According to the science and technology community on the 21st, as the scale of computation required for AI research grows rapidly, securing GPUs is emerging as the key determinant of whether AI model research can continue. Individual labs are producing a range of results with government support, but project periods are short, making it difficult to carry the work into follow-up research that retrains models or upgrades performance.
Through last year's supplementary budget and other measures, the government decided to secure 13,136 advanced GPUs, including Nvidia's B200 and H200, through public-private cooperation. The GPUs secured are being supplied to companies, universities and research institutes through multiple programs, including sovereign and specialized AI foundation models, "AI for All," support for advanced GPU utilization, and computing support for AI research. As a result, university labs and government-funded research institutes can now use dozens to hundreds of the latest GPUs — resources they would struggle to obtain on their own — with government support. K-Fold is a flagship result of such government GPU support.
The problem comes after the project period ends. Researchers selected for a program are allocated GPUs, use them for several months, and once support ends must apply to another program all over again to continue their work. AI models must be retrained and improved as new data accumulates, but when GPU support is cut off, follow-up research requiring large-scale computation also grinds to a halt.
Other research sites face the same situation. A research team at Seoul National University's College of Engineering developing next-generation omnimodal AI secured GPUs through government initiatives such as the "high-performance computing support program," but with the research period over, it is now looking for another national program. The research is a large-scale project requiring around 100 GPUs at a time, but it is not easy for a lab to secure computing resources on that scale on its own. "Small-scale research can be done with one or two GPUs, but for research that needs about 100, securing the resources all at once is not easy," the researchers said.
Procuring the missing GPUs from private cloud providers at the lab's own expense is also difficult. Rental rates for GPUs are rising as demand for AI model training and inference surges. For large-scale research that requires dozens to hundreds of GPUs simultaneously over long periods, covering the cost from research funding alone is virtually impossible.
Yet having the state provide hundreds of GPUs indefinitely to a team once it has been selected is not a realistic solution either. The computing resources the state has secured are limited, and new research teams must also be given opportunities.
For these reasons, some universities are expanding their own infrastructure. KAIST recently decided to build an AI data center (AIDC) at its main campus in Daejeon, aiming to meet surging on-campus demand for AI research and expand its own computing base that researchers can use at all times. KAIST plans to eventually expand the AIDC to as many as three buildings. Only the design of the first building has been finalized so far, but the plan is to scale up the infrastructure in stages so that two more buildings can be added as AI research demand grows. Sungkyunkwan University currently operates 40 Nvidia A100s as shared campus GPUs and plans to add 24 B300s this year with 3.5 billion to 4 billion won of its own budget. "Even after external GPU support periods end, on-campus researchers can continue to use the shared GPUs at about 30% of the cost of commercial cloud services," the university said.
Still, university-owned infrastructure alone has limits in supporting all large-scale AI research requiring dozens to hundreds of GPUs. Having every university permanently maintain hundreds of the latest GPUs, along with the servers and high-speed networks to connect them and the power and cooling facilities they require, is a heavy burden in terms of both cost and utilization efficiency. Some researchers are calling for a path to computing resources for follow-up research so that projects that have already produced results are not suspended while waiting for the next government program. An AI development researcher at a major domestic university said: "It is true that the government is working hard to secure GPUs. But supporting corporate commercialization and supporting academic research that creates new models and technologies need to be run on two tracks, with different timelines and budgets."
