[Reporter’s Notebook] For Korean-Style AI to Move Beyond the Nvidia-Centric Paradigm

"There is talk about promoting 'Sovereign AI,' but all we are hearing about right now is the supply of Nvidia's graphic processing units (GPUs)."

This was the recent comment made to a reporter by the head of a venture capital (VC) firm that has been consistently investing in major domestic AI semiconductor companies. While he welcomed the emphasis on AI sovereignty, he pointed out, "What we need first is a blueprint to increase the utilization of AI hardware led by domestic companies."

The Lee Jaemyung administration has announced plans to establish a 100 trillion won fund through public-private cooperation to build Sovereign AI. Sovereign AI refers to securing national capabilities and leadership throughout the AI industry ecosystem, including generative AI, semiconductors, and power and data infrastructure. As a result, the government is expected to pursue projects such as building a national AI data center, securing 50,000 GPUs, and developing domestically produced AI semiconductors and large language models (LLMs).

The challenge lies in expanding the reference cases for using domestic hardware. So far, OpenAI's ChatGPT has dominated the global market based on Nvidia GPUs, while Google's Gemini has been trained on its own proprietary hardware. In this situation, if government policy focuses solely on securing GPUs, the possibility of self-reliance for Korean AI could actually decrease. Leading domestic AI hardware companies such as Rebellions, FuriosaAI, and DeepX are developing neural processing units (NPUs). NPUs are semiconductors specialized for AI inference tasks, and their strengths include lower power consumption and lower cost compared to GPUs.

For an industry to grow from the hardware stage, government-level technology verification and securing use cases are essential. FuriosaAI, for example, failed to secure sufficient domestic demand for its NPU technology and consequently became an acquisition target for Meta in the United States. Although the acquisition ultimately fell through, to prevent similar situations and establish industrial security, support measures are needed to allow domestic companies to grow stably.

The private sector has already demonstrated the potential. SK Telecom is piloting the use of Rebellions' AI semiconductor 'Atom' in its own services, while AI service company Upstage is working on an on-premise AI system that combines FuriosaAI's next-generation NPU 'Renegade' with its LLM 'Solar.' In March, Rebellions CEO Park Sunghyun commented on social media regarding the supplementary budget discussions, saying, "It is very disappointing that the conversation is only about securing GPUs," and emphasized, "Korea is always lacking in scale and resources are absurdly insufficient in the AI arena, but as 'One Team Korea,' we must find a way to survive no matter what."

In 1996, Korea became the first in the world to commercialize code division multiple access (CDMA) technology, allowing it to lead the mobile communications market for decades. This was possible because Korea chose to focus and innovate beyond the standard technologies dominated by Europe and Nokia at the time. For Korean-style AI as well, true AI sovereignty can only be achieved by starting with a philosophy and strategy to design a self-reliant ecosystem based on domestic technology.