Since the unveiling of Chinese artificial intelligence (AI) startup Moonshot AI's large language model (LLM) "Kimi K3," concerns have been raised about a slowdown in AI investment and a drop in memory demand, but an analysis said this will remain a short-term worry. The outlook is that the more efficient AI models become, the more AI adoption will expand, driving bigger investments by Big Tech and higher demand for high-bandwidth memory (HBM).
Kim Dong-Won, head of research at KB Securities, said in a report on the 22nd, "After the recent unveiling of Kimi K3, there were concerns that algorithm optimization could lead to lower AI investment and slower memory demand, but this is expected to prove a short-lived worry similar to when DeepSeek and TurboQuant emerged in the past."
He compared gains in AI efficiency to improvements in car fuel economy. It may seem that higher fuel economy would reduce gasoline consumption, but in reality, lower operating expense increases vehicle use and thus total fuel consumption, he said.
Kim said, "From the perspective of corporations, if AI inference expense falls, they will perform more AI tasks by linking multiple agents," adding, "Even if the memory needed per inference drops to one-third, if usage expands twentyfold, total memory demand increases sevenfold." He added, "Improvements in AI models and semiconductor efficiency are not factors that shrink AI investment; they further stimulate AI demand by lowering service prices and broadening use cases."
KB Securities also expects Big Tech's AI investment race to continue for the time being. Next year, Google and Meta's AI investments are estimated at 270 trillion won and 220 trillion won, respectively, and the two companies' spending is projected to account for 46% of total AI investment by U.S. Big Tech. To secure an early lead in AI infrastructure, Google is expanding external data center lease contracts and pushing development of its next-generation AI server chip "Frozen v2," while Meta plans to secure a total of 14 GW of AI data center infrastructure by 2027.
Kim projected that this expansion of AI investment will translate into earnings stability for SK hynix. He said, "From next year, with a higher share of HBM production, capacity to supply commodity memory will be effectively constrained," adding, "As the share of long-term agreements (LTA) rises and sales to Big Tech and AI data centers expand to as much as 70%, profit volatility will ease and earnings visibility will improve."
KB Securities maintained a "Buy" rating on SK hynix and its target price of 4.2 million won.