According to South Korean media The Seoul Economic Daily, Samsung Electronics has decided that for NVIDIA’s next-generation custom high-bandwidth memory (HBM), NVHBM, it will allow NVIDIA to have the core component, the Base Die, manufactured by long-time rival TSMC using its 3nm process. This marks the first time Samsung will cooperate with TSMC at the HBM manufacturing stage.
NVIDIA’s Rigid Push for Standardization
The root cause of this shift in the partnership between Samsung and NVIDIA lies in NVIDIA’s relentless pursuit of supply chain consistency. In August this year, NVIDIA unveiled NVHBM, a custom product built on the HBM4E (7th-generation HBM) architecture. Unlike standard HBM, NVHBM integrates NVIDIA’s proprietary memory controller directly into the Base Die, aiming to deliver seamless and consistent performance across products from different suppliers on NVIDIA’s GPU platforms.
Previously, Samsung, SK Hynix and Micron each adopted distinct processes and designs for Base Dies. If suppliers were allowed to follow their own specifications, NVIDIA would have to conduct cumbersome compatibility and performance verification separately for each vendor’s products — an unacceptable efficiency drain amid the fiercely competitive race for AI computing power. As a result, NVIDIA issued a mandatory requirement to the three suppliers: the Base Die of NVHBM must uniformly adopt TSMC’s 3nm process, while the core chips will use the 6th-generation DRAM (1c) process.
For Samsung, this requirement directly undermines the vertical integration advantage that its Device Solutions (DS) division prides itself on. Previously, the Base Dies of standard HBM4 supplied to NVIDIA and HBM4E under validation were all fabricated on Samsung’s in-house 4nm process, and the core chips were produced at Samsung’s Hwaseong and Pyeongtaek 1c DRAM fabs. However, for the NVHBM project, after internal deliberation, Samsung ultimately opted to accept NVIDIA’s terms to retain its position in NVIDIA’s high-margin AI supply chain.
Calculated Compromise: Safeguarding Standard Products and Keeping Options Open
While conceding on the NVHBM front, Samsung has not surrendered its foundry interests entirely. Reports confirm that Samsung’s turnkey strategy remains intact for standard HBM products. For custom HBM demands from other clients such as AMD, Samsung plans to keep its fully in-house production model covering both core chips and Base Dies.
Notably, while agreeing to use TSMC’s 3nm process, Samsung has not completely ruled out its own foundry capacity. Industry insiders disclosed that Samsung is internally validating the feasibility of producing NVHBM Base Dies on its own wafer fabrication lines. This move is viewed as a backup capacity strategy: should market demand for NVHBM exceed TSMC’s 3nm capacity ceiling, Samsung stands ready to fill the gap with its domestic production lines.
Summary
From an industry landscape perspective, this collaboration between Samsung and TSMC in HBM manufacturing is effectively orchestrated by NVIDIA leveraging its buyer dominance. For TSMC, this represents not only an additional advanced process order but also deeper embedding of its technical standards into the underlying HBM architecture, reinforcing its role as a central hub within the AI chip manufacturing ecosystem.
For Samsung, although it forfeited in-house production of the Base Die for this single custom product, it gains the opportunity for deep integration with NVIDIA’s next-generation AI platform. While Samsung has taken the lead in mass production and customer sampling of standard HBM4E, SK Hynix holds a solid market share established during the HBM3/3E era. Samsung’s top priority now is to avoid being sidelined in NVIDIA’s custom product lineup.
(Reprinted from https://news.eccn.com/）