"Physical AI and on-device AI are no longer optional, but essential technologies, and the development of semiconductors for these technologies will determine the competitiveness of the entire industry."

At the '2025 e4ds Tech Day' event held on September 9, Hongseop Choi, CEO of Mind AI, presented on 'Physical AI Era, On-Device AI Technology and Strategy', and gave an in-depth presentation on the definition of physical AI, market outlook, and the importance of on-device AI semiconductors.

CEO Choi Hong-seop first explained the concept of physical AI, defining it as “AI that acts in an actual physical environment, such as Tesla’s Optimus or Figure AI’s humanoid robots.”

While existing robotics relied on sophisticated programming and hardware control, physical AI is characterized by autonomous decision-making through data-based learning.

These technologies are not limited to humanoids.

Agriculture, defense, It can be utilized in various industrial fields such as construction, and in fact, Mind AI developed the first commercialized AI agricultural machine in Korea and signed a contract for mass production of 100 units.

The machine is a 'speed sprayer' that sprays pesticides in orchards and can be controlled in real time through on-device AI.

The core of physical AI is the 'Vision-Language-Action Model (VLA)'.

This is a form that integrates visual recognition and behavior control functions into the existing LLM (Large Language Model), and a single model performs the entire process of seeing, understanding, and acting.

“This model is not simply a combination of three functions, but operates as a single, integrated AI model,” CEO Choi emphasized.

On the other hand, commercializing these technologies requires overcoming the limitations of cloud-based AI.

Especially in areas with unstable communication environments, such as agriculture or defense, a loss of cloud connection can cause robot control to be interrupted, which can be a critical problem.

Accordingly, CEO Choi argues that on-device AI, that is, a method of performing AI calculations within the device, is essential.

High-performance AI semiconductors are required to implement on-device AI.

CEO Choi explained, “Currently, the computational performance of edge AI chips is typically around 10 to 20 TOPS, but to implement actual physical AI, performance of over 100 TOPS is required.”

Mind AI is preparing for mass production using Qualcomm's 100 TOPS chip, which is the choice for running LLM-level models directly on the device. />
Collecting data for physical AI is also a major challenge.

To mimic human behavior, we wear full-body sensor suits and collect data by having robots follow these actions. We are also actively utilizing digital twin-based simulation environments to reflect various environments and scenarios.

Mind AI calls this the ‘Digital Training Ground.’

CEO Choi predicted, “Physical AI could become the largest market in human history beyond a simple technology trend,” adding, “It has the potential to grow into an industry worth trillions of won, surpassing the automobile and smartphone markets.”

In particular, humanoid and autonomous driving technologies are considered key fields, and it is expected that robots with built-in AI will be deployed in all industrial sites in the future.

Finally, he advised semiconductor companies to “actively develop on-device AI semiconductors to prepare for the era of physical AI.”

We are currently collaborating with domestic and international semiconductor startups, sharing various technical requirements such as computational performance, power efficiency, and heat management.

Representative domestic companies such as DeepX, Mobilint, FuriosaAI, and Rebellions are challenging this field.