According to the information from Zhitong Finance APP, CITIC Securities published a research report stating that at the model level, whether the technical iteration curve is slowing down still needs to be verified after the release of GPT-5. However, at present, leading manufacturers continue to show a competitive trend in the development of multimodal large models, and the arms race remains intense. At the application level, after about a year of exploration and deployment, it is expected that GenAI's penetration in B-end enterprise operations, C-end video applications, and intelligent terminal aspects will accelerate in the second half of this year. The rapid decline in inference costs is expected to further accelerate the arrival of the application inflection point. For the market, on the one hand, the narrative of upstream training computing power is not over, still representing the track of outstanding growth; on the other hand, downstream needs to pay attention to the changes in multimodal applications and end-side narratives, as well as the overflow of embodied intelligence and autonomous driving after further breakthroughs in models, waiting for the application to switch from thematic investment to prosperous growth.
Model Level: The model's technical iteration curve is difficult to slow down, and the international competition for large models is still intense.
Since OpenAI released GPT-4o in May, some in the industry believe that GPT-4o did not have a substantial breakthrough in reasoning ability compared to GPT-4 a year ago, and therefore some people believe that the technical iteration curve of large models is clearly slowing down. CITIC Securities believes that the focus of GPT-4o's research and development is on human-computer interaction, and one of the reasons for choosing this time to release it may be to pave the way for subsequent collaborations with Apple and Microsoft on the end side. Whether the model's technical iteration curve is slowing down still needs to be verified after the release of GPT-5. Currently, the arms race among model manufacturers remains intense, especially the narrowing gap with OpenAI may encourage the current first/second-tier manufacturers to continue to increase development investment, and the trend of chasing each other is expected to continue. However, with high development costs, unless it is well-funded internet giants or a few star startups with smooth financing, more AI large model companies are unable to support the development of the next generation of models, and the industry landscape may further converge.
Application Level: B-end empowerment, end-side agents, and video applications are showing positive signals.
As there have been no other GenAI &#34;popular&#34; applications on the market besides ChatGPT, and the API call ratio of most models is still very low, some investors have doubts about the actual demand and follow-up momentum of this wave of AI. CITIC Securities believes that there is a certain time lag between the invention of technology and the landing of applications, and defining demand, preparing data, deploying models, and building applications all require a certain amount of time.
Recently, there have been some positive changes in GenAI applications. In the ToB scenario, after about a year of exploration and deployment, more and more internet giants, software companies, and various small enterprises are accelerating the use of AI in their businesses. The most common functions include customer service, information security, digital personal assistants, customer relationship management,
Inventory management, etc. In the ToC scenario, Apple and Microsoft are leading the landing of AI on the edge. CITIC Securities believes that the AI Agent integrated in the edge operating system is one of the most suitable scenarios for GenAI to land in the C-end, and the huge consumer base of the two giants is expected to lead GenAI to break through. In addition, the rapid progress of video modalities also has positive implications for application landing. CITIC Securities believes that video modalities are more in line with the entertainment needs of C-end users, especially with a high degree of matching with short video formats, thus more likely to give rise to high-heat applications, but further maturity of model performance and further exploration of product definition are still needed.
In addition, the rapid decline in inference prices in recent times will further accelerate the downstream penetration of GenAI. According to a report by First Financial on June 7th, &#34;Baidu's main large models are free for half a month, with daily call volume increasing by 10 times.&#34; Since Baidu announced the free use of its two main large models, ERNIE Speed and ERNIE Lite, on May 21st, the daily call volume of these two models has increased by 10 times. Not only domestic manufacturers, according to the OpenAI official website, the API call price of GPT-4o has also decreased by 50% compared to GPT-4-turbo. In addition to manufacturers competing for market share through price wars, large models themselves are also reducing the actual inference costs through a series of new architectures and engineering optimization methods.
Computing power layer: Prospects for supply and demand remain relatively positive.
On the demand side, the Scaling Law in the training phase is still effective, and the proportion of the inference phase is continuously increasing. In addition, from the CAPEX investment of North American cloud computing giants, although the growth rate may have reached its peak, the upward trend is expected to continue, with the incremental investment mainly going to data centers, AI, and other areas. CITIC Securities estimates that the combined growth rate of CAPEX for the four major cloud giants in North America in 2024 is expected to reach 40%, and in 2025, it will continue to maintain high single-digit to double-digit growth. On the supply side, NVIDIA's B-series shipment progress exceeds market expectations, with future product plans for &#34;one iteration per year.&#34;
Under the expectation of loose liquidity, although US tech giants are at high levels, they still have momentum to move forward.
Under the paradigm mapped overseas, concerns about the sustainability of the AI market, led by US tech giants, are somewhat constrained, so it is necessary to answer whether the AI market led by US tech giants can be sustained. CITIC Securities believes that while concerns about the bursting of the current AI &#34;baby bubble&#34; do exist, compared to previous rounds of representative tech market trends, this round of AI market has not yet reached its peak. From an index perspective, as of June 13, 2024, the Nasdaq index has accumulated a 67% increase since 2023, still some distance from the peak level of 100% increase in the post-epidemic online economy
Compared to the peak increase of the 2000 Internet bubble and the mobile Internet wave, the current valuation dimensions of the top companies in terms of overall PE and PB changes are still far from extreme. From a macro perspective, the expected loose liquidity is also expected to provide some support to the US stock market and growth style.
Risk Factors:
Risks include the development of AI core technology falling short of expectations; continuous tightening of policy regulation in the technology sector; global macroeconomic recovery falling short of expectations; macroeconomic fluctuations leading to lower-than-expected IT spending by European and American enterprises; global cloud computing market development falling short of expectations; slower-than-expected expansion of downstream AI application scenarios; risks of enterprise data leakage and information security; and continuous intensification of industry competition.
Investment Advice:
The global AI market has been booming for more than a year. Concerns in the market currently revolve around the slowing down of model iteration speed and the lack of downstream applications. CITIC Securities believes that at the model level, whether the technology iteration curve will slow down still needs to be verified after the release of GPT-5. However, at present, it is expected that leading manufacturers will continue to compete fiercely in the development of multimodal large models, and the arms race will remain intense. At the application level, after about a year of exploration and deployment, CITIC Securities expects that in the second half of this year, GenAI's penetration in B-end enterprise operations, C-end video applications, and intelligent terminals will accelerate.
With continued training demand and the outbreak of inference demand, the computing power industry chain is expected to maintain a high level of prosperity. It is recommended to focus on the computing power layer, such as chips, optical modules, servers, cloud computing, etc., and there are rich investment opportunities to be explored in liquid cooling/power equipment/domestic computing power. On the application side, it is recommended to focus on leading AI application vendors in the fields of office, enterprise management, education, content, and to pay attention to opportunities in the consumer electronics and other smart terminal industry chains

06030

600030

CITIC Securities Co., Ltd. believes that model manufacturers are accelerating the development of multimodal large models, and the reduction in inference costs is expected to accelerate application development. The market needs to pay attention to the changes in multimodal applications and end-side narratives, as well as the spillover effects of models on embodied intelligence and autonomous driving. At the same time, upstream training computing power remains a representative track for outstanding growth. It is expected that in the second half of this year, GenAI's penetration in B-end enterprise operations, C-end video applications, and intelligent terminals will accelerate

Zhitong

CITIC SEC

CITIC Securities Co., Ltd.

CITIC Securities Co., Ltd.: Leading manufacturers are competing to develop multimodal large models, reducing inference costs and accelerating the arrival of application inflection points