Hong Kong Stock Concept Tracking | DeepSeek training cost is less than one-twentieth of GPT, AI applications may usher in low-cost expansion (with concept stocks)

The AI company DeepSeek has been making headlines across major media and social networks in the United States. Some speculate that this is due to the release of the R1 model on January 20, which can rival OpenAI's O1 model.

According to Tianyancha, DeepSeek is developed by Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co., Ltd., which was established in July 2023 and is jointly held by Ningbo Cheng'en Enterprise Management Consulting Partnership and Liang Wenfeng.

It is understood that Liang Wenfeng, the mastermind behind DeepSeek, is a well-known expert in the financial circle. His founded company, Huanfang Quantitative, is one of the domestic quantitative private equity giants.

Notably, according to news from Zhanjiang, the DeepSeek-V3, which performs on par with GPT-4o, had a research and development cost of $5.58 million, with training costs less than 1/20 of GPT-4o, and only used a cluster of 2048 H100 GPUs, taking just 53 days. At the same level, even leading companies in the global AI first tier would need at least 16,000 GPUs for training.

According to Zhitong Finance APP, DeepSeek's large model achieved performance comparable to giants like OpenAI at a very low cost ($6 million) and with a small number of chips (2000), challenging the industry consensus that "only tech giants can develop cutting-edge AI."

Its success stems from the innovative breakthroughs of Chinese engineers under U.S. chip restrictions: through open-source ecological collaboration, algorithm optimization, and resource integration, transforming black market chips and the computing power accumulation of quantitative trading companies into technological advantages.

This model not only reveals the unintended consequences of U.S. technology blockades but also suggests that China may reshape the global AI competitive landscape through open-source dominance, making low-cost innovation a new path to disrupt the industry.

The market understands that low costs can achieve the best model performance, negating the logic of stacking computing power for large models in the early stages. At the same time, it constitutes a substantial benefit for small and medium-sized AI innovation enterprises.

CITIC Securities released a research report stating that the recent official release of DeepSeek-V3 has garnered widespread attention in the AI industry, significantly improving training efficiency and inference speed while ensuring model capability.

CITIC Securities believes that the release of DeepSeek's new generation model signifies that the application of AI large models will gradually become more inclusive, facilitating the widespread implementation of AI applications; at the same time, the significant improvement in training efficiency will also boost the demand for inference computing power.

With the upgrade of model capabilities, applications such as Agent are expected to accelerate landing, driving companies in fields like office, education, management, edge computing, and embodied intelligence to realize performance elasticity more quickly. CITIC Securities continues to be optimistic about the accelerated landing opportunities in the AI industry chain.

Hong Kong stock companies related to AI applications:

Yidu Technology (02185), Innovation Works (02121), HuiLiang Technology (01860), Fourth Paradigm (06682), FenBi (02469), JingTai Holdings-P (02228), etc