NVIDIA’s Custom NVHBM Could Push Samsung and TSMC Into a Rare Manufacturing Partnership
NVIDIA is reportedly taking a more controlled approach to its next-generation high-bandwidth memory strategy, and the move could reshape how major memory makers build advanced HBM products. The company’s custom NVHBM technology is said to require a more standardized manufacturing process, with Samsung and TSMC working together on the base die using TSMC’s 3nm process.
Samsung has traditionally preferred a turnkey model for HBM production, handling both the core die and base die development through its own ecosystem. However, NVIDIA’s custom memory requirements appear to be pushing the company toward a different path. For NVHBM, Samsung is reportedly joining forces with TSMC to create a more unified production flow that reduces complexity and improves consistency.
The goal is simple: NVIDIA wants predictable performance across suppliers. To achieve that, the company has reportedly asked Samsung, Micron, and SK hynix to use the same manufacturing process for the NVHBM base die. By relying on TSMC’s 3nm node for this part of the memory stack, NVIDIA could avoid the need to separately qualify and verify each vendor’s version of the base die.
That could save significant time, especially as demand for AI accelerators and high-performance GPUs continues to surge. In the AI hardware market, even small delays in memory validation can affect product launches, shipment schedules, and customer deployments. A standardized NVHBM base die could help NVIDIA streamline production while maintaining tighter control over performance and efficiency.
NVIDIA’s custom HBM is expected to play an important role in future GPU platforms, including upcoming Feynman-generation products. The company has reportedly claimed that its in-house memory solution may deliver up to 30 percent higher memory bandwidth than standard HBM4E while reducing power consumption by as much as 15 percent.
Those gains are especially important for AI and data center workloads, where memory bandwidth and energy efficiency are critical. Modern AI models require massive amounts of data to move quickly between memory and compute units. Faster, more efficient memory can improve training and inference performance while helping reduce power and cooling demands in large-scale server environments.
For the core die, NVIDIA is reportedly requiring the use of sixth-generation DRAM technology, also known as 1c DRAM. Combining a 1c DRAM core die with a 3nm base die could give NVHBM a strong technical foundation, particularly as AI chips continue to push memory systems to their limits.
Samsung’s reported agreement to use TSMC’s 3nm process for the NVHBM base die marks a notable shift. It would be one of the first times the company steps away from its usual fully integrated HBM manufacturing strategy for a major custom memory product. Still, Samsung is not abandoning its own capabilities. The company is reportedly keeping its foundry options open in case NVHBM demand grows beyond what TSMC can supply.
This approach gives Samsung flexibility. If TSMC’s production capacity becomes constrained, Samsung could potentially use its own foundry resources to support additional demand. At the same time, standard HBM products are expected to remain under Samsung’s existing development model, with the company continuing to manage both core and base die production internally.
For Samsung, the stakes are high. HBM has become one of the most important growth areas in the semiconductor industry, fueled by explosive demand for AI accelerators, advanced GPUs, and high-performance computing systems. Strong HBM demand is also helping Samsung reduce pressure on its foundry business, which has faced losses in recent years.
Earlier estimates suggested that Samsung’s foundry losses could shrink significantly, falling from 6.74 trillion won, or around $5.1 billion, in 2025 to 3.92 trillion won, or around $2.95 billion, in 2026. That would represent a 41 percent reduction in losses, with HBM demand playing a major role in improving the company’s overall outlook.
The reported collaboration between NVIDIA, Samsung, TSMC, Micron, and SK hynix highlights how competitive and complex the AI memory market has become. NVIDIA wants more than just raw supply; it wants consistent quality, unified manufacturing standards, and memory products optimized specifically for its GPU roadmap.
If successful, NVHBM could become a key differentiator for NVIDIA’s future AI hardware. Higher bandwidth, lower power consumption, and a standardized production process may help the company maintain its lead in data center GPUs while giving memory suppliers a new path to participate in the booming AI infrastructure market.
At the same time, this shift shows that even the world’s largest semiconductor companies may need to cooperate in new ways to meet the technical demands of next-generation AI systems. For Samsung, working with TSMC on a custom NVIDIA memory product may be unusual, but the potential rewards could be too significant to ignore.
