Nvidia Wants TSMC to Make Its Memory, and TSMC Is Reportedly Drafting Plans for 3D DRAM

Nvidia has run out of patience with the memory market. Its biggest supplier, TSMC, has started drafting plans to make a new kind of memory chip called 3D DRAM at Nvidia's request, two people with direct knowledge of the plans told The Information. If it happens, the company that already builds nearly every advanced AI processor would also start making the memory that sits next to them.

According to the report, TSMC could build new factories or buy an existing plant and convert it for memory, the same route Micron took this year when it bought a fab and turned it over to DRAM. There is no production date. The effort is led by Hsu Kuo-chin, the vice president who runs advanced packaging at TSMC and who used to head Micron's operations in Taiwan.

What 3D DRAM changes

Ordinary DRAM lays its memory cells out in a single flat layer. 3D DRAM stacks them vertically, the way flash storage has done for years, to pack more capacity into the same area. Nobody ships it in volume yet, which is part of the appeal for a newcomer: TSMC would not be trying to beat Samsung, SK Hynix and Micron at a game they have spent decades perfecting.

It would still be expensive. Buying a small specialty memory maker would give TSMC a fab, but not a leading-edge DRAM process. Building one would take billions of dollars and years of research.

TSMC's careful answer

TSMC did not confirm the story. A spokesperson said the company focuses on its core business as a dedicated foundry and wants to remain the trusted capacity provider for the logic chip industry. That is a statement of strategy more than a flat denial, and it does not mention memory at all.

There are good reasons for caution. Memory is a boom and bust business: shortages trigger new factories, and new factories end in gluts that crush prices. Making DRAM would also put TSMC in competition with the memory makers it works with today. SK Hynix partnered with TSMC in 2024 to produce the logic base dies for next-generation HBM, and in October Nvidia reportedly asked Samsung, SK Hynix and Micron to build the base layer of its custom HBM on TSMC's 3nm process.

Why Nvidia is asking

The shortage is the whole story. AI data centers are buying memory as fast as it is made, prices have soared, and Micron expects the squeeze to last into 2028. Apple raised Mac and iPad prices in June because of memory costs. Sam Altman has even argued that RAM sitting in personal computers is a poor use of the world's limited supply.

For Nvidia, three memory suppliers setting the pace of its business is a risk it would rather not carry. A fourth one, run by the partner it already trusts with its processors, would change that math. Whether TSMC wants the job is a different question, and so far its answer is that it makes logic chips.

Some offers on this page may be paid placements or contain affiliate links.