Source: CITIC SEC Research

On January 20, 2025, DeepSeek-R1 was officially released, performance benchmarked against the official version of OpenAI o1, attracting global attention. As discussions about DeepSeek continue to unfold in the market, concerns have arisen regarding whether the tech giants' enormous computing power in the AI field is "worth it," as well as whether the performance growth logic of AI computing power can be sustained.

We believe: Deepseek's best practices in reducing model training costs are expected to stimulate tech giants to adopt more economical ways to accelerate the exploration and research of cutting-edge models, while also unlocking and implementing a large number of AI applications. The increasing returns to scale from algorithm training and the corresponding Jevons Paradox related to reduced unit computing costs all indicate that in the medium to short term, it is a highly certain event for tech giants to continue investing continuously and at scale in the AI computing power field.

We are Bullish on investment opportunities related to the growth of domestic computing power demand, the continuous expansion of AI+ application scenarios, and the ongoing increase in the popularity of edge AI.

DeepSeek quickly breaks boundaries and promotes the process of AI equality.

On January 20, 2025, DeepSeek-R1 was officially released and open-sourced, attracting global attention due to its low cost + high performance + open-source characteristics. DeepSeek-R1 achieved performance levels comparable to global top closed-source models through innovations at the model level, even under the constraints of limited chips/low computing power investment.

Model Overview: Inference capabilities benchmarked against OpenAI o1, with notable cost-performance advantages. According to the paper officially published by DeepSeek, DeepSeek-R1, as an open-source reinforcement learning large model, demonstrates strong reasoning abilities in areas such as mathematics, programming, and complex problem-solving; overall, the reasoning capabilities of the model are comparable to OpenAI o1. In terms of pricing, the API pricing of R1 presents a magnitude decrease compared to OpenAI o1.

Technical Interpretation: R1-zero takes subtraction to the extreme. The R1 plan adopts two-stage reinforcement learning and is expected to open up a new Scaling Law. We believe that R1's technical solution can be rapidly applied to scenarios requiring high complex reasoning capabilities, such as mathematics, coding, and scientific research, supporting the comprehensive implementation of application forms represented by Agents.

The greatest value of DeepSeek lies in its successful replication of the O1 model for the first time and its open-source release, promoting the process of AI equality. With the rapid breakout of DeepSeek, domestic hardware and application ecosystems are quickly gathering around it, expected to build a large domestic AI software and hardware ecosystem based on DeepSeek.

Table 1: DeepSeek-R1 has a cost-performance advantage compared to OpenAI O1.

DeepSeek is expected to drive the development of the entire AI industry chain.

Through technological breakthroughs and ecological openness, DeepSeek is expected to promote the rapid development of the entire domestic AI industry. Its low cost, high performance, and open-source features are changing the global AI industry chain landscape, driving China from being a "computing power follower" to an "ecosystem leader." CSP providers like Huawei Ascend Cloud, Tencent Cloud, Alibaba Cloud, and Baidu Intelligent Cloud have successively launched DeepSeek's large model on their cloud service platforms. AI computing chip manufacturers like Huawei Ascend, Haiguang Information, and Tensu Zhixun have also adapted to DeepSeek, indicating a positive outlook for DeepSeek’s impact on the domestic AI industry chain.

AI Computing Power: DeepSeek significantly enhances training efficiency, helping to meet the soaring demand for inference computing power.

It is believed that the smaller parameter count of the DeepSeek model compared to the GPT4 model means lower inference costs, and the reduction in inference costs will be a prelude to the popularization of AI applications. The current penetration rate of AI applications remains low, and the industry is still in its early stages. According to calculations from CITIC SEC Research Department's Cloud Infrastructure Group, as Agents become traffic entry points, if penetration reaches around 30%, the number of inference tokens for global Agents will increase by more than three orders of magnitude. Moreover, forms of AI such as video generation and real-time video understanding are also expected to result in significant demand growth. Based on calculations of video generation models, the demand for video inference computing power is expected to surpass the training computing power demand by 2-3 orders of magnitude. The exponential increase in computing power demand coupled with the decline in model training and inference costs will create an industry flywheel, driving the expansion of the entire industry chain. As AI becomes more integrated into daily life and industries, it is believed that the demand for inference computing power will continue to drive growth in the AI computing power industry chain.

Table 2: Calculation of computing power demands for global Agents.

The Computing Industry Chain has many links, is large in scale, and has good growth potential. Computing chips, servers, and network communications are the largest subfields of the Computing Industry Chain. According to the external report from CITIC SEC Research Department, "AI Series Report on Technology Industry - Computing Industry Chain Research Framework 2024" (2024-09-03), it is estimated that by 2025, the performance elasticity of chips and AI servers will be more than four times that of 2023, while optical modules will have more than double the performance elasticity. For domestically produced large models in China, domestic computing power is a relatively stable and reliable option, which can support the transition of these models from research and development to commercialization.

Figure 1: Main Links of the Computing Industry Chain.

AI Applications: DeepSeek's low cost + strong capabilities are driving the comprehensive implementation of AI applications. Under open-source, the ecosystem is expected to continue to grow.

As the cost-performance ratio of the DeepSeek model continues to improve, domestic AI applications are accelerating their implementation in various fields due to a rich ecosystem and mature traffic, while also having a significant driving effect on complex inference scenarios. Among these, the Agent model is expected to become one of the best carriers for all AI applications to land with longer task processes, better scene understanding, and higher autonomy. Furthermore, the open-source nature of the DeepSeek model facilitates the application of different scenarios, and the supporting ecosystem is expected to continuously enrich and grow. It is recommended to focus on scenarios in enterprise management (employee assistant, interviews, marketing), Education (mathematics, competitions), scientific research (drug and material development), law (contract case analysis), medical (longitudinal follow-up of medical records), and finance.

(1) AI + Finance: Embracing AI transformation will be essential.

Compared to the impact brought by the release of ChatGPT two years ago, this time, the innovation of DeepSeek promotes the transition of large models from closed-source to open-source, greatly reducing the cost and threshold for localized deployment. It has become essential for traditional Financial Institutions to embrace AI transformation, and it is expected to achieve results more quickly in reducing costs and increasing efficiency, risk control, customer service, etc., while the innovation of business models still needs observation.

Insurance: Deepseek helps to reduce costs and improve efficiency in the property insurance and health insurance sectors, while the impact on the sales-driven life insurance sector is expected to be relatively limited.

Securities: In the short term, AI is mainly applied in areas such as intelligent customer service, data organization, and code generation, which enhances the operational efficiency of wealth management and investment banking departments in securities companies as well as the middle and back offices. In the long term, the potential for in-depth thinking in the investment and research fields remains worthy of exploration. Several brokerages have successively announced the completion of the localized deployment of the DeepSeek-R1 model, and the next step is to observe the implementation effects of each brokerage in general business scenarios and specialized business scenarios.

Asset Management: Some representative companies have advantageous positions in platform development and the trend towards passive management, expected to benefit from enhanced AI capabilities.

Fintech: The emergence of Deepseek phase proves the feasibility of the continuous development of Chinese AI technology and ongoing cost optimization under the export restrictions on computing power from the USA. We are Bullish on the potential for business model innovation and sustained operational efficiency improvements for leading fintech platforms that already have a foundation in research and technical strength amid this wave of AI.

Figure 2: Traditional financial institutions embracing AI transformation becomes a necessity.

(2) AI + Education: Expected to accelerate into the commercialization phase by 2025.

Recently, companies in the education sector have been embracing Deepseek, and the support of a powerful reasoning AI model is expected to deeply empower complex reasoning scenarios. The agent model implementation in the education sector will become more diverse, while the understanding of application scenarios and data accumulation in related fields along with the level of productization capabilities may become new competitive advantages in commercialization. We are Bullish on the current momentum catalyzed by Deepseek, leading to commercialization in the education + AI landscape across three tracks: educational hardware, software subscriptions, and smart campuses, and we are also optimistic about the efficiency improvements in the traditional education and training sectors.

Figure 3: Three major commercialization markets for Education + AI

(3) AI + Medical: Large models drive advancements in AI diagnostic technology

Medical vertical large models, with stronger understanding, generation, and multimodal capabilities, have expanded the AI + Medical market space in more complex scenarios through two paths: quality improvement and efficiency enhancement. The number of AI medical vertical large models released by domestic companies has exceeded 50, all adopting the same B-to-C strategy as the USA in the application landing phase.

The domestic application of large models to assist diagnosis has two driving factors: 1) the scarcity of primary healthcare resources; 2) AI diagnostic intelligent products can improve diagnostic efficiency in a cost-effective manner. According to CITIC SEC Research Department's Computer & Medical Industry Chain Group calculations, from 2025 to 2029, the cumulative market space for domestic large model intelligent diagnostic products for B-end and G-end is nearly 20 billion yuan, while the theoretical annual market space for the C-end exceeds 70 billion yuan. Companies with underlying models, data barriers, and accumulated product customers are expected to benefit first.

Figure 4: Full-scenario AI diagnostic solution

Figure 5: Main stages of AI medical development and representative diagnostic auxiliary software products

Figure 6: China’s AI Medical Market Size 2019-2028 (including forecast).

Among them, medical imaging AI is currently one of the most important development directions in the medical field. After nearly ten years of development, imaging AI has been widely applied and has entered the commercialization 1.0 era. In November 2024, the National Healthcare Security Administration will list AI-assisted diagnosis as a separate expansion item. It is believed that the commercialization of imaging AI will continue to receive strong support. In early 2025, the DeepSeek series of models, with their advantages of high performance, low cost, and open-source nature, will become a phenomenon-level large language model (LLM). It is expected that LLM technology will continue to migrate to the medical imaging field, assisting platform enterprises with medical imaging CNI Data Factor Index to develop high-performance base models. These base models are expected to bring profound changes to the industry in areas such as model training and commercialization deployment, and imaging AI may enter the 2.0 era.

Figure 7: Schematic diagram of the medical imaging base model.

Edge AI: DeepSeek is a breakthrough in the domestic AI ecosystem that is expected to accelerate the implementation of edge AI.

DeepSeek will empower AI software, AI hardware, and other application manufacturers, accelerating the rapid development of AI applications at home and abroad and promoting the swift landing of edge AI. At the same time, edge AI in the Internet of Things, as one of the important carriers of AI applications, is expected to usher in accelerated implementation under the trend of model miniaturization and open-source brought about by DeepSeek. Counterpoint predicts that in the future, edge AI will drive significant growth in cellular module shipments, with a compound annual growth rate of 73% between 2023 and 2027.

Against the backdrop of rapid improvements in large model understanding and interaction capabilities, along with the rapid decline in external API calling costs, it is believed that AI will blossom in various forms. AI smartphones remain a core direction, while it is suggested to focus on opportunities in the AIoT (glasses, Smart Home), AI PC, and robotics industry chains.

Figure 8: The application of large models from cloud to edge, strengthening the dominance of hardware manufacturers.

Figure 9: Forecast of the rhythm of AI landing in mobile phones, PCs, and AIoT.

Investment Suggestions

The release of DeepSeek's new generation model means that the application of AI large models will gradually become popular, accelerating the comprehensive landing of AI applications; at the same time, it is expected to open up a new Scaling Law, with the model focus gradually shifting from pre-training to reinforcement learning and reasoning stages, supporting the continuous growth of computing power demand. It is recommended to pay attention to three sub-themes:

① The domestic computing power industry construction is about to enter an explosive period, focus on related computing power chips, supporting copper interconnection, and AIDC sectors;

② The application scenarios of AI+ will continue to expand, focus on related AI+ office software, AI+ industry, and data service sectors;

The popularity of edge AI will continue to increase, focusing on related AI mobile phones, AIoT supporting industry chains, and Brain-computer Interfaces among other Sectors.

In the US stock market related to Technology, considering the unexpected progress of AI Agents and the latest algorithm of Deepseek disrupting the global market, the CAPEX expenditures of North American cloud computing giants and their detailed structure have also become one of the current focuses of Capital Markets. Analyzing factors such as comprehensive demand and investment direction, it is believed that the Technology computing power industry chain in the US stock market will continue to maintain a high level of prosperity in 2025.

Figure 10: Sorting out AI investment opportunities.

Risk Factors

DeepSeek's development is below expectations; risks of AI computing power demand below expectations; risks of edge AI demand below expectations; serious social impacts caused by misuse of AI; AI application product expansion below expectations; potential ethical, moral, and user privacy risks of AI. Risks related to enterprise data security; information security risks; increased competition in the Industry; US sanctions on China's Semiconductor industry exceed expectations; global supply chain disruptions; geopolitical risks; ongoing tightening of policies and regulations in the Technology field; risks of policy regulations related to private data; macroeconomic recovery globally below expectations; macroeconomic fluctuations causing lower than expected IT expenditures (especially AI expenditures) by European and American companies; risks of domestic new Infrastructure falling short of expectations and digital economy policies not being implemented as expected; traffic growth below expectations; cloud vendors and operators' capital expenditures below expectations; risks of technology solution iterations. Risks in the computing power chip supply chain; risks of insufficient chip production capacity; capital expenditures of major Internet firms below expectations; slow progress and insufficient strength in the introduction of related industrial policies; chip technology iteration not meeting expectations; insufficient progress in the mass production of domestic advanced manufacturing processes. Automotive companies' self-research chip progress below expectations; self-developed chip performance below expectations; lower than expected penetration rates for advanced intelligent driving; risks in the decline of automobile industry sales; downstream demand for Consumer Electronics below expectations; technology iterations below expectations; lower than expected demands in the Robotics market; lower than expected progress in mass production of AR/VR products; increased competition risks in the financial information service industry; licensed operation risks of securities trading information; data security risks; risks of AI applications landing below expectations; compliance risks brought by large models; insufficient landing capabilities of traditional companies in AI; failure to anticipate potential disruptive innovations; overvaluation of some fintech companies; customer expansion below expectations; policy changes in the AI-assisted medical diagnosis industry risks; further intensification of market competition risks; macroeconomic volatility risks; reputational risks in the Industry and companies; ethical risks in medical imaging data; risks of failure in medical instruments registration; risks of restricted downstream medical insurance payments; policy change risks in the Education sector.

Editor/rice