Quick Summary
- The reported arrangement involves TSMC’s 3nm process for base dies in Nvidia’s new NVHBM memory design.
- Nvidia reportedly asked Samsung, SK Hynix and Micron to use TSMC’s process for base dies and sixth-generation DRAM for memory core dies.
- Samsung is reportedly keeping its in-house approach for standard HBM and may assess its own foundry capacity if demand exceeds TSMC’s.
Samsung Electronics is expected to use Taiwan Semiconductor Manufacturing Company’s 3-nanometre process for base dies in Nvidia’s new custom high-bandwidth memory, according to a report from Seoul Economic Daily cited by Reuters and The Indian Express on October 19. If confirmed, the arrangement would mark a notable collaboration between companies that compete across important parts of the semiconductor supply chain. The information is based on reported industry sources rather than a detailed public product announcement, so the precise scope and timetable should be treated as reported plans.
The memory design at the centre of the report is Nvidia’s NVHBM, unveiled in August. According to the cited report, it integrates Nvidia’s custom memory controller into the base die and aims to standardise implementation across multiple memory suppliers. Nvidia reportedly asked Samsung, SK Hynix and Micron Technology to use TSMC’s process for the base dies and sixth-generation DRAM technology for the core dies. A standardised approach could make it easier to coordinate memory products around Nvidia’s platform, but the manufacturing details and performance implications depend on the final design.
The base die is a separate part of the memory stack
High-bandwidth memory is important because AI accelerators need to move large volumes of data between processors and memory. As models and workloads grow, the memory subsystem can influence overall performance as much as raw compute capability. HBM stacks combine memory dies with a base die and advanced packaging, which creates a demanding manufacturing process. The base die has a different role from the memory core itself, so using a particular foundry process for one part does not mean the entire memory product is being manufactured by the same company.
The report says Samsung accepted Nvidia’s requirements after an internal review, while planning to keep its own production model for standard HBM and custom products for other customers. That suggests the reported cooperation is targeted at a particular Nvidia design rather than a wholesale change in Samsung’s semiconductor strategy. Samsung remains a major memory supplier and has its own manufacturing capabilities. Working with TSMC for one component could allow it to meet a customer’s design preference without abandoning its wider manufacturing footprint, although the cost and capacity implications have not been disclosed.
A targeted collaboration, not a full manufacturing shift
Samsung is also reportedly assessing whether its own foundry could produce the NVHBM base dies if demand exceeds TSMC’s capacity. That possibility shows a central constraint in advanced semiconductor production: even when a design is settled, manufacturers must secure the right process technology, packaging capacity and providey schedule. A supplier may prefer to keep more of the process in-house, but a customer with a large roadmap can influence how components are sourced. The report does not establish that Samsung will switch to its own foundry, only that it is assessing the option.
For Nvidia, the goal of standardising memory implementation across suppliers could help coordinate a complex hardware roadmap. For memory vendors, the opportunity may come with requirements that differ from their usual production arrangements. The benefits will depend on whether the process meets performance, yield, cost and reliability targets. A reported manufacturing partnership should not be treated as proof of final product specifications or supply volumes. Those details usually become clearer as a design moves through qualification and production planning.
The semiconductor industry increasingly depends on collaboration among companies that are also direct competitors. Advanced AI systems require processors, memory, packaging, networking and software to work together, and no single supplier necessarily controls every step. Such cooperation can improve compatibility, but it can also create dependencies on some specialised factories. The next evidence to watch is official confirmation, qualification progress, production timing and whether Samsung’s own foundry becomes part of the plan. Until those details emerge, this remains a reported sourcing arrangement rather than a confirmed commercial product launch.