China Securities Co., Ltd.: From strategic defense to proactive attack - Value upgrade of domestic AI infrastructure

China Securities Co., Ltd. released a research report stating that with the continuous tightening of export controls and the resonance of the institutional dividend of domestic substitution, domestic AI chips are transitioning from "policy procurement targets" to "market-based proactive choices", and the next few years are a key window for transitioning from scale verification to profit realization. The company sees continuous release of domestic chip production capacity on the supply side, explosive growth in demand-side reasoning loads (daily average Token call volume increased by over 1,000 times in two years), and continued capital expenditure by leading CSPs (Alibaba FY26 reached 126.1 billion yuan, ByteDance 2025E approximately 160 billion yuan) providing a certainty of demand anchor for domestic chips. On the realization side, profit margins follow a three-stage path of "scale dilutionstructural optimizationecosystem premiums", currently in stage one of exchanging quantity for price, with the future realization of system-level solutions through software+hardware binding leading to a significant optimization of profit margins. The main points of China Securities Co., Ltd. are as follows: Against the backdrop of the continued tightening of export controls and the resonance of the institutional dividend of domestic substitution, domestic AI chips are transitioning from policy procurement targets to market-based proactive choices. The company believes that the next 2-3 years will be a key window period for domestic AI chips to transition from scale verification to profit realization, with the core contradiction shifting to "whether it can be delivered at scale and iterated continuously". This report answers two core questions: 1) Spatial issue: training/inference demand growthdomestic substitution rateeffective supply (process/packaging/HBM/cluster delivery)=potential market space. The three are multipliers, with any one being limited constituting a bottleneck. 2) Realization problem: Profit margin improvement follows a three-stage path of "scale-drivenstructural optimizationecosystem returns", currently in the phase of exchanging quantity for price and diluting fixed costs, with the medium-term transition to stage two (commercialization of software stack, premium on cluster solutions), and stage three (increase in the proportion of high-margin revenue from IP/software, continuous improvement in net profit margin). On the market demand side, the total shipments of Chinese AI accelerators in 2025 reached approximately 4 million units, of which about 1.65 million units (41%) were domestically produced. Nvidia's market share in China has decreased from about 95% to about 55%, indicating a clear structural relinquishment space. On the demand side, the daily average Token call volume increased from approximately 100 billion in early 2024 to over 14 trillion in March 2026, with explosive growth in reasoning loads being a major marginal driver for the mass production of domestic chips in 2026-2027. Leading CSPs continue to increase capital expenditures, with Alibaba expecting an annualized Capex of 130 billion yuan for the next three years and Tencent expecting around 1300-1600 billion yuan in 2026, with ByteDance set to spend about 160 billion yuan in 2025 (with 90 billion for AI computing power), providing a strong anchor of certainty for domestic chip demand. On the supply constraint side, there are bottlenecks in process, HBM, and advanced packaging. In terms of process, SMIC N+2 has entered mass production but with lower yield than TSMC, and N+3 requires multiple graphics and significantly higher lithography costs, with mid-term availability of capacity improvement. HBM is the weakest link in the supply chain, accounting for nearly 50% of the AI accelerator BOM, with ChangXin's HBM2 already in mass production but HBM3/3E still in the research stage. Advanced packaging such as CoWoS is a prerequisite for mass production, with ChangXin planning to produce 5,000-8,000 pieces per month by 2026 (yielding 75-80%), and the availability of packaging capacity is becoming a bottleneck earlier than yield rate, and a prerequisite for the realization of shipment guidance. Pingtouge (Alibaba 100% owned) The core differentiation lies in the internal business loads of the parent company (Tongyi series models) and the zero marginal resistance on computing power digestion. The Zhu Wu series has shipped a total of 560,000 pieces (as of May 2026), with a customer customization rate of over 60% and ranking as the second in the country in 2025. The release of ICNSwitch 1.0 in May 2026 marks the competition dimension expanding from single card to cluster system computing power layer. The product roadmap is clear, with 810E (in production) M890 (start production delivery) V900 (Q3 2027, significant performance improvement) J900 (Q3 2028, significant performance improvement). The main risks lie in the maturity of software compatibility with non-standard architectures such as DeepSeek V4 (UE8M0, MLA, MoE All-to-All), PPU integration with Alibaba Cloud MaaS selling, potential loss of flexibility in facing external customer deep customization, and the possibility of bandwidth becoming a bottleneck in inter-card communication in scenarios with tens of thousands of cards, with the risk of technology route transformation after crossing 128 cards in ultra-node solutions. Kunlun Core (Baidu Holdings) The core advantages lie in cluster engineering capacity verification (P800 achieving 32,000 cards lighting up, training efficiency>98%) and the strong adherence brought by the hard-coded binding of the PaddlePaddle ecosystem. The "Five Years, Five Cores" roadmap (2025 P8002026 M1002027 M3002028 thousand-card super node2029 N series2030 million-card single cluster) is the most complete and clearest public roadmap among domestic AI chip manufacturers. The capitalization process is leading, with HKEx A1 submitted (January 2026) and Sci-Tech Board coaching and filing completed (May 2026). The main risks lie in the lack of a new generation flagship training card in 2026 (M300 to be waited until 2027), facing competition pressure from CambriconMLU690 and Ascend 950; Tianchi ultra-node is a physical carrier bound by XTCL+PaddlePaddle hardcoded, facing resistance in promoting outside the PaddlePaddle ecosystem (for native PyTorch customers); the performance of Baidu series models still lags behind the top domestic players, with differences in internal and external technical routes possibly involving team development efforts, and facing challenges in fully transitioning to serve external customers in the future. Key risks include EDA discontinuation (still lacking back-end sign-off tools for 7nm), insufficient maturity of software ecosystems, constraints on neutrality within the parent company, competition pressure if Nvidia/AMD are approved to enter, and continuous challenges in chip adaptation layers due to rapid iteration of model architectures.