The Korean memory industry is closely monitoring global market trends following the emergence of the Chinese artificial intelligence (AI) model 'DeepSeek R1'.
According to the industry on the 30th, AI semiconductors are specialized in performing simple calculations rapidly and simultaneously. Nvidia's graphics processing unit (GPU) is a prime example. High-bandwidth memory (HBM) made by memory companies such as SK Hynix plays a role in assisting large-scale calculations alongside the GPU. The 4th generation HBM3 is installed in Nvidia's flagship product, the 'H100'. It is known that 'HBM2e', the 3rd generation HBM, is used in Nvidia's lower-spec GPU H800 utilized by DeepSeek.
DeepSeek is designed to deliver the same level of performance with fewer high-performance chips compared to models like OpenAI's ChatGPT. Until now, big tech companies like OpenAI, Meta, and Amazon have been paying astronomical sums to Nvidia to purchase chips to secure AI computing power. DeepSeek has demonstrated that it can reduce hardware dependence through coding optimization.
As a result, the outlook for Nvidia chip demand is expected to be negatively affected in the short term. This is also linked to HBM demand.
SK Hynix is virtually exclusively supplying HBM to Nvidia. SK Hynix's revenue from HBM in the fourth quarter of last year was 5.851 trillion won, accounting for 40% of total DRAM sales. If global AI companies, influenced by DeepSeek's optimization design, turn to reducing chip and memory investment, SK Hynix's sales could also be hit. Samsung Electronics is also conducting quality tests to supply the 5th generation HBM3E to Nvidia.
U.S. containment can also act as a variable. The Donald Trump administration is considering blocking the export of Nvidia's lower-spec chip 'H20' to China. Samsung Electronics HBM is installed in the H20.
On the other hand, there is also a prospect that the DeepSeek shock will act as an opportunity to expand the AI ecosystem and lead to the expansion of the base of domestic companies. Lim Ji-yong, a researcher at NH Investment & Securities, said, "The AI development timetable is accelerating, and small and latecomers are showing new possibilities, which can create additional demand."
Mark Zuckerberg, CEO of Meta, said on the 29th (local time), "It is too early to be sure what DeepSeek means for the future of AI," but added, "I think that investing a lot in capital expenditure and infrastructure will be a strategic advantage over time."
[Copyright (c) Global Economic Times. All Rights Reserved.]
