The rise of Chinese AI startup DeepSeek, which has demonstrated the ability to deliver high-performance AI technology at a low cost, is pushing South Korean semiconductor giants SK Hynix and Samsung Electronics to rethink their strategies. Both companies have long relied on supplying high-bandwidth memory (HBM) to Nvidia, a major player in the AI market.
DeepSeek’s breakthrough has partially deflated the AI infrastructure investment “bubble” built around Nvidia’s GPU technology. By optimizing software, DeepSeek has shown that advanced AI services can run on lower-performance chips without the need for Nvidia’s increasingly expensive high-spec GPUs.
According to industry sources, major players such as Amazon Web Services, Microsoft, and Meta are expected to invest heavily in AI infrastructure this year, incorporating DeepSeek-R1 at some data centers to reduce costs. The model reportedly uses lower-spec HBM than SK Hynix’s latest HBM3E, which is supplied to Nvidia.
This development presents a short-term challenge for SK Hynix and Samsung Electronics, as reduced demand for high-end GPUs could impact HBM sales. However, analysts see it as an opportunity for the companies to diversify their customer base and reduce dependence on Nvidia.
“Despite concerns about DeepSeek, the rise of cost-efficient AI models will likely lead to an increase in AI applications,” said Cho Yun-ju, a researcher at Mirae Asset Securities. “This will sustain overall demand for AI computing and support the structural trend of rising power consumption.”
Cho added that major tech firms like Microsoft and Meta are not scaling back but rather expanding AI infrastructure investments.
DeepSeek’s success signals a shift from capital-intensive AI technologies, fostering greater ecosystem diversity. “DeepSeek’s emergence symbolizes a startup’s ability to develop independent AI applications without relying on major AI firms,” an IT industry official said. “This could mark the beginning of a paradigm shift.”
The changing landscape also calls for strategic adjustments from semiconductor giants like Samsung and SK Hynix, which have traditionally operated under a mass-production model focused on general-purpose memory.
“Samsung’s memory business largely depends on bulk supplies to major clients like Intel, Nvidia, and AMD,” a semiconductor industry official said. “Going forward, the company needs to adapt to a customized approach with a diverse product lineup for AI services.”
A domestic neural processing unit (NPU) company executive highlighted the significance of DeepSeek’s use of the “Mixture of Experts” (MoE) architecture.
“This technique isn’t new, but DeepSeek has proven that combining it with the latest technologies can deliver advanced AI services at a low cost,” the executive said. “This success will likely spur the rise of numerous AI startups that will forge close partnerships with semiconductor companies for memory and foundry services.”
