Samsung has spent years promoting its ability to manufacture both memory chips and advanced processors under one roof. Now, Nvidia's demanding requirements for its next generation of AI hardware could see the South Korean company relying on one of its fiercest manufacturing rivals.
Samsung Electronics reportedly plans to use Taiwan Semiconductor Manufacturing Company's (TSMC) 3-nanometre manufacturing process to produce the base dies for Nvidia's upcoming custom high-bandwidth memory, known as NVHBM.
According to reports, Samsung would continue manufacturing the DRAM memory layers, while TSMC would produce the underlying logic component that manages communication between the memory and Nvidia's processors. The arrangement would bring two competing chipmakers together at a crucial stage of advanced AI memory production.
For a company that has long advertised its integrated manufacturing capabilities, outsourcing that component would be an unusual concession. It also illustrates how Nvidia's influence is extending into decisions about which factories produce individual parts of its AI hardware.
The reason lies in the changing design of high-bandwidth memory.
High-bandwidth memory, or HBM, consists of multiple DRAM chips stacked vertically and connected through tiny electrical pathways. Its ability to transfer large volumes of data quickly makes it essential for AI processors, which must continuously retrieve information while training models or generating responses.
Beneath those memory layers sits the base die, a logic chip that handles communication with the processor. Traditionally, much of the memory-control circuitry has occupied space on the main computing chip.
Nvidia's NVHBM architecture, introduced in August, relocates its custom memory controller into the HBM base die. The company says the design can deliver up to 30% greater memory bandwidth and reduce HBM power consumption by 15% compared with standard HBM4E. It could also free up to 25% more space on the main processor die for computing circuits.
Those are Nvidia's stated design targets, not independently verified performance figures for mass-produced NVHBM systems.
The reported manufacturing arrangement reflects the increasing complexity of building such memory. Advanced logic processing is becoming essential even in products traditionally associated with DRAM manufacturing, an area where Samsung and its South Korean rival SK Hynix have long competed.
Samsung already manufactures HBM4 using its own 4-nanometre base die technology. The company began commercial shipments in February and has outlined plans to expand production as demand grows. Custom NVHBM, however, would require adapting manufacturing to Nvidia's architecture and production specifications.
The October 18 report also suggests Samsung is evaluating its own 2-nanometre foundry process as another potential source of NVHBM base dies. The reported use of TSMC therefore need not mean abandoning in-house manufacturing altogether.
TSMC has considerable capacity and expertise in advanced logic fabrication. Its third-quarter financial results show that 3-nanometre technology accounted for 33% of wafer revenue, reflecting the commercial scale of a process increasingly central to advanced computing.
Samsung, TSMC and Nvidia have not publicly confirmed the reported manufacturing arrangement or disclosed its commercial terms.
For Nvidia, the proposed approach could make it easier to source memory from multiple manufacturers while maintaining tighter control over performance. Samsung, SK Hynix and Micron would compete to supply memory built around Nvidia's custom specifications.
The development comes as semiconductor companies face renewed scrutiny over the costs of AI investment and volatile valuations across Asian chip markets.
Samsung's challenge is increasingly about securing a place in the supply chain that Nvidia is helping redesign. Producing advanced memory remains essential, but the logic beneath it is becoming just as consequential.
