[Column] Development of 'Korean-Style ChatGPT' Faces Hurdles from the Start

GPU contract process takes several months
Without realistic planning, budget waste is inevitable

In February, the Ministry of Science and ICT, the main government body responsible for artificial intelligence (AI), announced its plan to select a "national AI representative" and create a world-class large language model (LLM) on par with ChatGPT, dubbing the initiative the "World Best LLM Selection." The ministry also stated, "As soon as the supplementary budget, which includes funding for the purchase of graphics processing units (GPUs), is passed in the National Assembly, we will immediately begin the process of selecting the national representative." With the supplementary budget bill set to be reviewed by the National Assembly in May, this policy has become the most closely watched issue in the AI industry.

The "World Best LLM" policy, which was expected to proceed smoothly once the budget was secured, has recently encountered obstacles from the National Assembly Budget Office. The office revealed in its analysis of the government’s supplementary budget proposal that the GPU rental period, which is crucial for LLM development, is significantly shorter than what the government had planned.

According to the report, "The government planned to spend 158.6 billion won to allow companies selected as the national AI representative to freely use 2,000 GPUs by finding companies that own GPUs and renting them for 11 months. However, the actual period of use is only about 7 months. Furthermore, there are almost no companies in Korea that possess Nvidia’s latest B200 GPUs, which the government intends to rent."

The core issue is that the actual time the national AI representative company can use the GPUs is only about 7 months, which is 4 months shorter than the government’s stated 11 months, and it remains uncertain whether they will even be able to use the high-performance GPU models. The report explained, "In order for the government to secure and provide GPUs, a series of administrative procedures is required, including selecting an operating agency, finding a supplier with the necessary equipment, signing contracts, and then selecting the user companies. This process alone takes at least 4 months." The report also cited similar cases from the "High-Performance Computing Support Project," where this amount of time was already required.

Even after a nationwide search, there is only one company in Korea that possesses the B200: the ad tech company Filer, which imported the B200 last month. This means that even if the government secures the budget, the product they intend to purchase is not actually available in the market. While there are domestic companies that own the H100 GPU, making it relatively easier to procure, it is expected to take considerable time just to find a supplier for the B200. The government’s plan to develop a "Korean-style ChatGPT" has thus been revealed as an overly optimistic proposal included in the supplementary budget without sufficient prior research.

The government has declared AI a "national core strategic technology" and pledged to provide strong support. However, this analysis raises questions about the feasibility of the project. An unrealistically set rental period and insufficient equipment, without consideration of practical constraints, could ultimately result in wasteful spending. The budget consists of taxpayers’ money. It is only right that it should be spent appropriately where needed. There are already warnings that the domestic AI industry is lagging behind its international counterparts. We cannot afford to waste precious time.