that of the front-end network by more than 3 times.

Risk Alerts

The development of AI applications fell short of expectations; the iteration of large models fell short of expectations; and the capital expenses of major AI companies and major Internet companies fell short of expectations.