Demand for advanced graphics processing units (GPUs) is rising sharply at universities, public research institutes and industry-academia cooperation foundations amid the spread of artificial intelligence (AI), but a growing number of organizations are failing to secure GPUs because of market shortages and surging prices. Research communities warn that if the situation continues, declines in research quality and efficiency will be unavoidable.
According to an analysis by The Seoul Economic Daily of tender listings on the Public Procurement Service's Nara Marketplace, 142 "GPU purchase" service contracts were registered between January 1 and October 27 this year, with a combined budget of 24.64 billion won ($18.2 million).
Most of the ordering organizations were universities, public research institutes and industry-academia cooperation foundations that needed to build high-performance GPU servers for research projects in areas such as quantum computing, AI model development and machine learning.
That level of demand is three times the average of the past three years. Between 2023 and 2025, the number of registered service contracts stood at just 40 per year (3.39 billion won), 39 (3.33 billion won) and 50 (7.79 billion won), respectively.
The problem is that finding contract partners has become very difficult as AI demand has surged across academia and industry in recent months. An official at a public institution in charge of developing security-related AI models said, "These days, companies that have already secured GPUs don't readily bid for Public Procurement Service contracts. Prices are rising by the day, just like 'crypto,' so they wait unless they get the most favorable deal. For those without existing business relationships, the atmosphere is such that you can only secure a private contract with one supplier after several rounds of failed bidding."
It is not uncommon for requests for proposals to be revised multiple times as GPU prices rise during repeated failed biddings and re-tendering. According to SemiAnalysis, a research firm specializing in semiconductors and AI, the one-year rental contract price for Nvidia's H100 GPU has soared about 40%, from $1.70 per hour as of October last year to $2.35 last month.
Student A, a doctoral candidate at a graduate school in Seoul who placed a GPU purchase order with the Public Procurement Service this year, said, "GPU and RAM prices are rising by 20 to 30 percent every month. While going through the complex bidding paperwork and review process, research gets delayed and costs go up by several million won. In the end, to stay within the pre-set budget, we have no choice but to downgrade GPU performance, which leads directly to lower research quality and efficiency."
The government is aware of the problem and is running a GPU leasing program for industry, academia and research institutes. Last month, the Ministry of Science and ICT (MSIT) announced that it had selected recipients for support covering 4,000 GPUs and would begin supply in March. However, critics point out that the program suffers from time gaps between December and March each year because of budget execution cycles, and that rental periods are limited to a few months. As a result, securing GPUs independently remains unavoidable for continuous research.
