[Feature Story] The Rise of DeepSeek: The AI War Begins

   The generative AI model released by the Chinese startup “DeepSeek” in January has caused a major disruption in the AI ecosystem. As a result, Nvidia stock prices dropped sharply, and other stocks that were benefiting from the AI industry also experienced a significant decline.

   It was once believed that generative AI required large-scale manpower and costs reaching trillions of won, with training done using high-performance GPUs (graphics processing units). According to the Stanford 2024 AI Index Report, Google’s Gemini Ultra cost around 2.7 trillion won, and OpenAI’s GPT-4 cost about 1.1 trillion won. However, DeepSeek claims it only spent about 8 billion won, primarily by using low-spec GPUs. If true, this would mean DeepSeek developed an AI model at a significantly lower cost than OpenAI. While some argue this claim may be exaggerated, factors like Chinese government support and existing cloud infrastructure could have played a role.

   One of the key techniques behind DeepSeek’s cost efficiency is knowledge distillation. This method involves training a smaller “student” model to mimic the behavior of a larger, pre-trained “teacher” model. Rather than training the student on raw data, it learns from the teacher’s outputs, allowing the smaller model to achieve similar performance with far less computational power. This approach reduces the reliance on expensive, high-performance GPUs, making AI development more affordable. In addition, it enhances the model’s ability to generalize to new data, improving overall efficiency.

   Through this innovation, DeepSeek has shown that world-class AI models can be developed without high-performance GPUs. As a result, the competition in AI development may shift from hardware to software, with the focus moving from scale to creativity and optimization.

   In Silicon Valley, data centers adopting DeepSeek’s AI model are emerging, reshaping the AI market, which once seemed monopolistic. Amazon Web Services, following companies like Perplexity and Microsoft, has announced adopting DeepSeek’s AI model in its services. This allows companies to use the latest generative AI models more effectively, enjoying both cost savings and improved performance. In addition, there is an expectation that “distillation” technology will break down the barriers of AI monopolies. Through this, AI startups may quickly mimic the technologies of leading companies, causing the effectiveness of existing closed strategies to diminish. However, with closed-model companies having raised significant investments, predicting how future AI competition will unfold remains uncertain. Ultimately, the key to survival will depend on the end users and consumers. While the development of AI made the U.S. AI dominance in AI seem secure, China’s open-source strategy could reverse the market dynamics.

   The adoption of AI will fundamentally change our lives, as well as existing knowledge and frameworks. In the case of medicine, rapid research will allow many previously incurable diseases to be treated, and the pace of scientific research will accelerate significantly. Such achievements are already reflected in the Nobel Prize, and AI will continue to assist and improve our lives. However, there are concerns about the use of personal data and the potential monopolization of information by a few individuals, institutions, or countries. DeepSeek was accused of passing domestic user data to a third-party server during a meeting of the National Assembly’s Strategy and Finance Committee on the 19th. Similar concerns have been raised with ChatGPT. Some companies are strengthening security through their own servers, but this is only a temporary solution and may not be a sufficient one. In a situation where AI could become a weapon, countries must develop their own AI systems and create the infrastructure to utilize them. While the government is accelerating the development of computing centers, companies argue that this pace must be even faster. In addition, there is a need to train and cultivate AI experts to gain an advantage in the ongoing AI competition.

   When the iPhone was first released in South Korea in 2007, we could not have imagined living our lives carrying small devices. The introduction of smartphones led to a major shift in our lives. While there are concerns that AI is a bubble industry, like the impact smartphones had before and after, AI is likely to bring a rapid change to our lives. As a result, what we once took for granted will no longer be so, and the paradigm of life will change. While it is difficult to predict this change, both governments and companies must quickly adapt to these shifts and build global competitiveness.