TSMC is reportedly planning to make memory chips. On November 6, 2026 The Information reported that the world's largest contract chipmaker has begun planning production of 3D DRAM at the request of Nvidia, its biggest customer. Reuters relayed the report. TSMC and Nvidia have not commented.
Update (November 7, 2026): TSMC says it remains a logic foundry
On November 7, a TSMC spokesperson told Tom's Hardware the company focuses on its core dedicated IC foundry business and intends to stay the trusted technology and capacity provider of the global logic IC industry. The statement does not confirm or deny internal planning reported by The Information; TSMC repeated that it does not comment on market rumors. Tom's Hardware and other trade press read the wording as a public rejection of entering commodity DRAM, while noting TSMC already partners with memory makers on HBM base dies and advanced packaging.
Industry commentary on November 7 — including remarks reported by TechNews from Winbond chairman Jiao You-jun — emphasized that DRAM requires different process IP than logic and that TSMC's tight relationships with Samsung, SK hynix, and Micron make head-on competition unlikely. Analysts cited in the same coverage argue 3D DRAM is years from easing shortages even for established memory vendors.
This is a single-source report with no timeline. This post sets out what it says, what it leaves out, and how it fits the memory story we have been tracking.
TL;DR: what is reported
| Question | What the reports say |
|---|---|
| What is the claim? | TSMC recently began planning to make 3D DRAM |
| Who asked? | Nvidia, TSMC's biggest customer, according to the report |
| Source | Two people with direct knowledge of the matter, per The Information |
| Who leads it? | Hsu Kuo-chin, a TSMC vice president who oversees advanced packaging |
| How would TSMC do it? | New factories, or buying an existing plant and converting it to memory |
| When? | Not stated |
| Confirmed by TSMC or Nvidia? | No. Neither responded to Reuters |
| Why now? | An AI-driven memory shortage has pushed up prices and sent chip designers looking for supply |
What is 3D DRAM, and how is it different from HBM?
Reuters describes 3D DRAM as a memory architecture that stacks memory cells vertically to increase capacity. Ordinary DRAM lays cells out on a flat plane, and it is hard to shrink them further. Stacking cells upward is one way to keep adding capacity per chip.
High-bandwidth memory, or HBM, is a different idea. HBM stacks several finished DRAM dies on top of each other and connects them with very wide links, then sits next to the AI processor. The bottom of each stack is a base die. Nvidia asked its HBM suppliers to make that base die on TSMC's 3nm process, as we covered in Samsung, TSMC and the NVHBM base die. So TSMC is already part of the HBM story, but as a logic maker for the base, not as a producer of the memory cells.
That distinction is the key point. Planning 3D DRAM would move TSMC into the cell-making business that Samsung, SK Hynix and Micron dominate. Reuters notes that if TSMC enters, it would compete with those three. Studying a new memory architecture does not by itself mean entering the existing HBM market.
How does this fit the timeline of Nvidia, TSMC and memory?
The report lands in a run of memory-supply stories. In October, a Seoul newspaper report relayed by Reuters said Nvidia had asked its three HBM suppliers, Samsung, SK Hynix and Micron, to make the base die for its custom NVHBM memory on TSMC's 3nm process, and wanted the memory dies standardised on 1c DRAM. Samsung reportedly accepted after an internal review while keeping its in-house model for standard HBM and other customers' custom products. That earlier report put TSMC deeper inside the memory stack through logic.
The new report goes one step further. It says TSMC is looking at the memory cells themselves. Read together, the two stories describe a customer that wants more of its memory supply chain to run through a partner it already works with closely. Neither story shows that this will happen. Both are reports with unnamed sources, and the October one was also only partly confirmed by the companies involved.
How can you check a report like this?
Supply-chain scoops are often right in direction and wrong in detail. A few habits help:
- Find the original. Ask which outlet broke it and how many sources it cites. Here it is The Information, with two people with direct knowledge.
- Separate the verbs. "Began planning" and "considering" are weaker than "will build" or "has ordered". Only the first two appear in the coverage.
- Look for numbers. Capex, wafer capacity, a site and a date are what turn a rumour into a project. None has been reported.
- Wait for the counterparty. The company named as the customer or the supplier often says something at the next earnings call or trade show.
- Watch the second wave. If Taiwanese or Korean trade press adds equipment or site details independently, confidence rises.
What does the report not say?
The gaps are large, and they matter more than the headline.
- No timeline. Reuters says the report did not say when production would start.
- No location or size. There is no fab site, no capacity, and no investment figure in what we could read.
- No decision. "Began planning" and "considering" are the words used. A plan can be dropped.
- No comment from the companies. Neither TSMC nor Nvidia responded to Reuters.
- One origin. Every outlet we found traces back to The Information. We found no second independent report, and we have not read The Information's article, which is paywalled.
- Share moves are not proof. Reuters' page showed TSMC and Nvidia shares higher in a live ticker, but the page did not tie the moves to the report, and a ticker can reflect many things. We treat market reaction as context only.
Reuters adds one comparison: TSMC is also reportedly exploring buying an existing plant and converting it, similar to Micron's move this year to buy and convert a fab for DRAM. That detail is part of the same report, so it carries the same single-source caveat.
Why would Nvidia want this?
The reported motive is supply. AI accelerators need huge amounts of fast memory, and an AI-driven shortage has pushed memory prices up. Our history of the cycle is in Stanford's memory price history for DRAM, HBM and NAND, and the supply pressure shows up in other stories: Samsung's second Taylor fab and HBM base dies, Micron pausing a Taoyuan plant, and YMTC's fourth Wuhan fab for NAND.
A customer that large has two levers when memory is tight. It can sign long contracts with existing suppliers, or it can encourage a new supplier it already trusts. Nvidia already depends on TSMC for its processors and for advanced packaging. Our report on TSMC's reported CoWoS price rise covers how tight that packaging step is. Asking the same partner to look at memory would keep more of the stack under one roof. That reading is ours, not something the report states.
For the technical side of why memory dominates AI system design, see our explainer on Nvidia's five co-design ideas around HBM bytes and NVFP4.
How hard is it for a foundry to make memory?
Memory is not logic. A logic foundry like TSMC builds custom chips to a customer's design. A DRAM maker builds standard products at huge volume, tuned for cell density, yield and cost per bit over many years. The processes, equipment and know-how differ, which is why only a few companies make DRAM at scale.
Two things follow from that, and both are our analysis and not a report finding:
- Any new fab takes years. Whatever TSMC decides, nothing from this plan is likely to ease the current shortage soon. Buying and converting a plant is quicker than building one, but still not a matter of weeks.
- Entry would pressure margins over time. Today's memory makers set prices in a shortage. A well-funded entrant backed by the biggest AI buyer could change that, though only if TSMC reaches competitive yields and costs.
None of this is certain. The report may describe an early study that goes nowhere, which happens often in chip planning.
What should you watch next?
- A statement from TSMC. The company holds regular earnings calls and may be asked directly. A "no comment" and a clear denial mean different things.
- A second source. Taiwanese and Korean trade press often follow up on supply-chain reports with equipment-order or site details.
- Memory-maker reactions. Samsung, SK Hynix and Micron may speak about capacity and pricing in their own reports.
- Equipment orders. A real fab plan shows up in tool orders long before chips ship.
- Any Nvidia comment on its memory supply strategy.
What this means for what you build or pay
For most builders, nothing changes today. GPU and memory prices are set by current supply, and a plan for a future fab does not alter them. The report matters as a signal. It suggests the biggest AI hardware buyer sees memory, not compute, as the thing most likely to cap growth, and that the supply chain may reshape around that view. If you plan hardware purchases for 2027 and beyond, assume memory stays a constrained, expensive part of the bill until new capacity actually appears. Treat claims of relief from a single-source report as premature.
Related reading
- Samsung, TSMC and the NVHBM base die
- TSMC CoWoS price rise report
- Samsung second Taylor fab and HBM base dies
- Micron pauses Taoyuan plant
- YMTC fourth Wuhan fab
- Stanford memory price history: DRAM, HBM and NAND
- Nvidia five ways of co-design around HBM
This post reflects reporting dated November 6, 2026. The claim is unconfirmed by TSMC and Nvidia and may change.