Samsung to Use TSMC's 3nm Process for Nvidia's Custom HBM Base Dies

At the request of Nvidia, Samsung Electronics will utilize TSMC's 3-nanometer manufacturing process to produce the base die for Nvidia's next generation of custom high-bandwidth memory (HBM).¹ This marks the first instance of Samsung collaborating with TSMC during the HBM manufacturing phase.¹ The decision represents a departure from Samsung's established one-stop internal production model, where both the core memory and base logic dies were fabricated entirely on its own production lines.² This strategic change was prompted by specific technical demands from Nvidia for its upcoming products.¹³

The primary driver for this shift is Nvidia's new NVHBM architecture, which was officially announced on August 26, 2026.¹⁵ This framework standardizes the memory system by integrating a custom memory controller directly into the base die.¹ Nvidia has directed its three main HBM suppliers—Samsung, SK Hynix, and Micron—to use TSMC's manufacturing process for all NVHBM base dies to ensure uniformity.¹ The core memory dies stacked on top of this base will continue to use sixth-generation DRAM technology.¹

Samsung consented to Nvidia's requirement after completing an internal review of its own manufacturing capabilities.¹³ This internal assessment determined that the production yield of Samsung's own SF2 process could not meet the scale and reliability levels demanded by Nvidia.⁵ The substantial gap in advanced logic process yields between Samsung and TSMC is considered a key reason for Samsung's move to outsource the NVHBM base die.² Estimates from TechInsights suggest Samsung's SF2 process yield is between 50% and 60%, while TSMC's N3 process yield is reported to be higher than 90%.⁵

This change in strategy is specifically for Nvidia's custom NVHBM orders.¹ Samsung plans to continue its integrated, in-house manufacturing approach for standard HBM products and for orders from its other customers.¹ The company is also assessing a contingency plan to use its own foundry division to produce the NVHBM base dies.¹⁴ This backup option would be considered if demand for the component surpasses TSMC's available production capacity, according to unnamed sources cited by the Seoul Economic Daily.⁴

Samsung's decision aligns with a broader trend seen among major HBM producers.¹ Competitor SK Hynix, for example, had manufactured its base dies internally for all products leading up to HBM3E.¹ However, upon advancing to the HBM4 generation, SK Hynix also transitioned to using TSMC's logic foundry services for its base dies.¹ This shift by multiple top-tier memory makers indicates a growing reliance on specialized foundries for the advanced logic components required in next-generation HBM.¹

These strategic moves are taking place within a rapidly expanding market for high-performance memory.⁵ Data from Counterpoint Research projects that the global HBM market will reach a value of approximately $100 billion by 2028.⁵ As of the middle of 2026, SK Hynix was the market leader, holding a share of about 50%.⁵ Nvidia's NVHBM initiative, by mandating a single foundry for the base die, aims to create a more consistent and interchangeable supply from its key memory partners.¹

Nvidia's proprietary NVHBM architecture provides significant technical benefits when compared to the JEDEC HBM4E industry standard.⁵ It can deliver an increase in memory bandwidth of up to 30% per stack and reduce power consumption by 15%.⁵ The design also enables a reduction of up to 25% in the die area needed on the accelerator chip itself.⁵ By integrating the physical layer interface into the 3D stack, NVHBM reduces the area for the PHY and its supporting circuits on the accelerator by as much as 67% relative to the HBM4E standard.⁵

Concurrently with these strategic shifts in its memory business, Samsung is making substantial investments in its U.S. semiconductor manufacturing base.⁷ The company has pledged more than $22.8 million for workforce development and community programs in Central Texas.⁷ In 2025, Samsung's Austin and Taylor campuses are projected to inject $10.9 billion into the Central Texas economy.⁷ This economic activity is expected to support 28,746 jobs and generate $2 billion in local wages.⁷

Samsung's American expansion includes breaking ground in November for a second fabrication plant in Taylor, Texas.⁷ This new facility will span approximately 2.7 million square feet, with a target for mass production to begin in 2030.⁷ These investments are being made as the U.S. semiconductor industry anticipates a significant labor deficit.⁷ The industry group SEMI has estimated that the sector could face a shortage of up to 157,000 workers by the year 2030.⁷

TSMC, Samsung's partner and rival in this arrangement, is also expanding its manufacturing presence in the United States.⁷ The company's second fab in Arizona, which was recently completed, is scheduled to start mass production in the third quarter of 2027.⁷ TSMC CEO C.C. Wei also confirmed that the company has sent a team to assess Texas as a potential site for another factory.⁷ These parallel expansions highlight the intensifying investment and competition between the world's leading chipmakers on American soil.⁷

While navigating these complex changes in its core semiconductor business, Samsung continues to pursue ventures in other technology areas.⁶ For example, leaked specifications for a potential Samsung smart glasses device suggest it will use a Snapdragon AR1 Gen 1 chip.⁶ The specifications describe a device weighing 41 grams and equipped with 2GB of memory and 64GB of storage.⁶ This illustrates the company's continued diversification into consumer electronics alongside its strategic maneuvering in the foundry and memory markets.⁶