“On-device AI semiconductors are better suited to specific use cases than general-purpose designs. If we focus our capabilities on developing high-performance, high-memory, low-power AI chips within the next five years, Korea can also take the lead in the physical AI market.

At the '2025 Artificial Intelligence Semiconductor Future Technology Conference' held at the Lotte Hotel on December 10, Choi Hong-seop, CEO of Mind AI, gave a presentation titled 'The Core of Physical AI is the Brain: Take the Lead in the On-Device AI Semiconductor Market.'

In his presentation that day, CEO Choi Hong-seop focused on the importance of 'physical AI,' a new paradigm in artificial intelligence (AI) technology, and the on-device AI semiconductors that will support it.

CEO Choi emphasized, “Unlike existing robotics or autonomous driving technologies, the key to physical AI is its evolution through data-driven learning.”

While existing robots rely on rule-based programming, physical AI learns from vast amounts of data to make decisions and take action on its own.

Representative examples include Tesla's FSD (Full Self-Driving) and Optimus robots, as well as their potential use in various industrial fields such as agriculture, defense, and construction.

He predicted, “While the U.S. and China are ahead in some areas, such as autonomous driving, our country can also be sufficiently competitive in areas such as agriculture and manufacturing,” and “The physical AI market has enormous growth potential that surpasses the existing automobile or smartphone markets.”

The role of AI semiconductors is crucial for the full-scale commercialization of physical AI.

CEO Choi Hong-seop explained, “Physical AI requires AI models to be run directly on actual devices (on-device), rather than through cloud-based computation.” This is due to various practical constraints such as real-time, communication dead zone issues, power and heat generation.

In particular, the model applied to physical AI is 'VLA (Vision-Language-Action)', which is an advanced form of the existing LLM (Large Language Model).

VLA is an advanced AI that understands language and visual information simultaneously and even performs actual robot control.

Incorporating such complex models into small semiconductors presents various technical challenges, including computational performance, memory, power efficiency, and heat management.

Based on his experience applying AI semiconductors to actual agricultural machinery, CEO Choi emphasized, “The computational performance (TOPS) figures disclosed by manufacturers may differ from the actual AI model operation performance,” and “It is essential to design considering the LLM’s unique repetitive computation structure, such as the memory structure and key-value cache.”

He also added, “Since humanoid robots have limited battery and cooling space, low-power and low-heat design is of utmost importance.”

In fact, in the global market, various companies such as Qualcomm, Nvidia, and Apple are jumping into the on-device AI semiconductor competition, and China's Price Robotics is also rapidly catching up in technology.

CEO Choi Hong-seop advised, “Our country’s AI semiconductor companies are also approaching the global level,” and “We need to clearly define our goals through practical comparisons with products from competitors in countries like China and the United States, and target large-scale markets such as physical AI humanoids.”