Samsung to tap TSMC’s 3nm foundry for Nvidia’s next-generation NVHBM base die
In an unprecedented manufacturing shift, Samsung accepts Nvidia’s standard to produce custom AI memory on rival silicon.
Technology news
Samsung Electronics will adopt TSMC’s 3-nanometer foundry process to manufacture the base die for Nvidia’s next-generation custom high-bandwidth memory, known as NVHBM. The strategic shift, reported Sunday by the Seoul Economic Daily and Reuters, marks the first time Samsung has partnered with its chief foundry rival at the HBM fabrication level.[1]
Nvidia requested that all three of its primary memory suppliers—Samsung Electronics, SK Hynix, and Micron Technology—standardize the NVHBM base die on TSMC’s 3nm line while fabricating the core memory dies on sixth-generation 1c DRAM processes. NVHBM, based on seventh-generation HBM4E architecture, integrates Nvidia’s proprietary memory controller directly into the base logic layer to streamline validation across heterogeneous accelerators.[1]
Samsung’s Device Solutions division agreed to Nvidia’s technical mandate following an extensive internal review. While Samsung will retain its proprietary turnkey model—using in-house 4nm foundry lines for standard HBM4/HBM4E and custom chips for customers like AMD—it is also qualifying its own advanced lines to produce NVHBM base dies as a supply-chain hedge if TSMC capacity falls short.[1]
Our take: standardizing custom base dies on TSMC eliminates packaging discrepancies for Nvidia’s Rubin-generation AI platforms, but it represents a significant concession for Samsung Foundry. Relinquishing the logic layer to its primary competitor underscores Nvidia’s immense pricing and architectural leverage across the semiconductor supply chain. However, Samsung's core die memory output will remain vital to satisfying unyielding datacenter demand.