According to the news from Zhitong Finance APP, Orient has released a Research Report stating that NVIDIA's GTC conference introduced several AI hardware and software products, marking the official entry of AI technology into the Agentic AI era, with computational power demand expected to see a hundredfold increase. Domestic Internet giants are accelerating capital expenditure on AI Infrastructure, with TENCENT's capital expenditure in Q4 2024 showing a year-on-year increase of over 200%, and Alibaba planning to invest 380 billion yuan over the next three years to build Cloud and AI infrastructure. The firm is Bullish on the new AI cycle driving the computational power - algorithm - application ecosystem to continue progressing successively, recommending an increase in the allocation to the Hong Kong Internet Sector. Key recommendations include Alibaba-W (09988), TENCENT (00700), and KUAISHOU-W (01024).

The main points from Orient are as follows:

NVIDIA's GTC conference released multiple heavyweight AI hardware and software products.

At the 2025 GTC conference, NVIDIA released several significant products and technologies, with notable optimizations in hardware, model architecture, and algorithms. The scale and utilization efficiency of AI computational power continue to improve, and the Scaling law of large models is expected to push breakthroughs in AI large models. Additionally, Jensen Huang delivered a keynote speech themed on the Agentic AI era at the GTC conference, announcing that AI will transition from the generative AI era to the Agentic AI era, at which point computational power demand will see a hundredfold increase. The firm believes that, on one hand, Deepseek's efficient and low-cost powerful inference capability, along with the high margins of its MaaS service, validates the feasibility of the business model for inference-based AI. On the other hand, the emergence of AI agents like Manus and the new Quark has showcased the paradigms of AI agent products for the first time. As mentioned at the GTC conference, computational power will become an indispensable resource in the AI era. Therefore, the firm remains Bullish on domestic cloud computing leader Alibaba and TENCENT, which has a solid user base and forward-looking AI application deployment.

TENCENT is significantly increasing investment in AI, with capital expenditure growing sharply, as the domestic AI Industry Chain enters an upturn cycle.

This week, TENCENT announced its Q4 2024 Earnings Reports, showing a year-on-year increase of more than double in capital expenditure for 2024. Based on institutional expectations, assuming the capital expenditure in 2025 accounts for 12%-15% of income, it corresponds to TENCENT's capital expenditure of 86.4-108 billion. Coupled with Alibaba's ambitious strategy to invest 380 billion in capital expenditure for Cloud and AI hardware infrastructure over the next three years, and the positive and optimistic attitudes of major internet companies like ByteDance and Baidu, as well as telecom operators like Unicom and Mobile towards AI investment, the firm believes that the current AI industry chain boom led by Deepseek has propelled the Industry into a growth cycle, with significant capital support for both duration and intensity. The investment on the upstream computational power side will be effectively converted into growth in downstream AI computational power and large model product supply, and the cost reductions driven by efficiency improvements will significantly enhance AI demand and B/C penetration rates. Cloud Computing and AI application services are expected to experience high growth, leading to a re-evaluation of the cloud and AI businesses of relevant Internet leading companies.

The AI foundational model application is leading, and TENCENT is considered bullish on its position as a leader in AI ToC applications.

TENCENT's AI large model layout is comprehensive, possessing several industry-leading large models including the Hongyuan series of large language models and multimodal models such as Hongyuan video and Hongyuan images. The newly released deep thinking large model Hongyuan T1 has reached an industry-leading level, and the AI foundational model technology is domestically advanced. On the application side, Yuanbao topped the App Store's free chart in February, with DAU increasing nearly 20 times. Additionally, WeChat, a national-level application, supports the user base. The group's internal advertising, gaming, and WeChat multiple businesses can rely on AI empowerment. It is believed that TENCENT has significant advantages in AI application scenarios and users, and is likely to achieve the leading position in AI ToC applications.

Key developments in the AI industry this week.

1) On the industrial policy side: The Shanghai municipal government has released "Several Measures to Promote Innovation and Development of the Service Industry," explicitly mentioning support for AI industrial development, and has issued policies for industrial funding and technical talent support, likely driving the AI industry to experience systemic policy benefits. 2) On the large model side: Kunlun Tech has released the first multimodal deep reasoning large model R1V, which introduces reasoning paradigms into multimodal large models for the first time. 3) On the algorithm side: The METR team has found that large models can double their sustainable task duration on average every seven months, with accuracy rates significantly improving, as AI intelligence continues to accelerate. The Zoom team has developed a new draft chain reasoning paradigm that effectively enhances the efficiency of large AI models in solving simple problems while significantly reducing computational power consumption. 4) On the AI application side: The ByteDance team has released an open-source integrated AI agent product TARS similar to Manus, performing remarkably well in multiple benchmark tests. It is recommended to pay attention to the subsequent developments of AI agent products.

Risk Warning

Industry development and AI application implementation have not met expectations, changes in assumptions affect calculation results, macroeconomic risks, and overseas policy risks.