Samsung and SK Hynix have been selected as Nvidia Rubin HBM4 suppliers, with shipments expected to begin in March.

```

The supply landscape for high-bandwidth memory (HBM) for Nvidia’s next-generation AI accelerator, Vera Rubin, is taking shape, with both Samsung Electronics and SK Hynix selected as HBM4 suppliers. The delivery window may be opening soon.

On Monday, according to TrendForce citing sources in Korea’s Economic Daily, both Korean memory giants have been included in the supplier list for Nvidia’s Vera Rubin components. Since the manufacturing cycle of HBM4, from DRAM wafers to final packaging, exceeds six months, the two companies are expected to begin mass production as early as this month. This time frame aligns closely with Nvidia’s GTC conference scheduled for next week, drawing significant market attention to HBM4 supply dynamics for Vera Rubin.

In terms of supply share, insiders expect SK Hynix to handle more than half of Nvidia’s total HBM supply in 2026 (including HBM3E), while Samsung is anticipated to take the lead in HBM4 business specifically for Vera Rubin.

According to TrendForce’s forecast, SK Hynix will maintain its industry-leading position with 50% of the global HBM bit output, though that’s down from 59% in 2025; Samsung’s share will climb from 20% to 28%, showing clear momentum.

Technical barriers: Nvidia sets ultra-high performance requirements

The core dividing line in this HBM4 supply competition is data transfer rate. According to Korea Economic Daily, Nvidia has required data speeds exceeding 10Gb/s for HBM4 in Vera Rubin, much higher than the 8Gb/s ceiling set by JEDEC’s industry standards.

In terms of qualification progress, Samsung has practically passed Nvidia’s two-level HBM4 qualification test—corresponding to data rates of 10Gb/s and 11Gb/s. In comparison, SK Hynix is still optimizing its product to pass the more advanced 11Gb/s test. Analysts point out that HBM4’s data processing speed may become a key differentiating factor between the two suppliers.

At the product level, HBM4 adopts a structure of stacking 8 to 16 DRAM chips atop a control substrate. Nvidia’s Vera Rubin is expected to feature 16 HBM4 stacks, achieving a total capacity of 576GB, surpassing the 432GB limit of AMD’s upcoming MI450 and giving it an edge in flagship AI accelerator memory capacity competition.

Supplier structure: Micron targets the mid-range, SK Group’s top executive to attend GTC in person

Among the three major HBM suppliers, Micron’s positioning is different. According to Korea Economic Daily, Micron has not entirely exited the HBM4 market and is expected to provide HBM4 for mid-range AI accelerators focused on inference (such as Rubin CPX), rather than the flagship Vera Rubin.

It is noteworthy that SK Group Chairman Chey Tae-won will attend Nvidia’s GTC conference in person for the first time. According to industry sources, his on-site presence underscores SK Group’s high level of attention, and Chey will directly oversee the supply schedule and negotiations regarding SK Hynix’s deliveries.

In terms of delivery progress, Samsung Electronics began HBM4 shipments in February this year, while SK Hynix has yet to announce a delivery schedule, and the timing gap between the two has raised market concerns.

Samsung’s leverage: Regular DRAM price surge reshapes supplier negotiating positions

The sharp rise in regular DRAM prices is changing HBM suppliers’ strategic calculations and gives Samsung extra bargaining leverage. According to TrendForce, since the fourth quarter of 2025, standard DRAM prices have surged, narrowing HBM’s traditional advantage in profitability. As a result, memory manufacturers are now adjusting the capacity allocation between HBM and standard DRAM to balance overall revenue growth and customer commitments.

Against this backdrop, the risks of Nvidia relying too heavily on a single supplier have become more apparent—if a supplier shifts capacity to regular DRAM, the Vera Rubin platform’s ramp-up progress could be constrained.

Currently, server-class standard DRAM modules (such as SOCAMM2) sell for about $1.3 per Gb, already close to the price of Nvidia’s flagship HBM3E. This means that for Samsung, producing regular DRAM without expensive stacking technology may be just as profitable as HBM4. Hence, by holding both HBM4 and regular DRAM capacity, Samsung can offer diversified price options to Nvidia and holds a more favorable position in supply negotiations.

Risk DisclaimerThe market is risky, and investments should be made with caution. This article does not constitute personal investment advice, nor does it consider the specific investment objectives, financial situation, or needs of any particular user. Users should consider whether any opinions, views, or conclusions in this article are appropriate for their circumstances. Investment decisions made based on this content are at your own risk. ```