Key points
- TSMC reportedly weighs 3D DRAM at Nvidia's request
- Samsung, SK Hynix, and Micron are working on 3D DRAM too
- Micron says its timing depends on cost and performance
TSMC (TSM) has begun developing plans to make 3D DRAM after a request from Nvidia (NVDA), its largest customer, The Information reported on Nov. 6, according to a summary of its report.
3D DRAM is a proposed way to increase memory capacity by building layers of storage cells within a chip. That's different from the high-bandwidth memory, or HBM, used with AI processors today, which stacks separate DRAM dies.
TSMC hasn't announced a plan. We found no timeline, investment figure, or production target in the coverage of The Information's report, and the full article is behind a paywall.
Why stack memory cells?
DRAM is the fast memory a processor uses while it works. For decades, chipmakers have added capacity by shrinking cells across the surface of a chip. That's getting harder.
"DRAM is nearing atomic limits at this point, although we have confidence in several more planar nodes," Scott DeBoer, president and chief technology and products officer at Micron (MU), said at an Oct. 21 webinar.
Here are four terms that come up in the discussion:
| Term | What it is | Where it stands |
|---|---|---|
| 6F² | The cell layout used in most DRAM today | In production |
| 4F² | A cell layout with a nominal area one-third smaller than 6F²; actual chip density and cost depend on the design and manufacturing process | CXMT plans server modules by year-end; Samsung says it's working on it |
| 3D DRAM | Layers of cells built inside one die | On the roadmaps of the three largest makers |
| HBM | DRAM dies stacked, typically on a base die | In production for AI chips |
Samsung, SK Hynix, and Micron have their own plans
The memory market is led by Samsung Electronics (005930), SK Hynix (SKHY), and Micron. None has given a firm date for 3D DRAM in the sources we read.
Micron put 3D DRAM on its roadmap between two future generations, 1-zeta and 1-eta, and called it a parallel path. "The timing of it will depend on Micron business decisions," DeBoer said. Micron said it has chosen not to pursue 4F² for its roadmap.
SK Hynix's roadmap listed 3D DRAM for about 2030, Tom's Hardware reported in November 2025. That's separate from the "3D Stacked DRAM" SK Hynix showed in October, which bonds DRAM chips onto a processor and targets 2028.
Samsung's chief technology officer for chips confirmed 4F² work on Oct. 14 and gave no timetable. ZDNet Korea reported that Samsung plans to use 4F² in a next-generation process for 3D DRAM.
TSMC already has a role in memory
TSMC is a foundry. It primarily manufactures chips designed by customers. It has been moving closer to memory, though.
The Seoul Economic Daily reported in October that Nvidia asked Samsung, SK Hynix, and Micron to use TSMC's 3-nanometer process for the base dies in its custom HBM, and that Samsung plans to do so. Making the memory cells themselves would be a much bigger step.
Memory has also been in short supply. The summary of The Information's report said that shortage has prompted Nvidia and Apple (AAPL) to look for more sources.
What isn't known yet
TSMC is weighing new plants or converting an existing one, and it says its focus remains its core logic foundry business, according to a summary drawn from social media posts about the report. We couldn't check that wording against the full article. The coverage doesn't say when production could begin.
Cost may decide the timing. DeBoer said Micron wants to introduce 3D DRAM "at the optimal time from a cost and performance point of view."
