Nvidia has reportedly conducted a series of closely spaced visits to Samsung Electronics' semiconductor packaging campus in Cheonan, South Korea, signaling that verification of the next-generation high-bandwidth memory (HBM) supply chain is entering a final stage.
According to South Korean outlet New Daily, citing industry sources, Nvidia is scheduled to conduct an audit of Samsung's Cheonan facility on March 12 after visiting the same site during the first week of March. The short interval between inspections at the same production base is unusual and is widely interpreted in the industry as a formal on-site verification rather than a routine partner meeting.
The inspection is viewed as part of Nvidia's effort to confirm whether Samsung's HBM4 mass-production and packaging lines can support the requirements of the company's next-generation AI accelerator platform, Vera Rubin.
Focus shifts to supply reliability
Industry observers say the evaluation is focused less on peak performance metrics than on the ability to sustain a stable supply. While Samsung has highlighted HBM4 technology capable of reaching speeds of up to 13Gbps, the key issues under review are believed to include production yield, packaging integration, and delivery reliability.
Samsung's Cheonan campus serves as a major hub for the company's semiconductor packaging and backend manufacturing operations. HBM devices are produced by stacking between eight and sixteen DRAM dies on a base die and completing the chip through complex advanced packaging processes.
Even when the underlying chip design and fabrication technology are identical, variations in packaging precision, thermal management, and power control can affect performance and delivery schedules in large-scale production. For that reason, industry observers often regard packaging capability as a decisive factor in HBM's competitiveness.
Rubin HBM4 supply narrows to two suppliers
Attention in the memory sector has increasingly centered on which suppliers will provide HBM4 for Nvidia's Rubin accelerator platform. Industry sources indicate the supplier list has narrowed primarily to Samsung Electronics and SK Hynix, while Micron Technology is expected to focus on mid-range products rather than Rubin's top-tier accelerators.
The Korea Economic Daily reported that Nvidia has encouraged memory suppliers to develop higher-performance HBM4 products capable of exceeding the 8Gbps transfer rate specified by JEDEC. The Rubin platform is expected to adopt HBM4 operating at speeds above 10Gbps and to use 16 stacks with a total memory capacity of about 576GB.
The accelerator is expected to be unveiled at Nvidia's GTC 2026 developer conference in Silicon Valley and could deliver performance more than five times higher than current products.
Qualification progress advances
Samsung and SK Hynix have emerged as the leading contenders to supply HBM4 for Rubin. According to the report, Samsung has effectively passed quality testing for HBM4 operating at 10Gbps and 11Gbps, and recently shipped a small volume of completed products to Nvidia. SK Hynix is continuing to optimize its products to meet the higher-speed test requirements.
Because the HBM4 production process—from DRAM wafer input to final packaging—typically takes more than six months, the two companies could begin production as early as this month.
Industry observers say Nvidia's repeated inspections of Samsung's Cheonan packaging operations suggest the competition is entering a final verification stage before supply allocations are determined.
Article edited by Jack Wu
