[Kim Daesik & Kim Hyeyeon's AHA] If AI Does Everything, When Do Humans Grow?

Park Sunghyun, CEO of Rebellion,
Pioneering the AI Semiconductor Era with Customized Inference Chips

Editor's NoteAsia Economy has decided to explore how the rapidly advancing generative artificial intelligence (AI) will transform the field of artistic creation and what concerns 'humans' should have, from the perspectives of engineers and artists. Accordingly, we have prepared a monthly column where Professor Kim Dae-sik of KAIST's Department of Electrical Engineering and Computer Science and choreographer Kim Hye-yeon (CEO of Yeonist) either hold dialogues with artists or discuss works. The title of the column, 'AHA,' stands for 'AI, Human & Art.' Through Professor Kim Dae-sik, who passionately explores the future of generative AI, and choreographer Kim Hye-yeon, who boldly integrates generative AI with dance, we hope you take a step closer to the profound themes of AI, humans, and art.

The competition in the artificial intelligence (AI) industry is evolving beyond the corporate level into a geopolitical rivalry among nations. Individual companies are enhancing their own technological capabilities while closely responding to national and institutional changes. 'Rebellion' is an AI semiconductor design startup navigating the AI field with these two challenges in mind.

Park Sung-hyun, CEO and founder of Rebellion, earned his Ph.D. in Computer Science from the Massachusetts Institute of Technology (MIT) and has accumulated semiconductor design experience at various U.S. companies including Intel, SpaceX, and Morgan Stanley. Returning to Korea in 2020, he founded Rebellion with four co-founders. Over the past four years, Rebellion has rapidly grown by mass-producing and commercializing the AI semiconductor Atom aimed at data centers, as well as securing global investments including from Aramco.

In 2024, Park led the merger with Sapion Korea, formerly a subsidiary of SK Telecom, marking a new chapter in Korea's AI infrastructure history. We met Park on the 13th at BOTBOTBOT, a unique space in Seongsu-dong where AI robots make coffee and people dance.

Park Sung-hyun, CEO of Rebellion, is having a conversation with KAIST Professor Kim Dae-sik and choreographer Kim Hye-yeon at a robot-themed cafe in Seongsu-dong, Seoul, on the 13th. CEO Park of Rebellion, a leading domestic NPU-based AI semiconductor startup, and Professor Kim discussed the overall AI semiconductor industry. Photo by Yoon Dong-ju

-Recently, China's semiconductor industry has rapidly developed, narrowing the gap with the U.S. How do you see China's future influence in the AI semiconductor market?

▲China is also growing rapidly in the AI semiconductor market. Especially after U.S. sanctions, it succeeded in building its own supply chain. Recently, Huawei developed its own AI chips and operates AI services without Nvidia. While China is likely to increase its influence in the AI semiconductor market, catching up with the U.S.'s monopolistic technology and software ecosystem will not be easy. Meanwhile, Korea still lacks competitiveness in system semiconductor design, so in the AI semiconductor era, strategies to overcome these limitations are necessary.

-What do you think is the most necessary change for the Korean semiconductor industry to gain competitiveness in the AI era?

▲The biggest problem in the Korean semiconductor industry is the shortage of design talent. Even in large companies like Samsung Electronics or SK Hynix, talent tends to concentrate on memory business rather than semiconductor design. Ultimately, the incentive structure for nurturing semiconductor design talent must be reformed. Promotion systems and research support policies should be adjusted to create an environment where design personnel can grow long-term. It is time to invest by seeing the future value of AI semiconductors rather than short-term profits.

-There is a forecast that as AI advances, human roles will gradually diminish. What do you think is the role of humans in the AI era, and how should companies prepare for this?

▲AI is already replacing a significant portion of human tasks. Especially, AI maximizes experts' work efficiency by summarizing research papers or finding code errors. However, the problem arises when inexperienced people overly rely on AI, potentially losing real growth opportunities. At the corporate level, AI should be used not merely as a tool for efficiency but as a complement that enables humans to take on more creative and strategic roles. Creating a structure where AI and humans develop together is essential to maintaining competitiveness in the future.

-What kind of education methods or experiences do you think are necessary to grow as global talents in the AI era?

▲In the AI era, it is most important to develop problem-solving skills rather than simply memorizing knowledge. From my experience raising children, if education only provides answers, children do not develop the ability to think and solve problems independently. Ultimately, the key is to experience difficulties and solve them through direct engagement. The same applies to AI research. It is not about problems with predetermined answers but about defining and solving new problems. Therefore, to grow as global talents, it is essential to go beyond simple learning, gain practical experience, and have opportunities to solve diverse problems independently.

"AI already replaces a significant portion of human tasks by summarizing papers or finding code errors, maximizing experts' work efficiency, but the problem is that inexperienced humans overly relying on AI may lose real growth opportunities. AI should be used as a complement enabling humans to take on more creative and strategic roles."

-How is Rebellion navigating the rapidly changing AI industry and AI semiconductor market?

▲Rebellion is an AI semiconductor startup developing inference-specialized semiconductors rather than the existing general-purpose GPU-based AI semiconductors. As AI computations increase, existing GPU-based systems are inefficient in power consumption and cost. Rebellion develops Neural Processing Units (NPUs) optimized for actual application after AI training, rather than handling general AI computations like Nvidia.

Simply put, Nvidia excels at training AI models, while Rebellion creates hardware that efficiently operates trained models. While the existing AI semiconductor market has been GPU-centric, Rebellion is taking a differentiated path through power efficiency, performance optimization, and customized semiconductor design.

Sung-Hyun Park (right), CEO of Rebellion. Photo by Dong-Joo Yoon

-Nvidia currently holds significant influence in the AI semiconductor market. What strategies does Rebellion have to secure competitiveness in this market?

▲Nvidia holds a monopoly in GPU-based AI training. Because AI model training requires enormous computation, GPU's parallel processing architecture is optimized. However, Rebellion employs a different strategy.

We focus on AI inference. Since computation methods differ when AI is used after training, NPU-based semiconductors are much more efficient than GPUs. GPUs are highly versatile but include unnecessary computations. In contrast, Rebellion's semiconductors are designed with computation structures optimized for specific AI models, resulting in significant power efficiency and cost reduction. Although Nvidia currently leads the market, the AI semiconductor market will become more segmented, and Rebellion will pioneer new markets with customized AI inference chips.

-How do AI computations differ from traditional CPUs and GPUs, and what are the characteristics of the MPU and NPU developed by Rebellion?

▲Traditional Central Processing Units (CPUs) use sequential computation, making them unsuitable for AI computations. GPUs enable parallel computation and are used in AI training but remain general-purpose, leading to many unnecessary computations and high power consumption.

Rebellion's NPUs are customized semiconductors tailored to specific AI computation patterns. For example, in AI inference, certain matrix operations are repeatedly used, and Rebellion's NPUs optimize these to achieve faster computation speeds and lower power consumption. In other words, Nvidia's GPUs remain powerful for AI training, but Rebellion's NPU-based AI semiconductors provide a much more economical and efficient solution when AI is applied in real life.

Park Sung-hyun (right), CEO of Rebellion, is having a conversation with Kim Dae-sik (left), a KAIST professor, and choreographer Kim Hye-yeon at a robot-themed cafe in Seongsu-dong, Seoul. Photo by Yoon Dong-ju

-The possibility of Artificial General Intelligence (AGI) is increasingly discussed. How do you expect the AI semiconductor industry to change if AGI becomes a reality?

▲If AGI emerges, AI computation demands will increase exponentially compared to now. Current large language models (LLMs) already require enormous computation, but AGI will multiply that demand several times over. Computation optimization will become even more critical in the AI semiconductor market. Instead of relying solely on Nvidia's GPUs for AI training and operation, specialized semiconductors tailored to AI applications will be essential. Since Rebellion designs such customized semiconductors, we believe there is an opportunity to lead the high-performance, high-efficiency AI inference chip market in the AGI era.

Professor Kim Dae-sik, Department of Electrical Engineering and Computer Science, KAIST · Choreographer Kim Hye-yeon (CEO of Yeonist)