[Edaily Reporter Park Soon-yeop] While controversy over the reduction in SOCAMM capacity for NVIDIA’s next-generation artificial intelligence (AI) server platform, “Vera Rubin,” has served as a catalyst for a correction in memory semiconductor stocks, analysts in the securities industry have concluded that it is difficult to interpret this as a sign of slowing demand. Rather, the assessment is that this is closer to an adjustment aimed at increasing CPU and GPU shipments within limited memory supply, and that the trend of expanding demand for High Bandwidth Memory (HBM) remains intact.
According to the financial investment industry on the 5th, recent foreign media reports suggested that the capacity of the LPDDR5X-based SOCAMM modules installed in the Vera Rubin NVL72 server could be reduced. SOCAMM is a modular memory module installed in server CPUs, and looking solely at the reduction in capacity, there is room to interpret this as a decrease in memory demand per server. Following this report, short-term profit-taking pressure emerged on domestic memory semiconductor stocks such as #SamsungElectronics and #SKHynix.

However, Kim Young-geon, an analyst at Mirae Asset Securities, believes this adjustment does not signify a weakening of demand for the Vera CPU. Analyst Kim explained, “Channel checks confirm that low-capacity versions of SOCAMM were already being prepared,” adding, “This is judged to be a measure to reduce the memory capacity per unit of the Vera CPU in order to expand the number of Vera CPUs deployed within limited supply.”
Kim analyzed that an increase in the number of Vera CPUs could actually lead to expanded demand for Rubin GPUs and HBM4. His reasoning is that in a configuration where Vera CPUs and Rubin GPUs are paired, an increase in CPU supply could expand the capacity for GPU shipments; in this scenario, HBM4 demand could grow more significantly than the reduction in SOCAMM capacity.
Researcher Kim analyzed, “Since the Vera-to-Rubin ratio in the Vera-Rubin set is 1:2, if the LPDDR5X capacity of Vera is reduced by 50% to double the number of Vera CPUs, the number of Rubin GPUs would quadruple, and HBM4 demand could also increase fourfold.” However, he added that further confirmation is needed to determine whether this report is limited to the NVL72 or applies to the Vera Rack as well.
Accordingly, Researcher Kim assessed that the negative impact of this issue on the memory sector would be limited. Based on a reverse calculation of the price and total cost of ownership (TCO) mentioned in the report, the price of LPDDR5X for SOCAMM is estimated at $3.6 per Gb, which is significantly higher than the projected 2026 price of $1.5 per Gb for mobile LPDDR5X. This suggests the possibility of further price increases for mobile memory.
The report also noted that it is difficult to interpret this situation negatively based solely on a decline in rack prices or TCO savings. This is because even if TCO decreases by reducing SOCAMM, the utility for users may be limited if performance also declines. Analyst Kim stated, “We view the recent correction phase as a buying opportunity at lower prices,” and maintained a target price of 550,000 won for Samsung Electronics and 3.8 million won for SK Hynix.
Daishin Securities also offered a similar perspective. Ryu Hyung-geun, an analyst at Daishin Securities, estimated that the SOCAMM capacity for the Vera Rubin chip has been reduced from eight 192GB modules to eight 96GB modules, but noted that the overall market size for DRAM for NVIDIA CPUs remains unchanged. It projected that the combined supply of LPDDR5X standalone modules and SOCAMMs from Korean memory companies would increase from 18 billion Gb in 2026 to 29 billion Gb in 2027.
Researcher Ryu explained, “We must focus on the fact that the total addressable market (TAM) has not changed despite the reduction in per-unit capacity,” adding, “This suggests that GPU shipments could exceed market expectations.” In fact, efforts to expand production capacity at the supply chain level to resolve AI bottlenecks are continuing. Daishin Securities estimated that TSMC is pursuing a larger-than-expected expansion of its CoWoS production capacity.
Ultimately, the interpretation gaining traction is that this reduction in SOCAMM capacity is an adjustment aimed at maximizing shipments amid a memory supply shortage, rather than a downgrade in system specifications or a slowdown in AI server demand. As bottlenecks in the expansion of AI infrastructure investment shift from GPUs to memory and packaging, this implies that attention should be focused on the structural improvement trend in the memory market rather than short-term controversies.
Researcher Ryu stated, “It is reasonable to view this decision to reduce the memory capacity as stemming from the severe supply shortage in the memory semiconductor industry,” adding, “It cannot be seen as a change resulting from a downgrade in system specifications, nor is it an issue that warrants excessive concern.” Daishin Securities maintained its recommendation to increase exposure to the semiconductor sector and set a target price of 560,000 won for Samsung Electronics and 3.4 million won for SK Hynix.