The Zhitong Finance App learned that Hualong Securities released a research report saying that the 2026 World Artificial Intelligence Conference will be held in Shanghai from July 17 to 20, and that supernode products will be exhibited centrally, proving that domestic computing power hardware has moved from “easy to use” to “easy to use”, and that system-level competitiveness is taking shape. The release of Kimi K3 proved that the large domestic model already has the world's first-tier model capabilities, and that this capability continues to be iterated, and each round of iteration means greater computing power requirements. Maintain the computer industry's “Recommended” rating.
The main views of Hualong Securities are as follows:
WAIC Review: Focus on China's Supernode Solutions
The 2026 World Artificial Intelligence Conference will be held in Shanghai from July 17 to 20, and supernode products will be showcased. For the first time, Huawei exhibited the Shengteng 950 supernode offline. Based on the Lingqu interconnection protocol and supernode architecture, it realizes the industry's largest 1024 card scale interconnection, provides 1 EFLOPSFP8 and 2 EFLOPSFP4 surging computing power, has 256 TB of global unified memory addressing space, and has an RTT delay of only 3 microseconds. At the same time, Ali showcased the Pingtou Gezhenwu M890 AI chip with Alibaba Cloud's Panjiu AL128 supernode server; the Xinhua Third Group, a subsidiary of Ziguang Co., Ltd., showcased the Xinhua-3 UniPodS8000 supernode. The product covers a full range of products from 32 cards to 1024 cards, which can be flexibly expanded up to 16,384 cards, improving training performance by 70% and improving reasoning performance by 3 times; ZTE launched the OEX Super Node; Mu Xi Co., Ltd. showcased the Xijing S600 supernode; Zhongke Shuguang showcased the nationally produced 100,000-card Shuguang AI supernode. The bank believes that the transition from sample to large-scale display has released a clear signal that domestic computing power continues to break through. Currently, the competitive dimension of domestic AI chips has extended from single-card computing power parameters to cluster interconnection and system-level computing power delivery.
Kimi K3 is in the first tier of global open source, and the domestic model is expected to start the path of advancement
On July 17, Dark Side of the Moon released a new generation model, Kimi K3, with total parameters of 2.8 trillion yuan. It is currently the open source model with the largest parameters in the world. K3 is built based on KDA's hybrid linear attention mechanism and attention residual technology, and has a contextual window of 1 million tokens. Compared to K2, the overall expansion efficiency of K3 is increased by about 2.5 times. In terms of performance, K3 quickly entered the world's first tier after release. In various evaluation lists for various scenarios such as programming, K3 scored second only to Claude Fable 5 and GPT-5.6 Sol. In terms of pricing, the K3 output price is 100 yuan per million tokens (equivalent to about 15 US dollars). It is already in the same price system as the North American model, breaking the previous line where the domestic model focused on cost performance. Furthermore, at the end of June this year, Meituan released and open-sourced a next-generation trillion parameter model, LongCat-2.0. This model became the first trillion parameter model in the industry to complete a full process of training and inference on a 50,000 card domestic computing power cluster, verifying the viability of domestic computing power to support model training and closed loop reasoning.
Google's cloud business has grown dramatically while continuing to increase capital expenses
Google's parent company Alphabet released 2026 Q2 earnings report on July 22. In terms of cloud business, the company's Q2 revenue was US$24.8 billion, with a year-on-year growth rate of 82%. In terms of capital expenditure, the company raised its capital expenditure guidance for the full year 2026 from US$180-1900 billion to US$1950-205 billion, and made it clear that capital expenditure in 2027 will be significantly higher than 2026. The capital expenditure of overseas cloud companies continues to increase, providing a new anchor for future investment in domestic cloud companies, and domestic computing power is also expected to gain greater flexibility due to the acceleration of domestic replacement.
Risk warning: Domestic computing power construction falls short of expectations; risk of errors in quoted data; AI applications fall short of expectations; technology iteration speed falls short of expectations; focus on company performance falling short of expectations; introduction of policy standards falls short of expectations; declining risk appetite has led to increased capital market fluctuations.