Samsung Electronics plans to use rival chipmaker TSMC’s advanced manufacturing technology to produce a key component of NVIDIA’s next-generation AI memory chips, a major change in Samsung’s semiconductor strategy.

The agreement, first reported by the Seoul Economic Daily, is part of NVIDIA’s effort to improve the performance and supply of high-bandwidth memory, a crucial technology for AI computing. Samsung has traditionally manufactured its HBM components internally, but will now rely on TSMC’s 3-nanometer process for NVIDIA’s custom NVHBM chips.

NVIDIA is also requiring its other two HBM suppliers, SK Hynix and Micron Technology, to adopt TSMC’s 3nm manufacturing process for NVHBM. The three companies have been directed to use sixth-generation DRAM technology, known as 1c, for the memory dies.

For NVIDIA, the strategy addresses a challenge in AI infrastructure: obtaining consistent memory performance from multiple chip suppliers. Standardizing production could also reduce the testing required to qualify chips from different manufacturers.

On one level, the decision highlights TSMC’s leading role in the AI chip market, even as Samsung competes with the Taiwanese company in advanced chip manufacturing.

Big picture, the arrangement highlights trends in the AI hardware sector. NVIDIA’s advanced requirements are pushing competing semiconductor manufacturers toward a shared production process, while the industry’s ability to deliver faster AI systems remains closely tied to memory performance and manufacturing capacity.

“The future of HBM is custom and NVIDIA is now dictating how it gets built rather than just buying it,” said Brendan Burke, Research Director at The Futurum Group. “Putting all three memory suppliers on a TSMC base die shows how much leverage the world’s largest HBM buyer has and Samsung signing on keeps it inside NVIDIA’s custom HBM program. TSMC loses nothing here. It gains a wafer under every stack.”

More Space for Computing Circuits

NVIDIA unveiled NVHBM in August as a customized memory technology based on the HBM4E architecture. The design moves the memory controller, which manages communication between the GPU and memory, onto the base die at the bottom of the memory stack.

This change gives NVIDIA more room on its GPUs for computing circuits, which should boost overall processing performance. The company’s new design also reduces the physical interface layer by 60% compared with conventional HBM.

According to the reported technical specifications, NVHBM can deliver up to 30% greater memory bandwidth while consuming 15% less power than standard HBM4E.

These improvements matter for AI data centers, where memory bandwidth can limit the performance of expensive GPUs. Faster processors deliver less benefit when the memory system can’t supply data fast enough.

Manufacturing reliability is another factor driving NVIDIA’s decision. Industry estimates put TSMC’s 3nm manufacturing yield above 90%, compared with approximately 60% for Samsung’s competing 2nm process. Higher yields mean a greater percentage of manufactured chips meet quality requirements, a key consideration when suppliers are attempting to expand production of complex AI components.

The industry faces a substantial memory shortage. UBS forecasts that HBM demand will reach approximately 61.5 billion gigabits in 2027, a 91% increase from 2026. Memory manufacturers are expected to ship only 54.6 billion gigabits, leaving a supply gap of roughly 10%.

To keep up with growing demand, TSMC is expanding its manufacturing capacity. The company plans two additional 3nm production lines at its Fab 18 facility in Tainan, Taiwan, with manufacturing scheduled to begin in the first halves of 2027 and 2028. Additional 3nm production capacity is planned for TSMC facilities in Arizona and Kumamoto, Japan.

Samsung, meanwhile, will continue producing standard HBM4 and HBM4E memory using its own 4nm process for the base dies and sixth-generation DRAM technology for the memory dies, along with custom HBM products for other customers, including AMD.

Samsung is also evaluating whether its own 2nm manufacturing process could eventually produce NVHBM base dies if TSMC cannot meet NVIDIA’s demand.