A report on Friday put the world's largest chipmaker into a business it has never been in. The Information, citing two people familiar with the discussions, said TSMC had begun drawing up plans to manufacture 3D DRAM — memory cells stacked vertically, the DRAM analogue of what NAND flash has done for two decades — at the request of Nvidia, its largest customer. The effort is reportedly overseen by Hsu Kuo-chin, a TSMC vice president for advanced packaging and formerly the head of Micron's Taiwan operations.
By Saturday, TSMC had answered, sort of. "Please note that TSMC focuses on our core business at all times, which is dedicated IC foundry," a spokesperson told Tom's Hardware. "Our mission is to be the trusted technology and capacity provider of the global logic IC industry for years to come." The statement names neither 3D DRAM nor Nvidia. It does not say the planning work does not exist. Tom's Hardware read it as a denial, and a denial of a rumour is not quite the same document as a denial of a roadmap.
That ambiguity was enough for the market. Reports put the share move at roughly six percent, which is what a plausible second source for memory is worth right now.
Why the market moved
DRAM is a three-company business. Samsung, SK hynix and Micron hold roughly 87 percent of it between them; China's CXMT has taken about a tenth. No foundry has ever supplied DRAM at scale, because DRAM is not a logic process: it is a capacitor business, and the yield learning is measured in decades. A credible fourth supplier with TSMC's balance sheet and capacity would change the bargaining position of every buyer in the industry — including Nvidia, which currently negotiates high-bandwidth memory allocation with the same three companies it is asking to be disrupted.
Today's DRAM is one monolithic die, and its capacity is set by its area. The cells are approaching the point where they can no longer be shrunk without losing the charge they store. 3D DRAM stacks layers of capacitors and transistors on a single die, and in principle allows die-to-die or wafer-to-wafer bonding on top, exactly as NAND does. Density stops being a lithography problem and becomes a stacking problem.
The counterweight
Everyone is already working on it. Micron's public roadmap still includes at least four more generations of conventional planar DRAM before 3D arrives; Samsung and SK hynix have their own programmes under their own names; the Californian startup Neo Semiconductor has been showing "3D X-DRAM" since 2023. Nobody ships it commercially.
TSMC has an extra obstacle the memory makers do not: no planar DRAM process to build on. No DRAM fab, no cell library, no accumulated yield data. Tom's Hardware's assessment is that 3D DRAM could not ease the current shortage for at least five years — which is roughly the horizon on which it would matter, not the horizon on which it helps.
The reason it still counts
The AI buildout is now constrained less by logic than by memory. DRAM and NAND output has been absorbed by accelerator demand, and the bill is arriving in consumer prices for laptops and phones in the Christmas quarter. A second source of supply is the only structural answer; allocation games and price ceilings are not.
There is also a packaging argument, and it is probably the real one. TSMC already places memory makers' die next to its own logic in advanced packages, and in October a Seoul newspaper reported that Samsung would use TSMC's 3-nanometer process to make base dies for Nvidia's next-generation custom high-bandwidth memory. If compute and memory are going to be co-packaged regardless, the company that owns the packaging has a reasonable claim on the memory die too. Owning the whole stack is how you stop being a subcontractor.
What to watch
TSMC does not need to announce anything for this story to progress. Watch for DRAM process engineering hires in Hsinchu, for any filing or land purchase that looks like a memory fab, for Nvidia saying one word about it (it has said nothing), and for whether the spot price of commodity DRAM keeps climbing through the first quarter. The sceptical reading is simple and unglamorous: a customer asking a supplier to study something is not a customer placing an order, and a company that will not name the product in its own denial has not committed to making it.




