Lee Yeon-woo is a financial journalist at The Korea Times. Her wide range of reporting includes policies, macroeconomics, stock market, companies and even crypto. She is passionate about connecting the dots in Korean finance and making it easier for foreign nationals to understand. Based on her previous experience as a national reporter, she also has a keen interest in social issues within the sector, including gender equality and ESG. Your tips and insights are always appreciated. You can send them to yanu@koreatimes.co.kr.
Samsung Electronics, SK hynix to benefit further from AI memory bottleneck
Nvidia CEO Jensen Huang, center, leaves the SK hynix booth during the annual COMPUTEX expo in Taipei, Taiwan, June 2. Reuters-Yonhap
By Lee Yeon-woo
Sanjeev Rana, head of Korea research at CLSA
Memory chips remain the key bottleneck in the artificial intelligence (AI) supply chain, with Samsung Electronics and SK hynix still positioned as strong beneficiaries of the cycle, Sanjeev Rana, head of Korea research at CLSA, said Tuesday.
During a press briefing on the sidelines of the CITIC CLSA Northeast Asia Forum held at Conrad Seoul, Rana noted that over the past 10 years, memory chips’ contribution to semiconductor industry revenue has jumped from 28 percent to 52 percent.
“It’s all driven by global AI infrastructure spending. Companies are looking to expand capacity across foundry, advanced packaging, production equipment, interconnects and other areas,” Rana said. “But these bottlenecks are still likely to persist for the next several years and are unlikely to go away anytime soon.”
According to CLSA, global semiconductor industry revenue is expected to reach $2.5 trillion by 2030, up 80 percent year-on-year, with the majority — approximately $1.4 trillion — coming from the memory chip sector.
Rana said that as AI models grow, limitations on memory bandwidth increasingly constrain overall system performance and scalability. In AI data centers, 60 percent of energy is spent on moving data across chips and only 40 percent on compute. While the industry has pursued multiple mitigation strategies, including faster memory, larger caches and improved data-movement architectures, the memory wall remains a persistent and costly challenge, he noted.
The emergence of agentic AI, though still in its early stages, is also raising the need for CPUs, which consume a large amount of server DRAM.
“In iPhones, Galaxy phones or PCs, companies in the past could lower memory content, and demand and supply would eventually come into balance,” Rana said. “But if we try to lower the memory content in AI systems, it is going to have a very negative impact on the consumer experience and performance. So that kind of trade-off just isn’t there.”
This outlook has prompted hyperscalers to sign multiyear supply contracts with Samsung Electronics and SK hynix. Rana forecast that more than 50 percent of their capacity will be tied to long-term agreements, providing better visibility for the companies’ performance. CLSA has a target price of 540,000 won ($351.84) for Samsung Electronics and 3.7 million won for SK hynix.
“The valuation multiple rerating story for Korean memory stocks is underway, and it will continue for some time,” Rana said.
Regarding the rise of Chinese memory chip makers, such as CXMT in DRAM and YMTC in NAND, Rana said he does not expect them to compete directly with Samsung and SK hynix in the near term, as the performance and power efficiency of their chips remain significantly lower.
“But it is also a fact that Samsung and SK hynix can meet only 60 percent of customer demand. For the remaining 40 percent, Chinese memory chip makers do have opportunities,” he said, noting that Apple CEO Tim Cook has hinted at easing restrictions on Chinese suppliers.