An estimate found that memory semiconductors account for about 62% of the cost of "Vera Rubin," Nvidia's next-generation artificial intelligence (AI) platform now being shipped in earnest. The core computing unit, the superchip, which bundles one Vera central processing unit (CPU) and two Rubin graphics processing units (GPUs), costs $38,902 (about 55.4 million won), of which sixth-generation high-bandwidth memory (HBM), HBM4, and SOCAMM2 account for $24,297 (about 34.6 million won).
According to an analysis of the Vera Rubin superchip bill of materials (BOM) from global investment bank (IB) UBS obtained by ChosunBiz on the 10th, the cost related to Rubin GPUs—including HBM4, the interposer, advanced packaging, and peripheral components—was estimated at $9,247 (about 1.317 million won). The Vera CPU, including SOCAMM2, was analyzed at $20,059 (about 2.856 million won), and other board components at $350 (about 500,000 won).
The expense of HBM4 mounted on the Rubin GPU was estimated at $4,943 (about 704,000 won), or 12.7% of the total cost. SOCAMM2 connected to the Vera CPU accounted for $19,355 (about 27.56 million won), or 49.8%. It is a structure in which CPU memory represents a far larger share of the cost than HBM4.
◇ Musk: "the best architecture"… strong sales expected
Market reaction to Vera Rubin is positive. Nvidia said last month that Vera Rubin NVL72 is running on Google Cloud, Microsoft (MS) Azure, Oracle Cloud, CoreWeave, and Nevius, among others.
SpaceX CEO Elon Musk also called Vera Rubin "the best architecture" during an earnings announcement on the 4th and laid out a plan to build AI infrastructure centered on Nvidia GPUs. SpaceX aims to secure more than 2 gigawatts (GW) of computing infrastructure by the end of this year and increase it to nearly 10 GW by the end of next year.
In the market, such system builds and purchasing plans by major customers are seen as signals that strong sales of Vera Rubin could continue. If sales volumes remain high for Vera Rubin following Blackwell, memory sales to Nvidia by the three memory companies—Samsung Electronics, SK hynix, and Micron—will also grow. The three memory makers are expanding their product lines beyond HBM4 to SOCAMM2 and enterprise solid-state drives (eSSD).
◇ GB300 53% → Vera Rubin 62%… SOCAMM2 widened the gap
Memory was estimated to account for 53% of the total cost of the previous Blackwell Ultra–based GB300. Compared with Vera Rubin, the share of memory semiconductors rose by 9 percentage points (P). UBS estimated the total cost of the GB300 superchip at $18,203 (about 25.92 million won) and the memory cost at $9,714 (about 13.83 million won). While Vera Rubin's total cost increased about 2.1 times from its predecessor, memory expense rose 2.5 times, a steeper climb.
The key driver of the rising memory cost share is SOCAMM2. Each Rubin GPU contains eight 12‑high HBM4 stacks, providing a total of 288 gigabytes (GB). That is 50% more than the Blackwell B200's HBM3E 192 GB, but the same capacity as the Blackwell Ultra B300's 288 GB. In contrast, the Vera CPU supports up to 1.5 terabytes (TB) of LPDDR5X. That is more than triple the maximum 480 GB of the previous Grace CPU.
SOCAMM2 is a product that modularizes low-power DRAM LPDDR5X—mainly used in smartphones, tablets, and laptops—for AI servers. Multiple DRAMs are bundled and placed close to the CPU and designed to be replaceable. Although it has lower bandwidth than HBM, it can deliver large capacity with less power and offers higher bandwidth and power efficiency than server DDR5 registered dual in-line memory modules (RDIMMs).
Vera Rubin NVL72 is an AI system that ties 72 Rubin GPUs and 36 Vera CPUs into a single server rack. One rack contains a total of 74.7 TB of DRAM, including 20.7 TB of HBM4 and 54 TB of CPU LPDDR5X. A premium smartphone typically carries 12 GB of LPDDR5X, which means the CPU LPDDR5X in a single Vera Rubin NVL72 rack alone is roughly equivalent to about 4,500 smartphones.
◇ From HBM to SOCAMM2 and eSSD… widening supply scope
Samsung Electronics, SK hynix, and Micron are widening their supply scope from HBM to low-power DRAM and storage. Samsung Electronics introduced HBM4, SOCAMM2, and the PM1763 for storage as "memory products for Nvidia." SK hynix is mass-producing a 192 GB SOCAMM2 applying sixth-generation (1c) LPDDR5X at the 10‑nanometer (nm, one-billionth of a meter) class. Micron also said in March that it began mass production of 12‑high 36 GB HBM4 and 192 GB SOCAMM2 for Vera Rubin.
However, Nvidia's spec adjustments due to memory supply shortages are cited as a variable for the earnings of the three memory companies. IT outlet The Information reported that Nvidia is considering reducing the HBM capacity of the next-generation GPU "Rubin Ultra" from the original plan. It tested at least three samples in recent weeks, and some reportedly carry less memory than initially disclosed.
Nvidia initially unveiled a plan to mount 1 TB of seventh-generation HBM (HBM4E) on a single Rubin Ultra GPU. The design applies 16 HBM4E stacks. But some samples under test are said to be at the 192 GB level, and others at 256 GB. That is less than the 288 GB of HBM4 on the Rubin GPUs now in mass production. One cited reason for the spec change is the possibility that suppliers will find it hard to produce sufficient HBM4E by the launch date.
Market research firm TrendForce also analyzed that Nvidia is reviewing several HBM configurations for Rubin Ultra. It reportedly included, in addition to the existing 12‑high HBM4E, 8‑high HBM4E as well as 12‑high and 8‑high HBM4 among the candidates. It also said that, considering the possibility of LPDDR5X shortages continuing through 2027, the SOCAMM capacity of the Vera Rubin superchip has been lowered to about half of the initial plan.
A semiconductor industry official said, "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," adding, "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."