President Lee Jae-myung visits Furiosa AI in Gangnam, Seoul, in April and listens to CEO Baek Jun-ho explain about the server./Courtesy of National Assembly Photo Reporters

One of President Lee Jae-myung's key promises, 'securing 50,000 graphic processing units (GPUs),' is drawing increasing calls from academia and industry to utilize resources from domestic artificial intelligence (AI) corporations, including Intel and AMD, to avoid dependency on Nvidia.

On the 16th, Professor Kim Jeong-ho, a global authority on high-bandwidth memory (HBM) at the Korea Advanced Institute of Science and Technology (KAIST), noted in a conversation with ChosunBiz that "if the government's promise of 50,000 GPUs is fulfilled, the total expense will likely be around 5 trillion won for hardware and software combined," and added, "While Nvidia's products are the most proven options, competitors and domestic neural processing unit (NPU) corporations should also join in to strengthen the domestic AI ecosystem through this opportunity."

Professor Kim said, "If President Lee's promise is realized, a balanced purchasing plan is needed, with 50-60% of the 50,000 GPUs being Nvidia's H100 or Blackwell, 30% Intel Gaudi or AMD AI accelerators, and 10% from domestic NPU corporations." This aims to lower overall expenses while escaping dependency on specific companies. He explained that it is necessary to encourage the domestic AI semiconductor industry to foster self-sustainability by using products from local companies like Furiosa AI and Rebellion for inference NPUs.

He stated, "While purchasing GPU hardware and software is essential, AI semiconductors must be designed flexibly by connecting through optical communication," adding, "Only then can researchers connect directly to the AI data center even if they conduct research on the same framework. Otherwise, it will be nothing but a shell." He emphasized the need for proper investment planning for an efficient AI research platform.

Earlier, President Lee prioritized aggressive investments in AI as one of his key promises during the presidential campaign. Securing AI accelerators like GPUs is crucial for implementing various AI-based technologies such as generative AI and robotics. The new government decided to secure more than 50,000 high-performance GPUs to meet this projected demand and realize an 'AI basic society.' They also promised to invest 100 trillion won jointly with corporations and the government over five years to establish necessary data centers for the smooth operation of AI.

The government organization has also begun adjustments accordingly. It has assigned a key coordinating role to the National Artificial Intelligence Committee directly under the president and established the 'Chief of AI Future Planning.'

Importantly, it is essential for the government, industry, and academia to align with a common purpose. With Nvidia's GPU hardware dominating the global AI accelerator market, there are increasing calls for a ratio adjustment among the 50,000 GPUs. Beyond Nvidia, it is crucial for domestic AI semiconductor companies to gain self-sustainability for the growth of Korea's AI ecosystem. In fact, China is strengthening its technological capabilities centered around companies like Huawei amid controls on Nvidia GPU imports.

Even if the plan to secure 50,000 GPUs succeeds, concerns have been raised about the importance of subsequent operations. A university professor, who requested anonymity, stated, "It is clear that the promise of securing 50,000 GPUs is a very important measure for the domestic AI industry and academia," but noted, "The refinement of operations and utilization methods is as crucial as securing GPUs." He pointed out that despite securing a total of 2,000 GPUs in the Gwangju AI data center project promoted by the Yoon Suk-yeol administration, the operating rate dropped below 50% this year due to budget constraints.

Professor Kim Jeong-ho also emphasized the importance of government support for operational infrastructure such as power supply and cooling facilities. Even if GPUs are secured, if the labor costs, electricity fees, or technical support are cut off, high-performance resources become 'useless.' He stated, "After securing GPUs, it is crucial to reflect the opinions of corporations as much as possible on how to operate them stably and sustainably."