SEOUL, Feb 17 (Reuters) - South Korea announced plans to secure 10,000 high-performance graphics processing units (GPUs) this year to stay competitive in the escalating global AI race. "As competition in the AI industry heats up, the rivalry is shifting from corporate battles to full-scale competition between national innovation ecosystems," said South Korea's acting President Choi Sang-mok.

The government aims to acquire these GPUs through public-private partnerships to kick-start services at its national AI computing center sooner. Last month, the U.S. introduced new regulations to control the flow of American-made AI chips and technology, essential for the most advanced AI applications. These regulations limit the export of GPUs, which were originally developed to enhance graphics rendering.

The GPU requirements for AI models vary based on the GPU's performance, the data volume used for training, the size of the model, and the training time desired. The new U.S. rule classifies countries into tiers, with South Korea among about 18 nations exempt from restrictions. However, 120 countries will face export limits, while nations like Iran, China, and Russia are completely banned from receiving them.

Although the South Korean government hasn't decided on the specific GPUs to purchase, it plans to finalize details, including budget, GPU models, and involved private companies, by September. Nvidia, the leading U.S. chip designer, holds around 80% of the global GPU market, surpassing competitors like Intel and AMD.

Meanwhile, Microsoft-backed OpenAI is working on reducing its dependence on Nvidia by developing its own AI chips. OpenAI's popular ChatGPT uses thousands of GPUs for training and improvements. The company is preparing to finalize the design of its in-house chip, which will be manufactured by Taiwan Semiconductor Manufacturing Co.

The AI race has also been impacted by the rise of Chinese startup DeepSeek, which uses AI models focused on optimizing computational efficiency instead of sheer processing power. This approach could help narrow the gap between Chinese-made AI processors and their more powerful U.S. counterparts.

Disclaimer: This image is taken from Reuters.