The world is too loud. Read what matters.

Asianometry

Should TSMC Make 3D DRAM? Nvidia Asks, But History Haunts

TSMC tried twice to make DRAM for customers and both times failed; Nvidia's latest request forces it to revisit a business untouched for two decades, but this time it's not traditional DRAM—it's new stacked memory based on oxide semiconductors.

SemiconductorsTSMCDRAMAdvanced PackagingNvidiaCompute-Memory Integration
Medium-high information density, suited for those trying to understand what this vague rumor actually means; the middle section's explanation of the fundamental difference between logic and memory processes is the episode's most rigorous part.

The argument · tap a timestamp to hear it

4:05

Taiwan's entire generation of DRAM makers fell to the same crash

TSMC invested in and shared executives with Vanguard, which initially aimed to make DRAM but was crippled by a brutal memory downturn. It was eventually forced to exit memory and pivot to logic. In the same crash, Inotera and Rexchip were acquired by Micron, ProMOS was sold to other foundries, leaving only Nanya and Winbond still making mature or niche DRAM—a share negligible at best. This history shaped TSMC leadership's conservative stance on memory.

5:09

Morris Chang's memoir hides regret over the memory business

Reading Morris Chang's memoir for research, the author could plainly sense his bitterness and regret over Vanguard's DRAM failure. This explains why, for nearly two decades after, TSMC executives have been cautious about entering memory—organizational memory outlasts policy.

6:10

Logic means stable profit; memory is pure price war

Taiwan's semiconductor industry widely holds that logic foundry work won't spike like memory, nor crash like memory either. The ‘customer service’, technical depth, and care that foundry work demands are what Taiwan believes it does best, while memory is a commodity business where customers only care about price—a game TSMC does not want to play. If forced to, it would price itself out of contention.

7:12

DRAM capacitor etch complexity means TSMC cannot use existing fabs

DRAM's core is a tall, narrow cylindrical capacitor (1–3 micrometers deep, only 20–30 nanometers wide), and high-aspect-ratio etch to make it is considered one of the hardest steps in semiconductors. Then comes alternating deposition of metal, insulation, metal inside the trench. TSMC has barely done this type of work, so it would almost certainly need entirely new etch equipment and an entirely new fab—meaning this carries higher financial risk from the start.

9:14

What Nvidia asked for might not be for its own use

The report says TSMC is planning 3D DRAM at Nvidia's request, but that doesn't necessarily mean Nvidia needs it for itself. Startup d-Matrix has a product called ‘3D-DRAM’ that stacks compute dies on top of DRAM, connected via vertical I/O microbumps, enabling cheaper inference economics. In September 2026, Nvidia and d-Matrix announced a partnership—d-Matrix's next-generation Raptor XPU will integrate into Nvidia's AI infrastructure initiative. This 3D DRAM rumor is likely part of that collaboration.

10:14

The report is not about packaging but TSMC making DRAM dies itself

The author initially assumed this was just TSMC packaging memory wafers from others with its own logic dies—routine work. But The Information reported clearly that TSMC is weighing whether to actually manufacture 3D DRAM dies itself, possibly even acquiring an existing fab like Micron recently did from Powerchip, or building a new one. That doesn't match the intuitive explanation that ‘TSMC only does packaging’.

12:18

Copper interconnects cannot withstand the temperature needed to stack logic onto traditional DRAM

Silicon transistor fabrication needs exposure to temperatures up to around 1,000 degrees Celsius, but copper interconnects can only tolerate up to 400 degrees. So after front-end-of-line, you cannot add silicon transistors needed for traditional 1T1C DRAM. Oxide semiconductors like IGZO can be made at low temperature, meaning you can stack multiple oxide memory cells above the silicon transistors—exactly the direction TSMC researchers have been exploring recently.

15:24

Oxide storage's decisive factor is threshold voltage drift from defects

Oxide semiconductor memory has three technical approaches: direct 1T1C DRAM transplant, capacitorless gain cells leveraging oxide's ultra-low leakage to store charge better, and the more radical FeFET with single-transistor non-volatile storage. But all face the same fundamental problem: oxide accumulates defects over time, causing threshold voltage drift and rising leakage. Whoever solves this wins. Even if TSMC pursues this, it would take five to ten years to see results—by then the memory market landscape would be entirely different.

16:26

HBM erases the line between logic and memory, forcing TSMC's hand

HBM4 chips already use advanced-logic base dies, meaning more computation is shifting to memory's side—Samsung might use those HBM base dies to improve its own logic process, catch up, or even grab share in compute-memory integration, threatening TSMC's dominance in logic. The author believes under that threat, TSMC might have no choice but to actually make DRAM.

In their own words · checked verbatim

In 1996, the two announced that TSMC would make 16 megabit DRAMs for Fujitsu inside a new 8-inch fab using a 350 nanometer node.

With commodities, customers care most about the price - and maybe that is not the sort of game that TSMC wants to compete in.

Fabbing DRAM is all about the capacitor. The vertical cylinder-shaped capacitor that must be as tall and skinny as possible - 1-3 micrometers deep but just 20 to 30 nanometers wide.

But what is so weird about the Information report is that it is clearly not meaning a packaging product. It is saying straight up that TSMC is weighing whether to make the actual 3D DRAM dies themselves.

Unfortunately though, the copper interconnects can only tolerate temps of up to 400 degrees Celsius. So we cannot make any more silicon transistors after the front end

This is a fundamental issue and the first company to crack it will win.

CC Wei is coming into his own as a CEO and Chairman. And after building a whole fab in the US, one can imagine TSMC being a company that is willing to do hard things if that is what the customers want.

Figures

TSMC's memory business revenue share in 2005less than 1%2:58
DRAM capacitor depth and width1–3 micrometers deep, 20–30 nanometers wide6:58
Silicon transistor manufacturing temperature vs copper interconnect maximum~1000 degrees Celsius vs ~400 degrees Celsius12:18
Estimated time for 3D DRAM production ramp-up5–10 years15:24

Glossary

High-Aspect Ratio etching
Carving extremely deep structures in extremely narrow holes; one of the hardest etching steps in DRAM capacitor fabrication.
oxide semiconductor (IGZO etc.)
Using metal oxides instead of silicon for transistor channels; can be made at low temperature, enabling stacking.
Back End of the Line / Front End of the Line
Front-end transistor fabrication needs high temperature; back-end copper interconnects only tolerate low temperature—the process temperatures conflict.
gain cell
A capacitorless memory structure storing data bits using two transistors instead of a capacitor.
FeFET
A single-transistor non-volatile memory cell that stores data via ferroelectric gate polarization direction.

How to listen

Who it's for

Investors and engineers tracking foundry competition and AI chip supply chains, plus anyone trying to understand what the TSMC-Nvidia memory rumor actually says.

Skip

The opening disclaimer about Huang Renxun-related events (0:00-1:04) can be skipped—it's unrelated to the main topic.