Nvidia begins shipping Vera Rubin GPUs while memory costs surge as Korean chipmakers scale output, driving component inflation.
Memory semiconductors account for approximately 62% of the cost of Nvidia's Vera Rubin, the next-generation artificial intelligence platform now shipping in volume. According to a bill of materials analysis from UBS obtained by ChosunBiz, the Vera Rubin superchip—which bundles one Vera central processing unit and two Rubin graphics processing units—carries a total cost of $38,902. Of this, memory components account for $24,297: sixth-generation high-bandwidth memory (HBM4) and SOCAMM2 represent the largest expense.
The cost breakdown reveals a striking imbalance between GPU and CPU memory. HBM4 mounted on each Rubin GPU costs $4,943, representing 12.7% of total system cost. In contrast, SOCAMM2 connected to the Vera CPU accounts for $19,355, or 49.8% of the total—making CPU memory the dominant cost driver. The Rubin GPU's full cost, including HBM4, the interposer, advanced packaging, and peripherals, was estimated at $9,247. The Vera CPU with SOCAMM2 cost $20,059, with other board components adding $350.
Market reception to Vera Rubin has been decidedly positive. Nvidia announced last month that Vera Rubin NVL72 systems are already running across Google Cloud, Microsoft Azure, Oracle Cloud, CoreWeave, and other major cloud platforms. SpaceX CEO Elon Musk called Vera Rubin "the best architecture" during an earnings announcement on August 4th and outlined plans to build AI infrastructure centered on Nvidia GPUs. SpaceX aims to secure more than 2 gigawatts of computing infrastructure by year-end and increase capacity to nearly 10 GW by the end of next year. Such deployment plans from major customers signal sustained demand and robust sales momentum for Vera Rubin following Blackwell.
The memory cost share has grown significantly compared to the previous generation. Memory represented 53% of the cost of Blackwell Ultra's GB300, a figure that rose 9 percentage points with Vera Rubin. While Vera Rubin's total cost increased approximately 2.1 times from its predecessor to $38,902, memory expense climbed 2.5 times to $24,297—a steeper trajectory driven primarily by SOCAMM2 adoption.
SOCAMM2 represents a novel approach to server memory. The technology modularizes low-power DRAM (LPDDR5X), traditionally used in smartphones and laptops, for AI servers. Multiple DRAM modules are bundled and positioned close to the CPU in a replaceable configuration. While it offers lower bandwidth than HBM, SOCAMM2 delivers large capacity with reduced power consumption and superior bandwidth and efficiency compared to server DDR5 registered dual in-line memory modules.
Vera Rubin's memory capacity underscores this shift. Each Rubin GPU contains eight 12-high HBM4 stacks, providing 288 gigabytes—50% more than Blackwell B200's 192 GB HBM3E but equivalent to Blackwell Ultra's capacity. The Vera CPU, however, supports up to 1.5 terabytes of LPDDR5X—more than triple the previous Grace CPU's 480 GB maximum. The Vera Rubin NVL72 system ties 72 Rubin GPUs and 36 Vera CPUs into a single server rack containing 74.7 TB of total DRAM: 20.7 TB of HBM4 and 54 TB of CPU LPDDR5X. To contextualize this scale, a premium smartphone typically carries 12 GB of LPDDR5X; a single Vera Rubin NVL72 rack contains roughly the equivalent of 4,500 smartphones in CPU LPDDR5X alone.
Samsung Electronics, SK hynix, and Micron are expanding their supply portfolios beyond HBM into low-power DRAM and storage. Samsung introduced HBM4, SOCAMM2, and the PM1763 storage product as "memory products for Nvidia." SK hynix is mass-producing a 192 GB SOCAMM2 using sixth-generation LPDDR5X at the 10-nanometer node. Micron began mass production in March of 12-high 36 GB HBM4 and 192 GB SOCAMM2 for Vera Rubin.
Supply dynamics present emerging risks. The Information reported that Nvidia is considering reducing HBM capacity on the next-generation "Rubin Ultra" GPU from its original specification. Nvidia initially planned to mount 1 terabyte of seventh-generation HBM (HBM4E) on each Rubin Ultra, using 16 HBM4E stacks. However, test samples reportedly carry lower capacities: some at 192 GB and others at 256 GB—both below the 288 GB currently shipping in Rubin GPUs. Supply constraints, particularly difficulties in producing sufficient HBM4E by the launch date, appear to be the primary driver of this downgrade.
Market research firm TrendForce identified additional configuration variations under review for Rubin Ultra, including 8-high HBM4E alongside the planned 12-high variant, and mixed configurations combining 12-high and 8-high HBM4. TrendForce also noted that anticipated LPDDR5X shortages continuing through 2027 have prompted Nvidia to reduce the SOCAMM capacity in the Vera Rubin superchip to approximately half its initial specification.
These specification adjustments reflect underlying supply-chain power dynamics. A semiconductor industry official observed: "The fact that Nvidia, which does not compromise on performance, yielded part of the specs can be interpreted as a result of the supplier-advantaged structure of memory corporations being reflected." While Nvidia's single-pole dominance in the AI accelerator market is gradually weakening, it remains the largest buyer in the AI memory market. How each of the three memory suppliers allocates production volumes will ultimately determine the magnitude of earnings improvement for Samsung, SK hynix, and Micron.