Research from TrendForce indicates that NVIDIA has decided to halve the SOCAMM capacity on its next-generation Vera Rubin Superchip module. This adjustment does not represent a reduction in NVIDIA's overall memory demand but is a response to the reality that initial supply capacity allocated to NVIDIA for 2027 is insufficient. In this context, NVIDIA has chosen to reduce the capacity per chip and increase the number of modules shipped to strengthen its market share. This move also underscores the trend that the LPDDR5X supply gap is difficult to fill and that medium- to long-term demand is on an upward trajectory.
Although memory manufacturers currently have expansion plans for 2027, the overall increase in bit supply still falls short of the demand projected by the buyer's market. Given that the initial LPDRAM supply plans from Samsung, SK hynix, and Micron can only meet about 60% of the brand's demand, with limited room for upward adjustment, NVIDIA recently decided to modify the SOCAMM module capacity on the Vera Rubin Superchip module. This is to meet the shipment requirements for more Vera CPUs and to proactively address the risk of continued tight supply in the future.
TrendForce emphasizes that this specification change is a supply-side-driven architectural compromise. It has no substantial negative impact on the overall DRAM supply and demand balance, nor does it indicate that NVIDIA is downgrading its total memory demand.
NVIDIA's new-generation AI server products are divided into two categories: the Vera Rubin Rack and the standalone Vera CPU Rack. In addition to the full-rack Vera Rubin Rack solution, NVIDIA has also clearly stated its intention to vigorously promote the standalone Vera CPU solution, aiming to expand its market share in the inference-type AI market. This is also one of the main factors driving LPDRAM demand.
Key Market Driver
With its excellent power efficiency and transfer speed performance, LPDRAM has expanded into diverse application scenarios such as AI servers and automotive. Coupled with the fact that ASIC chip manufacturers other than NVIDIA are also beginning to evaluate its adoption, LPDRAM will no longer be exclusive to smartphones in the future. As AI applications continue to expand, the overall supply-demand gap is expected to become more severe.
TrendForce projects that the overall AI server ecosystem has the potential to become the single largest outlet for global LPDRAM between 2028 and 2030, surpassing smartphone applications.
Comments