Nvidia is asking TSMC to turn the memory bottleneck into another foundry business

TSMC has begun drawing up plans to manufacture 3D DRAM, according to two people with direct knowledge of the work. Nvidia, its largest customer, requested the initiative as AI demand drives memory prices higher and leaves chip designers searching for additional supply. Unlike conventional DRAM, which arranges cells in a single plane, 3D DRAM builds them vertically to increase density. Samsung, SK Hynix and Micron are developing their own versions.

There is no production date, committed capacity or settled factory strategy. TSMC could build a new plant or acquire an existing fab and convert it; Micron paid $1.8bn for a Taiwanese Powerchip facility in March with a similar conversion in mind. The effort is reportedly led by advanced-packaging vice-president Hsu Kuo-chin, who previously ran Micron’s Taiwan operation. TSMC made ordinary memory chips until the late 2000s, but has since defined itself as a dedicated logic foundry.

The strategic logic extends beyond selling another kind of wafer. TSMC already manufactures the processors that consume high-bandwidth memory, supplies advanced packaging that connects those processors to HBM stacks and has a partnership with SK Hynix to manufacture the logic base die beneath next-generation HBM. Producing memory cells would let it coordinate more of that system—and would reduce Nvidia’s dependence on three suppliers whose capacity now limits the number and price of complete AI servers.

It would also import a business model TSMC has spent years avoiding. Memory is highly cyclical: shortages invite large capital programmes, new fabs arrive after demand has changed and standardized products then fall sharply in price. 3D structures add yield risk because every extra layer creates more opportunities for defects. Samsung, SK Hynix and Micron have decades of process knowledge in making enormous volumes cheaply.

The useful conclusion is therefore narrower than “TSMC enters memory.” Nvidia is probing whether its most capable manufacturing partner can remove a bottleneck that GPU design alone cannot solve. If TSMC proceeds, the AI hardware stack becomes more vertically coordinated around one foundry. If it declines, that will be evidence that even today’s extraordinary memory prices do not compensate for the cyclicality and execution risk.