Memory prices have skyrocketed unprecedentedly. Is NVIDIA also feeling the pressure? GF Securities: AI servers "downgrade" become a tool to control costs
Guangfa Securities research report points out that due to the surge in memory prices, NVIDIA may adopt "downgrading" measures to control AI server costs. Specific measures include reducing the SOCAMM capacity of the VR200 and Vera series CPUs, and lowering the memory configuration of general servers by about 50%. This move aims to keep the memory cost ratio around 20%, and it is expected that server shipments will see strong quarter-on-quarter growth in Q4 2026
According to Zhitong Finance APP, due to the "unprecedented" surge in memory chip prices continuing to drive up the overall cost of AI servers, Guangfa Securities believes that NVIDIA (NVDA.US) may introduce a series of measures to control costs.
Guangfa Securities analyst Jeff Pu stated in a research report: "We believe the supply chain has taken action and is also working to boost server shipments. Specific measures include: 1) According to a report by SemiAnalysis, the capacity of the SOCAMM (Small Compressed Attached Memory Module) configured with the new generation VR200 CPU has been reduced; 2) The Vera series CPU racks are expected to follow suit, with the SOCAMM for a single Vera CPU reduced to 96GB, meaning the total capacity of LPDDR5X per CPU is 768GB, lower than the 1.5TB specified in the official specifications; 3) General servers are synchronously lowering memory configurations, with RDIMM and/or MRDIMM capacities reduced by about 50%."
The analyst stated: "Therefore, we expect a slight quarter-on-quarter increase in shipments of general servers in the second and third quarters of 2026, but a strong quarter-on-quarter growth of 20-30% in the fourth quarter."
Pu added that reducing the capacity of LPDDR5X (Low Power Double Data Rate 5X memory) is "crucial," as this memory accounts for about 61% of the material costs for memory and storage in the Vera Rubin 200 rack.
Pu estimates that through the aforementioned downgrading operations, NVIDIA is expected to stabilize the proportion of memory costs in the overall cost of AI servers at around 20% before 2027.
According to the latest report from market research firm SigmaIntell, due to a severe imbalance between supply and demand for memory, consumer-grade DRAM prices surged across the board in the second quarter of this year, with LPDDR4X 4GB products rising 75% quarter-on-quarter, and LPDDR5X 12GB products seeing an increase of as much as 89%
