"Need to Confirm Whether Investment Will Be Maintained for Semiconductor and AI Performance Improvement"

Daishin Securities analyzed on the 31st that in order to sustain the upward trend in semiconductors, it is necessary to confirm that capital expenditures (CAPEX) for performance improvement are maintained.

Chinese AI company DeepSeek unveiled the DeepSeek-R1 model on the 20th. The DeepSeek-R1 model shows performance similar to OpenAI's latest model o1, and its API fee, which refers to the model usage fee, is only about 3-5%. While Chinese companies' AI technology is evaluated to be about two years behind that of the U.S., the DeepSeek-R1 model caused a significant stir among AI professionals in the U.S.

Seoyoungjae, a researcher at Daishin Securities, explained, "The U.S. government has been continuously strengthening export regulations on AI semiconductors," adding, "We did not expect China to catch up with AI technology at such a rapid pace."

He continued, "The reason they could create a model with similar performance at a low cost was due to reducing supervised learning and focusing on reinforcement learning," adding, "They enhanced COT (Chain of Thought), the reasoning process by which AI models find answers on their own."

He emphasized, "OpenAI claims that DeepSeek used its AI model for training."

Researcher Seo analyzed, "After DeepSeek's announcement, stock prices of semiconductor and power-related companies, including NVIDIA, declined," explaining, "This reflects concerns that if AI models can be trained at low cost, companies may not need to purchase AI accelerators through massive investments." He added, "The stock prices showed high volatility due to the possibility of overinvestment."

However, Seo emphasized, "The DeepSeek announcement does not change the fact that more computing power is needed to create improved models." He added, "To sustain the upward trend in the future, it is necessary to confirm that companies maintain investments to create more powerful AI models, even though efficient methods to develop AI models have emerged."