Zhongyuan Securities: DeepSeek-R1 brings breakthroughs in AI technology, continuously monitoring the launch progress of GPT-5

According to the report from Zhongyuan Securities, due to the technological breakthrough of DeepSeek-R1, China's capabilities in the large model field have gained global recognition, paving a new path for the development of large models. Because DeepSeek-R1 adopts an open-source strategy, it greatly accelerates the adaptation of domestic chips and enhances the release of downstream application capabilities. Although DeepSeek-R1 reduces the demand for computing power, the global demand for computing power will continue to grow rapidly as the inference demands from the application side are released in large quantities. Considering the upcoming release of GPT-5 and the continuous implementation of AI applications, the sector will continue to attract market attention.

In 2024, the growth rate of China's software industry shows a downward trend. The software business revenue in 2024 is expected to reach 13.73 trillion yuan, a year-on-year increase of 10.0%, which is a decline of 3.4 percentage points compared to the 13.4% growth rate in 2023. The total profit of the software business in 2024 is projected to be 1.6953 trillion yuan, a year-on-year increase of 8.7%, which is a decline of 4.9 percentage points compared to the 13.6% growth rate in 2023.

Key data and dynamics to focus on in sub-industries include:

Localization: In 2024, China's dependence on imported integrated circuits is 78% (i.e., the localization ratio is 22%), a decrease of 2 percentage points compared to January-November, indicating a significant downward trend, reflecting the ongoing impact of a series of U.S. sanctions. The rapid increase in ASML's export volume to China suggests that domestic chip production is expected to grow significantly in the future. Huawei's revenue in 2024 has returned to levels before U.S. sanctions, with a noticeable increase in R&D investment.

AI: R1 has extensively used reinforcement learning in the post-training phase to replace the supervised fine-tuning method traditionally used in large models, eliminating the need for large amounts of labeled data, saving computing power while achieving performance comparable to o1. This means that, under the constraints of the Scaling Law, humans have explored new paths for improving efficient architectures and training strategies beyond simply increasing training data and computing power, which may become one of the most important technological breakthroughs driving the emergence of OpenAI's GPT-5.

Computing Power: In Q4 2024, the capital expenditure of the five major U.S. technology companies reached a new high of $73.497 billion, a year-on-year increase of 66%, continuing to rise by 7 percentage points compared to Q3. After the release of DeepSeek-R1, Amazon, Google, and Meta have all provided positive capital investment plans for 2025, while domestic and foreign cloud providers have quickly integrated DeepSeek's models, and domestic chip manufacturers have rapidly achieved adaptation. Zhejiang Province is about to launch a special plan for artificial intelligence, striving to exceed 100 EFlops in computing power scale by 2025, with a clear acceleration trend.

Risk Warning: Uncertainty in the international situation; downstream companies cutting expenses