According to reports from Zhito Finance APP, Tianfeng released a research report stating that AI Agent is rapidly developing, and major overseas companies are actively laying out applications for Agent construction. From Claude 3.5 to GLM, the application of Agents on C-end hardware is reaching a critical point. The C-end entry is undergoing reshaping, and internet giants are expected to accelerate their reverse AI hardware positioning layout, while the cloud collaborative technology route is likely to bring an increase in cloud reasoning computing power.

Recommendations:

Domestic computing power: Cambricon (688256.SH), Haiguang Information (688041.SH), Shengke Communication (communication coverage), Digital China Group (00861), Dawning Information Industry (603019.SH), Changsha Jingjia Microelectronics (300474.SZ).

AI applications: Beijing Kingsoft Office Software, Inc, Kingdee Int'l, iflytek co.,ltd., shanghai weaver network, beijing seeyon internet software corp., Dingjie Smart, yonyou network technology, hand enterprise solutions, hundsun technologies inc., wondershare technology group, arcsoft corporation.

C-side hardware: Qualcomm (US stocks), Jingxin Technology, Zhongke Lanxun, bestechnic (shanghai) co., ltd.

Tianfeng Securities' main points are as follows:

AI Agent is rapidly developing, with major overseas companies actively laying out applications for Agent construction.

The global AI Agent market is reshaping the operation models and customer interaction experiences across various industries at a high-speed growth trend. AI Agent, through memory capability, planning capability, action capability, and tool capability, achieves efficient collaboration with human users, external environments, other agents, and system developers. LLM, as the core competency foundation of AI Agent, combines four capability modules: planning, memory, tools, and actions, realizing the understanding, decomposition, and execution of complex tasks. Leading overseas enterprises are continuously strengthening their efforts in AI Agents, with Microsoft announcing the world's largest enterprise-level AI Agent ecosystem at the Ignite conference, while M365 Copilot has also added more Agent features; Salesforce is set to release Agentforce on September 12, 2024, aiming to integrate AI agents with human collaboration, data cloud, CRM, and other core modules to provide comprehensive customer service and sales solutions for enterprises.

From Claude 3.5 to GLM, the implementation of Agents in consumer-end hardware is approaching a critical point.

Claude 3.5 launched the latest Sonnet model, adding Computer use capability, allowing it to use a computer similarly to a human. After inputting language commands, the model perceives interface interactions, further advancing natural language control of hardware. Meanwhile, Zhizhu launched AutoGLM, realizing natural language interaction capabilities on Android system phones and web platforms; among hardware manufacturers, Apple took the lead in launching Apple Intelligence, while Huawei also introduced Harmony Intelligence.

The consumer-end entry is undergoing a transformation, and internet giants are expected to accelerate the reverse layout of AI hardware positioning while cloud collaboration technology routes are expected to bring additional cloud inference computing power.

Edge-side AI is expected to become a key entry point for the next generation of traffic, with hardware manufacturers of mobile phones, headsets, etc., collectively joining the large model capabilities. Leading with Apple Intelligence, Samsung, Huawei, and other hardware manufacturers are exploring the integration of natural language processing capabilities into their new systems, which may simplify hardware interaction processes. Internet companies are expected to leverage models and software capabilities to reverse position for traffic entry, and in the long term, specialized edge-side and versatile cloud collaboration may be the optimal solution for edge-side AI, with AI Agent being an important part of edge-side AI. Considering that AI Agent requires planning and multiple calls to large models, for example, Anthropic's newly launched Claude Sonnet model shows relatively high costs when using the Computer Use capability. Tianfeng believes edge-side AI will bring a significant increase in cloud inference computing power.

Risk Warning: The progress of AI applications may not meet expectations, risks of computing power being sanctioned, and advancements in large model technology may lag behind expectations.