Samsung Taps TSMC's 3nm Process for Nvidia's Next-Gen HBM

Samsung Electronics will use TSMC’s 3-nanometer manufacturing process to produce the base die for Nvidia’s next generation of custom high-bandwidth memory (HBM), according to a report in the Seoul Economic Daily. The base die is the system-semiconductor layer that sits beneath the memory core and helps manage data flow. Until now, Samsung has made both the core die and the base die for every HBM product it supplies, keeping the entire stack in-house.

The decision marks the first time Samsung has partnered with TSMC, the world’s largest contract chipmaker, at the HBM manufacturing stage. It also breaks Samsung’s one-stop, turnkey model for HBM, in which one company handles every step from memory cells to the logic layer. Nvidia, which designs the graphics processors that dominate AI training, has reportedly asked all three major HBM suppliers — Samsung, SK Hynix and Micron — to use TSMC’s process for the base die of its NVHBM products, while using sixth-generation DRAM technology for the core chips.

The base die has become more important as HBM has evolved. In earlier generations, the base die was a relatively simple logic chip. In custom HBM, it takes on more of the work of routing data between the GPU and the stacked memory layers, and its performance and power efficiency increasingly affect the whole system. Nvidia’s push to standardize on TSMC’s advanced node suggests it wants tighter control over that logic layer across suppliers.

For Samsung, the arrangement is a pragmatic concession. The company has invested heavily in its own foundry business and has tried to sell customers on a fully integrated memory-plus-logic offering. But Nvidia is the largest customer in AI accelerators, and its technical requirements carry weight. Accepting TSMC as a partner on the base die may help Samsung win or keep a larger share of Nvidia’s HBM orders, even if it means giving up some of the value it once captured internally.

For TSMC, the deal extends its reach into the memory supply chain. The company is already the main manufacturer of Nvidia’s GPUs, and producing HBM base dies adds another stream of advanced-node revenue. It also deepens the dependence of the AI hardware ecosystem on a single foundry, a concentration that has drawn attention as demand for AI chips has surged.

The report did not specify volumes, timelines or financial terms. It also did not say whether SK Hynix and Micron have agreed to the same arrangement, though Nvidia’s request was described as applying to all three. Representatives for Samsung, TSMC and Nvidia did not immediately comment on the report.

The shift comes as HBM has become one of the most contested parts of the semiconductor industry. AI servers need vast amounts of the memory, which is stacked vertically and placed close to the processor to reduce bottlenecks. Supply has struggled to keep pace with demand, and the major memory makers are racing to qualify their next generations with Nvidia. How Samsung balances its own foundry ambitions with Nvidia’s preference for TSMC’s process will be one of the key questions for the memory industry in the coming year.