It has been confirmed that the South Korean government has set a budget of 27.5 billion won per rack for NVIDIA's next-generation AI chip "Rubin," to be introduced next year. This exceeds twice the price estimated by foreign media. Despite soaring AI chip prices in an era of "chipflation," the government is proceeding with securing GPUs (graphics processing units) at a cost.
According to the "Global AI Frontier Lab" project explanation document obtained on the 6th by the office of Democratic Party of Korea lawmaker Lee Jeong-heon, the Ministry of Science and ICT has allocated 3.85 trillion won for next year's GPU server expansion budget. The entire amount will be used to secure 140 Rubin NVL72 server racks. Each NVL72 rack consists of 72 Rubin GPUs and 36 Rubin CPUs (central processing units), totaling 18,000 Rubin GPUs when counted individually. This quantity is separate from the 52,000 advanced GPUs that the Ministry of Science and ICT has stated it aims to secure by 2028. The actual supply volume may vary slightly depending on NVIDIA's supply conditions.
A budget of 27.5 billion won per rack significantly exceeds foreign media estimates. In a June report, U.S. investment bank Bernstein estimated the cost of a Rubin NVL72 rack at $9.1 million (approximately 12.3 billion won), while Morgan Stanley predicted approximately $7.8 million (around 10.5 billion won). The government explained that it factored in potential price increases due to GPU shortages as well as infrastructure costs when setting the budget.
The government plans to deploy Rubin for developing next-generation frontier models such as AGI (Artificial General Intelligence). This move follows growing calls, after Anthropic's release of its super-large AI model "Toss," that South Korea must secure top-tier models. Deputy Prime Minister and Minister of Science and ICT Park Chan-ho previously stated in June that "approximately 10,000 Rubin units are needed to build a Toss-level model." The current import of 140 racks precisely matches that scale.
in TECH
