TSMC Weighs Building 3D DRAM Capacity at Nvidia's Request, Reports Say

TSMC has begun planning production of 3D DRAM at the request of Nvidia, according to The Information, which cited two people familiar with the matter in a report carried by Reuters. The development sits squarely within the AI compute bottleneck: surging demand for AI semiconductors has tightened DRAM supply and pushed prices sharply higher, prompting chip designers to seek additional sources, with AI Times reporting that Nvidia and Apple are among those looking. Nvidia is described as TSMC's largest customer, and ETtoday noted that memory manufacturing would widen TSMC's share of the AI hardware supply chain.

DRAM holds data temporarily inside devices. 3D DRAM stacks memory cells vertically rather than arranging them on a single horizontal plane, raising capacity. AI Times noted that the architecture differs structurally from high-bandwidth memory, which stacks whole DRAM chips to raise data-transfer speeds. Samsung Electronics, SK Hynix and Micron are also developing 3D DRAM to move past the scaling limits of conventional designs.

TSMC is weighing two routes to capacity: a dedicated new fab, or the purchase of an existing plant converted to memory output, which Liberty Times reported would mirror Micron's acquisition of a wafer fab this year for conversion to DRAM production. According to the reports, the effort is led by vice-president Hsu Kuo-chin, who oversees advanced packaging and previously ran Micron's Taiwan operations. Entry would place TSMC in a market dominated by Samsung, SK Hynix and Micron; the sources do not say which of the two Korean makers would be more exposed. AI Times noted that SK Hynix and TSMC agreed in 2024 to cooperate on HBM4 and advanced packaging, with foundry processes supplying the base die, and that Nvidia has asked the three memory makers to build NVHBM base dies on TSMC's 3-nanometre process — cooperation that could run alongside competition in next-generation DRAM.

AI Times cautioned that vertical stacking complicates manufacturing and raises defect risk, so securing volume production could take years, while the incumbents hold decades of memory experience. Memory cyclicality is a further variable: capacity added during a shortage can meet weaker demand and lower prices by the time output begins. TSMC told The Information it does not comment on market rumours or speculation and would keep focusing on its core foundry business. TSMC and Nvidia had not immediately responded to Reuters requests for comment, and Nvidia declined to comment to AI Times. According to the reports, no investment scale or production start date has been set.