The Zhitong Finance App learned that Guohai Securities released a research report saying that North American internet giants raised their 2024 AI capital expenditure guidelines to verify the increase in demand for AI computing power in the big model competition. It is expected that with the iteration of large models and the increase in the number of AI application users, the AI computing power market will open up more market space. AI computing power is expected to continue to grow, and sectors such as GPUs, AI servers, server cooling, optical modules, and data centers are expected to continue to benefit. Maintain a “recommended” rating for the computer industry.
▍ The main views of Guohai Securities are as follows:
Origin: Internet capital expenditure is invested in software and hardware, cloud computing, etc., and computing power is the main component
Internet capital expenditure mainly refers to ICT-related expenses in Internet companies. According to Gartner, global IT spending is expected to reach 5.1 trillion US dollars in 2024, 8% compared with the same period last year. Of these, computing power is the main component of expenditure. The big model competition will drive continued growth in Internet capital expenditure. According to the Dell'Oro Group, AI infrastructure spending is expected to drive data center capital expenditure to exceed 500 billion US dollars in 2027, and more than 20% of global server deployments may be an accelerated type.
Turning point: ChatGPT and other big models released to increase computing power demand and increase internet capital expenditure
1) Overseas: ① Microsoft: OpenAI (Microsoft) released pre-trained GPT-4v and GPTS. Microsoft expects FY2024Q3 capital expenditure to achieve substantial growth; ② Google: Releasing the conversational AI service Bard and the multi-modal large model Gemini, Google expects a gradual increase in capital expenditure in 2024; ③ Meta: Launched the open source big language model LLAma2, and Meta's 2024 capital expenditure guidelines adjusted the upper limit to US$30-37 billion.
2) Domestic: As of January 2024, 42 major domestic models have passed the filing and approval. Baidu Wenxin's big model 4.0 has been fully upgraded, and the parameter scale is expected to exceed trillion; Ali Tongyi Qianwen 2.0 was released, with parameters reaching 100 billion; ByteDance's “Coze Button” AI Bot development platform was launched. Both 2023Q3 and BAT said they will continue to increase investment in artificial intelligence, and indicated that they will not only successively empower the main business with AI models, but also that they will continue to increase investment in artificial intelligence models.
Outlook: Large model iteration and application innovation will increase computing power demand and Internet capital expenditure
1) Training: “Big model big data” has become a “new paradigm” for pre-trained models. According to SemiAnalysis, GPT-4 contains a total of 1.8 trillion parameters, and AI computing power will move from a stand-alone machine to a cluster era. As multi-modal models continue to develop, the computational power consumed by model training is expected to continue to increase.
2) Reasoning: On January 11, OpenAI announced that it had built 3 million GPTS. Meta, Microsoft, Google, etc. have a massive user base, and the potential penetration of large model applications is high. In the domestic market, iFLYTEK said that as of January 9, the number of Spark Assistant models created had surpassed 51,000. IDC expects 72.6% of inference workloads by 2027.
3) Internet AI capital expenditure is expected to continue to rise. According to Counterpoint, the AI expenditure of most global cloud service providers is expected to account for only 3-7% of total capital expenditure in 2023.
Related companies:
1) Complete server: IFF (601138.SH), etc.
2) Server components: AI chip: Haiguang Information (688041.SH), etc.
3) Optical module: Zhongji Xuchuang (300308.SZ), etc.
4) Data center: Aofei Data (300738.SZ), etc.
Risk warning:
Macroeconomics affects downstream demand, innovation policies fall short of expectations, market competition intensifies, Sino-US games intensify, and related companies' performance falls short of expectations, etc. Companies are not completely comparable, and the relevant materials and data on the target are for reference only.