Nvidia shipping Vera Rubin GPU while HBM costs surge across Korean suppliers, tightening memory supply amid demand spike.

Memory semiconductors account for approximately 62% of the cost of Vera Rubin, Nvidia's next-generation artificial intelligence (AI) platform now entering mass production. The superchip—which integrates one Vera CPU (central processing unit) and two Rubin GPUs (graphics processing units)—costs $38,902 (approximately 55.4 million won). Of this total, memory components account for $24,297 (approximately 34.6 million won): sixth-generation high-bandwidth memory (HBM4) and SOCAMM2.

According to a bill-of-materials analysis of the Vera Rubin superchip from global investment bank UBS, obtained by ChosunBiz on August 10th, Rubin GPU-related costs—including HBM4, interposer, advanced packaging, and peripheral components—total $9,247 (approximately 1.317 million won). The Vera CPU, including SOCAMM2, costs $20,059 (approximately 2.856 million won), with other board components adding $350 (approximately 500,000 won).

HBM4 mounted on the Rubin GPU costs $4,943 (approximately 704,000 won), or 12.7% of total cost. SOCAMM2 connected to the Vera CPU accounts for $19,355 (approximately 27.56 million won), or 49.8%—a structure where CPU memory represents a far larger cost share than HBM4.

Market reaction to Vera Rubin has been positive. Nvidia reported last month that Vera Rubin NVL72 systems are running on Google Cloud, Microsoft Azure, Oracle Cloud, CoreWeave, and Nevius, among others. 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 (GW) of computing infrastructure by year-end and increase it to nearly 10 GW by the end of next year. Industry observers view such deployments and purchasing plans from major customers as signals of continued strong Vera Rubin sales. If sales volumes remain high following Blackwell, memory shipments from Samsung Electronics, SK hynix, and Micron to Nvidia will also grow. All three are expanding their product lines beyond HBM4 to include SOCAMM2 and enterprise solid-state drives (eSSD).

Memory semiconductors accounted for 53% of the total cost of the previous Blackwell Ultra–based GB300. In Vera Rubin, the memory share has risen to 62%—a 9 percentage point increase. UBS estimated the GB300 superchip's total cost at $18,203 (approximately 25.92 million won) with memory costs of $9,714 (approximately 13.83 million won). While Vera Rubin's total cost increased approximately 2.1 times compared to its predecessor, memory expense rose 2.5 times—a steeper climb. SOCAMM2 is the primary driver of this rising cost share.

Each Rubin GPU contains eight 12-high HBM4 stacks, providing 288 gigabytes (GB) total—50% more than the Blackwell B200's HBM3E (192 GB) but equal to the Blackwell Ultra B300's capacity. By contrast, the Vera CPU supports up to 1.5 terabytes (TB) of LPDDR5X, more than triple the previous Grace CPU's maximum of 480 GB.

SOCAMM2 modularizes low-power DRAM (LPDDR5X)—traditionally used in smartphones, tablets, and laptops—for AI servers. Multiple DRAM modules are bundled, placed adjacent to the CPU, and designed to be replaceable. Though offering lower bandwidth than HBM, it delivers large capacity with less power consumption and provides higher bandwidth and power efficiency than server DDR5 registered dual in-line memory modules (RDIMMs).

The Vera Rubin NVL72 is an AI system combining 72 Rubin GPUs and 36 Vera CPUs in a single server rack. One rack contains 74.7 TB of DRAM total: 20.7 TB of HBM4 and 54 TB of CPU LPDDR5X. A premium smartphone typically carries 12 GB of LPDDR5X; the CPU LPDDR5X in a single Vera Rubin NVL72 rack alone is roughly equivalent to approximately 4,500 smartphones.

Samsung Electronics, SK hynix, and Micron are broadening their supply scope from HBM to low-power DRAM and storage. Samsung introduced HBM4, SOCAMM2, and the PM1763 storage drive as "memory products for Nvidia." SK hynix is mass-producing 192 GB SOCAMM2 modules with sixth-generation (1c) LPDDR5X at the 10-nanometer class. Micron announced in March that it began mass production of 12-high 36 GB HBM4 and 192 GB SOCAMM2 for Vera Rubin.

However, Nvidia's specification adjustments due to memory supply constraints present a variable for the three memory companies' earnings. IT outlet The Information reported that Nvidia is considering reducing the HBM capacity of the next-generation GPU "Rubin Ultra" from original plans. The company tested at least three samples in recent weeks, some carrying less memory than initially disclosed.

Nvidia initially planned to mount 1 TB of seventh-generation HBM (HBM4E) on a single Rubin Ultra GPU using 16 HBM4E stacks. Some samples under test reportedly carry 192 GB, others 256 GB—less than the 288 GB of HBM4 on current mass-production Rubin GPUs. One cited reason for the specification change is suppliers' potential difficulty producing sufficient HBM4E by the launch date.

Market research firm TrendForce also analyzed that Nvidia is reviewing multiple HBM configurations for Rubin Ultra, including existing 12-high HBM4E, 8-high HBM4E, and 12-high and 8-high HBM4 variants. Additionally, considering potential LPDDR5X shortages extending through 2027, the SOCAMM capacity of the Vera Rubin superchip has been reduced to approximately half of the initial plan.

A semiconductor industry official stated, "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. Although Nvidia's single-pole dominance in the AI accelerator market is gradually weakening, it remains the largest buyer in the AI memory market, so how each supplier allocates volumes will determine the degree of earnings improvement for the three memory companies."