China's artificial intelligence (AI) startup DeepSeek has proven that high-performance AI technology can be implemented at a low cost, prompting SK hynix and Samsung Electronics, which have relied significantly on the revenue from high-bandwidth memory (HBM) supply to NVIDIA, to seek new counterstrategies.
Most importantly, the 'DeepSeek Shock' has partially removed the 'bubble' of investment in NVIDIA's graphics processing unit (GPU)-based AI infrastructure. This is because it has been confirmed that advanced AI services can be achieved using low-spec semiconductors through software (SW) optimization without investing in NVIDIA's high-spec GPUs.
According to industry sources on the 6th, this year, Amazon Web Services (AWS), Microsoft (MS), and Meta are expected to implement large-scale investments in AI infrastructure and are projected to utilize DeepSeek's R1 model in some data centers to drive cost efficiency. It is known that the R1 model will be equipped with HBM of lower specifications than HBM3E, which SK hynix currently supplies to NVIDIA.
This situation poses a short-term challenge for SK hynix and Samsung Electronics, which supply high-performance HBM to NVIDIA. As demand for cutting-edge GPUs declines, it is highly likely that revenue from high-spec HBM will also be impacted. On the other hand, there is also the perspective that this could become an opportunity to break away from a structure that heavily relies on NVIDIA and diversify customer bases. To this end, voices are being raised that corporations should establish collaborative systems with various companies and strengthen 'customized' business strategies that meet each company's needs.
Jo Yeon-joo, a researcher at Mirae Asset Securities, noted, "Despite concerns regarding DeepSeek, the emergence of cost-efficient AI models will lead to more AI applications. Therefore, it is expected that overall AI computational demand will actually increase, and the structural trend for increasing power demand is unlikely to change." She added, "So far, the computational efficiency of AI models has continuously improved, but instead of reducing investments, big tech companies like MS and Meta are actually expanding their investments in AI infrastructure."
There are also views that DeepSeek's success could mark a turning point away from the existing AI technology paradigm that depended on large-scale capital and could expand ecosystem diversity and possibilities. An IT industry insider stated, "The emergence of DeepSeek symbolizes that startups can create independent AI applications without relying on gigantic AI companies," adding, "This is the beginning of a paradigm shift."
Samsung Electronics and SK hynix, which have been engaged for a long time in a business model centered on universal memory for low-variety mass production, also find themselves in a situation where strategies tailored to the changing AI paradigm have become necessary. A semiconductor industry insider said, "Currently, Samsung Electronics' memory business is standardized for mass supply of universal memory products to large clients such as Intel, NVIDIA, and AMD. Going forward, they must be able to respond with a customized approach to diversified small-volume supply based on AI services."
The representative of a domestic neural processing unit (NPU) company stated, "Technologies like DeepShik's 'Mixture of Experts' architecture are already commonly used in the market, but proving that it is possible to implement cutting-edge AI services at a low cost by well combining and optimizing it with the latest technology is significant. Many AI startups will emerge led by DeepShik, and these corporations will establish close partnerships with semiconductor companies in areas such as memory and foundry (contract semiconductor manufacturing) to achieve cost efficiency."