HuaChuang Securities: Strong demand for domestic intelligent computing chips, recommended to focus on the computing power industry direction

According to the Zhitong Finance APP, Huachuang Securities has released a research report stating that global demand for intelligent computing power continues to be robust, and the domestic intelligent computing market has vast potential. By 2025, the CAGR of China's intelligent computing chip market is expected to reach 53.7%, with GPU market share rising to 77.3%. Policies such as the second phase of the East Data West Computing initiative are driving the domestic GPU shipment ratio to exceed 30%, indicating that a commercialization inflection point has emerged. Hardware performance is presenting a multi-dimensional technological route, and the competitiveness of domestic computing power solutions is being validated. With the implementation of the "Computing Power Interconnection Capability Requirements," the "One Degree of Computing Power" measurement standard, and the 400G ROADM network, national computing power is upgrading from "available" to "tradeable, schedulable, and measurable," with a commercialization inflection point already evident. The firm suggests paying attention to the computing power industry direction.

The main viewpoints of Huachuang Securities are as follows:

Global demand for intelligent computing power remains robust, and the domestic intelligent computing market has vast potential

1. Demand side: In the United States, OpenAI's "Stargate" is working to build a $500 billion cluster, xAI Colossus with 200,000 cards, and Meta's 1.5 GW Prometheus, along with the UAE's 5 GW sovereign AI park and 13 AI factories in the EU, continuously validating the simultaneous expansion of ultra-large-scale training and sovereign intelligent computing investments. Domestically, the daily consumption of tokens for large models is expected to leap from the hundreds of billions level at the beginning of 2024 to the trillions level by March 2025, with vertical applications rapidly landing and inference demand expanding exponentially.

2. Capital side: According to Frost & Sullivan, the CAGR of China's intelligent computing chip market from 2025 to 2029 is 53.7%, with GPU market share rising from 69.9% to 77.3%; according to Kezhi Consulting, since 2023, new AIDC investments have mainly been driven by ByteDance, excluding basic telecom operators; by the end of 2024, domestic AIDC cumulative investment is expected to exceed 187.5 billion yuan. Looking ahead to 2025, leading internet companies such as Alibaba, Tencent, and Huawei plan to simultaneously expand capital expenditures, coupled with new procurement from ByteDance, laying the foundation for sustained industry expansion.

3. Policy side: The U.S. entity list continues to expand, and domestic substitution is expected to accelerate; in addition, the country is systematically promoting the national integrated computing power network work according to the "point, chain, network, surface" system, emphasizing "moderately advanced construction of network facilities and enriching computing power application scenarios"; the second phase of the East Data West Computing initiative, computing power vouchers, and local intelligent computing centers form rigid quotas on the demand side, providing vast space for the domestic GPU market. According to IDC data, in 2024, the shipment ratio of China's intelligent computing chips is expected to exceed 30%.

Hardware performance is presenting a multi-dimensional technological route, and the competitiveness of domestic computing power solutions is being validated

1. In the advanced process field, global competition in advanced chip manufacturing is becoming increasingly fierce, with international giants continuously breaking physical limits. Although there are still bottlenecks in domestic advanced manufacturing processes at 7nm and below, the pace of technological catch-up is accelerating. SMIC's 14nm yield is >95%, and 7nm mass production is steadily advancing; Huawei's Ascend 910C, using 7nm+EUV technology, pushes FP16 computing power to 448 TFLOPS, with a 15% reduction in power consumption compared to the 910B

2. In terms of single card benchmarking, Huawei's Ascend 910C adopts Chiplet dual-core packaging, with a 1024-chip cluster FP16 computing power reaching 819.2 PFLOPS, comparable to the NVIDIA H100 cluster, supporting the training of models with hundreds of billions of parameters; Cambricon's Siyuan 590 supports 8-chip cascading, with an FP16 cluster computing power of 2.048 PFLOPS.

3. In terms of architectural innovation, Huawei's CloudMatrix 384 single cabinet integrates 384 Ascend 910 chips and 192 Kunpeng CPUs, with an FP16 peak of 300 PFLOPS, cross-node latency <1 μs, linear scalability of 95%, and a pre-filled efficiency of 4.45 tokens/s/TFLOPS in DeepSeek-R1, outperforming SGLang's 3.18 tokens/s/TFLOPS under the default configuration of NVIDIA H100; validating that domestic supernodes have the capability for large-scale commercial use.

Software compatibility adapts to independent standards, domestic ecosystem gradually breaks through CUDA monopoly

NVIDIA has formed a "chip-cluster-cloud" closed loop with 15 years of CUDA accumulation, with millions of developers worldwide relying on its CUDA-XAI foundation, PyTorch/TensorFlow frameworks, and DGX Cloud services. In March 2024, EULA will further prohibit the operation of translation layers, blocking third-party compatibility paths. Domestic companies are breaking the blockade through independent research and collaborative innovation: Huawei CANN, Haiguang DTK, Cambricon BANG+MagicMind and other self-developed software stacks are being implemented simultaneously; SenseTime's DeepLink has completed uninterrupted training of a model with hundreds of billions of parameters across more than 10 domestic chips in 20 days, maintaining an efficiency of 95%, scoring full marks in four categories of the "New Generation Artificial Intelligence Infrastructure White Paper"; the China Academy of Information and Communications Technology's AISHPerf, Wuwen Chip Qiong's Infini-AI, and Jietiao Xingchen's "Model Core Alliance" are promoting the adaptation of domestic models and chips towards full-link collaboration. With the implementation of the "Computing Power Interconnection Capability Requirements," "One Degree of Computing Power" measurement standards, and 400G ROADM networks, national computing power is upgrading from "available" to "tradeable, schedulable, and measurable," with a commercial turning point already in sight.

Investment Recommendation: Focus on the computing power industry direction

1. Chips: Cambricon, Haiguang Information, Jingjia Micro, Chipone Technology, Rockchip, Loongson Technology; 2) Servers: Zhongke Shuguang, Inspur Information, Huaqin Technology, Wolle Material, Oulu Tong; 3) Data Centers: GDS Holdings, Kehua Data, Range Technology, UCloud, Baosight Software, Kingsoft Cloud, Wangsu Science and Technology, Guanghuan New Network, Century Internet; 4) Switches: RJWL, Unisplendour, Guangxun Technology, Huagong Technology; 5) Hardware/End-side: Hengxuan Technology, Rockchip, Quectel, Espressif Technology, Zhongke Chuangda, Intellifusion, Zhiwei Intelligent, etc.

Risk Warning: The application of large models may not meet expectations; technological development may not meet expectations; market competition may intensify