Zheshang Securities released a research report stating that Alphabet's (GOOGL.US) fourth-quarter 2025 results comprehensively exceeded expectations, fully validating that its AI investments have entered a harvest period. Alphabet's management has guided for 2026 capital expenditures to reach $1.75-$1.85 trillion to alleviate persistently tight computing power supply, reflecting that both internal and external AI demand remains extremely strong. Some ASIC chips, represented by Alphabet's TPU, have begun to compete directly with Nvidia on cost after optimizations in raw computing performance and system efficiency. Drawing parallels to the impact of Nvidia's GPU architecture on PCBs, the architectural upgrades of ASIC chips are also expected to replicate this trend, thereby driving demand for high-value PCBs. The main points from Zheshang Securities are as follows:

Alphabet's Q4 results significantly surpassed expectations, leading to a substantial upward revision of its 2026 capital expenditure guidance. For the fourth quarter of 2025, Alphabet's performance comprehensively beat expectations, with revenue rising 18% year-over-year to $113.8 billion. Full-year revenue surpassed the $400 billion mark for the first time, and net profit surged by 30%, providing strong evidence that AI investments are yielding returns. Core businesses showed strength across the board: Google Search and other revenues grew by 17%, YouTube's annual revenue exceeded $60 billion, and Google Cloud delivered the most impressive performance, with revenue soaring 48% year-over-year to $17.7 billion. Its operating margin improved significantly by 12.6 percentage points year-over-year to 30.1%, and the backlog of orders increased 55% sequentially to $240 billion, highlighting robust enterprise AI demand. The Gemini ecosystem is advancing rapidly; version 3.0 became the fastest-adopted model in the company's history, with monthly active users of applications exceeding 750 million and enterprise paid seats surpassing 8 million. Furthermore, the cost per service unit decreased by 78%, indicating continuously enhanced AI monetization capabilities. Notably, the company entered a significant collaboration with Apple to develop the next-generation foundational model, while Waymo secured new funding and accelerated its global expansion. Management's guidance for 2026 capital expenditures reaching $1.75-$1.85 trillion aims to ease the ongoing tight supply of computing power, reflecting that AI demand remains exceptionally strong both internally and externally.

ASIC chips are catching up strongly, with Alphabet's TPU achieving cost-performance parity with Nvidia's GPU. In the early stages of AI development, training compute power was almost the sole determinant, which is why Nvidia's GPU chips have dominated the global AI server chip market in recent years. However, as large models gradually enter the deployment and commercialization phase, inference workloads are beginning to surpass training, bringing cost considerations to the forefront. Some ASIC chips, exemplified by Alphabet's TPU, have now become competitive with Nvidia on cost after optimizations in raw computational performance and system efficiency. The upgrade from TPUv6 to TPUv7 reduced the inference cost per token by approximately 70%, bringing its absolute cost essentially in line with, and in some estimates even slightly advantageous to, Nvidia's GB200 NVL72 platform. Consequently, the proportion of TPU usage within Alphabet's internal workloads is continuously increasing, and it is now widely used for both training and inference of the Gemini model.

AI server architecture upgrades are driving increases in both the volume and value of PCBs. During the architectural evolution of Nvidia's GPUs, both the quantity and value of PCBs used increased. Prior to the Hopper architecture, the PCBs used in Nvidia server modules were primarily the OAM accelerator card (HDI) and the UBB backplane (high-layer count). In the Blackwell architecture, a switch board using HDI technology replaced the UBB backplane. Future architectures are expected to introduce orthogonal backplanes with over 70 layers for connecting the computer tray and switch tray. Referring to the impact of Nvidia's GPU architecture on PCBs, it is evident that architectural upgrades in ASIC chips are likely to follow a similar trend, driving demand for high-value PCBs. For instance, Alphabet's TPUv8 is expected to newly incorporate a switch board utilizing HDI technology. Looking ahead, the introduction of orthogonal backplanes, along with the substrate-like and modularization trends for PCB boards, are anticipated developments in the ASIC server domain.

Regarding investment targets, whether through the introduction of new board types or upgrades towards substrate-like and modular designs, the underlying essence is a significant increase in the value captured within the PCB manufacturing segment. Leading PCB enterprises possessing comprehensive process technology capabilities and key client relationships are clearly better positioned to capture this value increment amid the broader trend of high-performance computing growth driven by AI. In light of Alphabet's upward revision of its 2026 capital expenditure guidance, it is recommended to focus on the PCB industry chain, including companies such as Guanghe Technology (301389.SZ), Wus Printed Circuit (002463.SZ), Shennan Circuits (002916.SZ), and Victory Giant Technology (300476.SZ).

Risk factors include potential disruptions to shipments of Alphabet's TPU chips and the possibility that architectural upgrades to Alphabet's TPU chips may have a lesser-than-expected impact on driving PCB demand.