What is 3D DRAM — stacking cells, not chips; due around 2030
3D DRAM is a next-generation memory design that stacks memory cells in many layers inside one chip instead of laying them out on a single plane. It follows the path NAND flash took with 3D NAND, but DRAM is far harder to stack because every cell needs a capacitor to hold charge. It is not the same as 4F² DRAM, which stands the transistor upright to pack one layer more tightly, or HBM, which is a packaging method that stacks finished DRAM chips. Samsung said at a 2024 conference that commercial 3D DRAM would not come before 2030, SK hynix placed it in the 2029-2031 window of its 2025 roadmap, and Micron has given no date. No company mass-produces it. On November 6, 2026, The Information reported that TSMC had begun drafting plans to make 3D DRAM at the request of Nvidia. TSMC responded on November 7 that it remains focused on its core foundry business, in a statement that mentioned neither DRAM nor Nvidia.
The three lines
- Meaning — DRAM with cells laid on their side and stacked in layers inside one chip, adding capacity by layer count instead of line width; distinct from 4F² (denser single layer) and HBM (stacked chips in a package)
- Timing — Samsung: not before 2030 (2024 conference); SK hynix: 2029-2031 (2025 roadmap); Micron: no date. Disclosed research is at 5 and 16 layers, with no mass production
- Trigger — a November 6, 2026 report that TSMC began drafting 3D DRAM plans at Nvidia's request; TSMC said on November 7 it stays focused on foundry. No investment or schedule has been decided
Key questions
- what is 3D DRAM
- **A DRAM structure that stacks memory cells in multiple layers inside a single chip. Capacity grows by adding layers, not by shrinking line widths. No product is in mass production.** | Item | Detail | |---|---| | What changes | A single plane of cells becomes many stacked layers | | Why it is needed | As cells shrink, the capacitor has to get taller and thinner, and scaling runs out of room | | How it is built (research stage) | Alternating silicon and silicon-germanium layers are grown, then etched sideways to form cells on each layer | | Disclosed level | SK hynix 56.1% test yield on 5 layers (reported 2024), Samsung reportedly at 16 layers, imec grew 120 film pairs (2025) | | Layers needed | 32 to 192 (reported SK hynix view), 200 to 300 (2021 trade press estimate) | | Expected timing | Around 2030. No firm production schedule exists |
- 3D DRAM vs 4F2 DRAM vs HBM
- **4F² makes each cell smaller within one layer. 3D DRAM stacks those cells in many layers. HBM stacks finished DRAM chips and bundles them in a package.** | | 4F² DRAM (VCT) | 3D DRAM | HBM | |---|---|---|---| | What is stacked | Nothing. The transistor is stood upright | Cells inside a chip | Finished DRAM chips | | Where it is made | Wafer fabrication | Wafer fabrication | Packaging | | What it buys | About 33% less cell area (theoretical) | Capacity that scales with layer count | Wider data paths and faster transfer | | Status | Prototypes and announced launches | Research | In mass production | | Expected timing | CXMT targets end of 2026; Samsung reportedly 2028 or later | Around 2030 | HBM4 generation shipping |
- is TSMC making 3D DRAM
- **The Information reported that TSMC began drafting 3D DRAM production plans at Nvidia's request. TSMC answered that it stays focused on its foundry business. No investment or timing has been decided.** | Item | Detail | |---|---| | First report | The Information, November 6, 2026, citing two people | | What it said | Draft plans at Nvidia's request; options include a new fab or buying and converting an existing plant | | Person reported to lead it | The vice president in charge of advanced packaging | | TSMC's response | November 7, 2026: focused on its core foundry business. No mention of DRAM or Nvidia | | TrendForce (November 13) | Cites industry sources saying TSMC has no concrete plans to make DRAM itself | | What TSMC already does | HBM base dies and advanced packaging |
3D DRAM is a next-generation memory design that stacks memory cells in many layers inside one chip instead of spreading them across a single plane. Until now, DRAM capacity has grown by shrinking circuit line widths so that more cells fit in the same area. 3D DRAM changes direction. It is the difference between building ever-smaller single-story houses on expensive land and putting up an apartment block. NAND flash made that move in 2013. DRAM has not, and no company mass-produces 3D DRAM.
The term reached general news in November 2026 because of a foundry, not a memory maker. The Information reported on November 6, 2026, that TSMC had begun drafting plans to produce 3D DRAM at the request of Nvidia, its largest customer. On November 7 TSMC said it remains focused on its core foundry business. This article first explains why planar DRAM is running out of room and how 3D DRAM is structured, including how it differs from 3D NAND. It then sets 3D DRAM beside the two things it is most often confused with, 4F² DRAM and HBM, in one table, and collects what Samsung, SK hynix, Micron and Chinese makers have disclosed about timing. It closes with what the TSMC report means, why it matters for South Korea, where the world's two largest DRAM makers are based, and what is still unconfirmed.
1. Why planar DRAM is hitting a wall: the capacitor
A DRAM cell is one transistor and one capacitor. A charged capacitor is a 1 and an empty one is a 0. The transistor is the gate to that capacitor. The structure is simple and it scales badly, because a cell must hold a minimum amount of charge for a 1 to be told apart from a 0, and the charge a capacitor holds depends on its surface area.
Shrink the footprint of a cell and the capacitor's diameter shrinks with it. To hold the same charge it has to be built taller. Semiconductor Engineering describes this as the core constraint on DRAM scaling. A tall, narrow pillar is hard to etch, hard to coat evenly on the inside and hard to keep standing, and yield suffers. The equipment maker ASM likewise writes that the planar shrink path is reaching its limits and that maintaining capacitance while limiting leakage gets harder as geometries shrink.
This is why the industry has talked about the 10-nanometer class for so long. Since DRAM line widths entered the teens of nanometers, each generation has shaved off only a little. Korean makers label these generations 1x, 1y, 1z, 1a, 1b and 1c. Micron calls its recent ones 1-alpha, 1-beta and 1-gamma. At a technology webinar on October 21, 2026, Micron said 1-gamma is in volume production and the next node, 1-delta, ramps in the second half of 2027. Each step buys less than it used to.
| Scaling step | What is done | Where it stalls |
|---|---|---|
| Line-width shrink (10nm-class generations) | The same 6F² cell is drawn a little smaller | The capacitor keeps getting taller and thinner; extreme ultraviolet (EUV) lithography costs rise |
| 4F² cell (vertical channel transistor) | The transistor is stood upright, cutting cell area by about 33% | A one-time gain. The capacitor problem remains |
| 3D DRAM | Cells are stacked in layers | New processes are needed to stack tens to hundreds of layers without defects and to etch them sideways |
2. How 3D DRAM is built, and how it differs from 4F², HBM and 3D NAND
The basic idea of 3D DRAM is to lay the upright cell on its side. Tip the tall capacitor over and each layer becomes thin enough that many layers can be piled up. A Lam Research engineer told Semiconductor Engineering that continued DRAM scaling is driving a move to a 3D geometry built from a stack of horizontal capacitors. He also said that simply rotating the existing cell does not give a manufacturable design. The capacitor cannot be as long sideways as it is currently tall, so it has to become shorter and wider, and every layer needs an etch that cuts in laterally, which is difficult because the depth of each recess varies.
Published materials research shows the direction. In a paper in the Journal of Applied Physics in August 2025, the Belgian research center imec and Ghent University reported growing up to 120 pairs (241 layers) of 65-nanometer silicon and 10-nanometer silicon-germanium on 300 mm wafers. The silicon-germanium layers are sacrificial: they are meant to be selectively removed from the side later in fabrication. The paper notes that stacks of 100 or more pairs are being considered for 3D DRAM. This is a stack of films, not a working memory.
Here are the three technologies that are most often mixed up.
| 4F² DRAM (VCT) | 3D DRAM | HBM | |
|---|---|---|---|
| One-line definition | Stand the transistor upright to pack one layer of cells tighter | Stack the cells themselves in layers inside a chip | Stack finished DRAM chips into one bundle |
| What is stacked | Nothing | Cells | Chips (dies) |
| Where it is made | Wafer fabrication (front end) | Wafer fabrication (front end) | Packaging (back end); chips are joined by through-silicon vias |
| Purpose | Less area at the same line width | More capacity regardless of line width | A wider data path to the processor |
| Cell layout in each chip | One plane | Multiple layers | One plane, as in ordinary DRAM |
| Status | Prototypes, some launches announced | Research | In mass production |
| How they relate | Some see it as the step before 3D DRAM, others plan to skip it | Once it exists, the chips inside HBM gain capacity too | Not a rival to 3D DRAM but something to combine with it |
The 4F² design, covered in our October 13, 2026 explainer, keeps the building at one story and makes each room smaller. 3D DRAM adds stories. HBM is closer to stacking several finished buildings and connecting them with elevators. So when a headline says 3D memory, the first question is whether cells or chips are being stacked. Samsung's zHBM, shown at FMS 2026 in August 2026, and Micron's vHBM both place HBM on top of an accelerator. They are packaging ideas.
The comparison with 3D NAND matters too, because the most common question is why NAND can stack hundreds of layers and DRAM cannot.
| 3D NAND | 3D DRAM | |
|---|---|---|
| Where data is stored | A charge-trap film on the wall of a vertical hole. No capacitor | One capacitor per cell |
| How cells connect | Hundreds of cells hang in series on one vertical string | Each cell needs its own transistor and wiring |
| How layers are made | Insulating films are deposited alternately, then holes are etched through all of them at once | Single-crystal silicon layers are needed, so the stack is grown epitaxially and each layer is etched sideways |
| Layer count | More than 200 in production, products above 400 announced | 5 and 16 in disclosed research; 120 pairs as bare films |
| Commercialized | 2013, Samsung's V-NAND | Not yet |
At the memory conference IMW in May 2024, Samsung explained that stacking volatile DRAM is harder than stacking NAND and requires new materials for the capacitor and the bit line, according to Tom's Hardware, citing ZDNet Korea. Estimates of how many layers it takes to make economic sense run high. SK hynix was reported to have said at the 2024 VLSI symposium that widespread use needs 32 to 192 layers. A 2021 Semiconductor Engineering article relayed an industry view that replacing conventional DRAM economically would take about 200 to 300 layers. The gap between those numbers and the 5 and 16 layers disclosed so far is the time still to go.
3. Who is where: the disclosed roadmaps
| Company | What has been disclosed | Timing | Basis |
|---|---|---|---|
| Samsung Electronics | 4F² vertical channel transistor (VCT) DRAM first, then 3D DRAM; reportedly stacked up to 16 layers | No commercial 3D DRAM before 2030 | IMW 2024 presentation (May 2024); Business Korea report (June 2024) |
| Samsung Electronics | Showed a 16Gb prototype combining 4F² cells and wafer bonding at ISSCC 2026 | Not expected on the market before 2028 | TrendForce (October 12, 2026) |
| SK hynix | 56.1% test yield on 5-layer 3D DRAM; said substantial development remains before commercialization | — | Reports on its VLSI 2024 paper (June 2024) |
| SK hynix | Roadmap applying 4F² vertical gate (VG) and 3D DRAM at 10nm and below; its November 2025 roadmap placed 3D DRAM in the same window as DDR6 | 2029-2031 | VLSI 2025 (June 2025); SK AI Summit (November 3, 2025) |
| Micron | Skips 4F². Extends the 6F² planar cell through 1-delta and later nodes, adds wafer bonding at 1-epsilon, then moves to 3D DRAM | 1-delta in the second half of 2027; no 3D DRAM date | Technology webinar (October 21, 2026) |
| CXMT (China) | Announced server DDR5 modules built on a 4F² cell | 4F² targeted for end of 2026; no 3D DRAM schedule confirmed | TrendForce (October 12, 2026) |
| TSMC | Reported to have begun drafting production plans; the company says it is focused on foundry | No investment or timing decided | The Information (November 6, 2026) |
Two things stand out. First, the most specific numbers are from 2024. The 56.1% on five layers and the 16-layer figure came out of conferences and reports two and a half years ago, and no company has since disclosed how far layer counts or yields have moved. The 56.1% also means that about 561 of roughly 1,000 devices on a single test wafer worked, which is a different thing from a production yield.
Second, the makers disagree on the route. Samsung and SK hynix have laid out a sequence that goes through 4F² to 3D DRAM. Micron has not. Its chief technology and products officer, Scott DeBoer, said at the October 21, 2026 webinar that the density benefit of 4F² over 6F² is "a bit of a myth." His argument is that capacitor pitch, not the transistor, sets bit density, and that 4F² adds process complexity and cost. Micron therefore plans to skip 4F², carry the planar cell for several more generations and go directly to 3D. Which view holds up will be easier to judge once CXMT's 4F² products ship and Samsung's schedule firms up.
4. A foundry making memory: the report, the response and what it means for Korea
A foundry manufactures chips that other companies design, and TSMC has concentrated on logic chips, the processors that do the computing. DRAM is a separate industry in which Samsung, SK hynix and Micron both design and manufacture. By TrendForce's count, second-quarter 2026 DRAM revenue shares were 39.4% for Samsung, 24.9% for SK hynix, 23.3% for Micron and 9.5% for China's CXMT. The report was that TSMC might enter that field.
| Date | What happened |
|---|---|
| August 25, 2026 | Nvidia unveils NVHBM, its own custom HBM specification |
| October 2026 | Reports say Nvidia asked Samsung, SK hynix and Micron to have NVHBM base dies made on TSMC's 3-nanometer process |
| November 6, 2026 | The Information reports that TSMC began drafting 3D DRAM production plans at Nvidia's request |
| November 7, 2026 | TSMC says it remains focused on its core foundry business |
| November 10, 2026 | Korea JoongAng Daily publishes an analysis warning that Korean memory makers could be reduced to subcontractors |
| November 13, 2026 | TrendForce summarizes TSMC's pushback and lists three memory areas where TSMC already has influence |
The Information cited two people. As relayed by several outlets, TSMC is looking at building a new memory plant or buying and converting an existing one, and the work is led by the vice president who runs advanced packaging. Whether to invest, how much and when production might start have not been decided. The background given is that AI demand has left memory in short supply and Nvidia wants more sources.
TSMC's answer was brief. As reported by Kyunghyang Shinmun, citing Bloomberg, a spokesperson said TSMC always focuses on its core semiconductor foundry business and aims to remain a reliable provider of technology and capacity to the logic chip industry. Most outlets reported this as a denial. Others pointed out that the statement names neither DRAM nor Nvidia and does not settle whether an internal study took place. On November 13, 2026, TrendForce cited industry sources as saying TSMC has no concrete plans to manufacture DRAM itself.
The technical barrier is high as well. eeNews Europe noted that conventional DRAM is a different process from TSMC's leading-edge logic, so vertically stacked memory cells would require substantial new process development. TSMC has no DRAM production line. As the previous section showed, the three companies that have made DRAM for decades are themselves talking about 2030 or so.
Yet TSMC is already inside the memory business without making DRAM. TrendForce names three areas.
| Area | Detail |
|---|---|
| Embedded memory | Embedded DRAM (eDRAM) inside logic chips and non-volatile memories such as MRAM and RRAM, which TSMC has offered for years |
| HBM base dies | The chip at the bottom of an HBM stack that links the memory layers to the processor. SK hynix has TSMC make the base die for its HBM4, and Micron has said its NVHBM and HBM4E use foundry-made base dies |
| Memory bonded on logic | Qualcomm is pursuing a structure that bonds low-power DRAM directly on top of a compute die at wafer level, with TSMC reported to handle integration and packaging |
This is where the implications for Korean companies come from. Semiconductors are South Korea's largest export category, and Samsung and SK hynix supply the HBM in Nvidia's AI processors. In its November 10, 2026 analysis, Korea JoongAng Daily framed the issue as less about whether TSMC builds a DRAM plant and more about where control over design and manufacturing ends up. Kwon Seok-joon, a professor at Sungkyunkwan University, listed four ways TSMC could take the most profitable parts of memory without a DRAM fab: making more base dies, leading stacking and integration through hybrid bonding, partnering with Taiwanese DRAM makers, and producing the peripheral circuits that run 3D DRAM. Hybrid bonding is expected to become necessary once HBM stacks reach about 20 layers. A former SK hynix executive warned that if TSMC holds that step too, Korean makers could be left supplying only DRAM cells. Kwon's recommendations were for Korean firms to become co-designers of the base die, secure mass-production hybrid bonding before TSMC does, and widen their custom HBM customer base.
SK hynix's own statement at its Future Forum on September 8, 2026, that it would use logic foundry processes for DRAM peripheral circuits, belongs to the same trend. The larger the share of a memory chip that is logic, the blurrier the line between memory maker and foundry. For now, what is confirmed is one report, a statement about staying focused on foundry, and a base-die collaboration that was already under way. A TSMC 3D DRAM fab has not been decided even as a plan.
5. Frequently asked questions
Will 3D DRAM make HBM unnecessary? No. 3D DRAM raises the capacity of a single chip. HBM is packaging that puts several chips next to a processor with a wide data path. A 3D DRAM chip could itself be stacked into HBM.
Does 3D DRAM have to come after 4F² DRAM? The companies answer differently. Samsung and SK hynix have described 4F² as the earlier step. Micron says it will skip 4F² and go from the planar cell directly to 3D.
Will 3D DRAM bring memory prices down? Not in the near term. The earliest roadmaps point to around 2030, and an estimate attributed to Tom's Hardware puts 3D DRAM at least five years from helping with the shortage.
Is it true it could be made without EUV tools? The idea is that adding capacity through layers reduces dependence on the finest line widths. A 2024 report relayed the view that, if 3D DRAM is achieved, it could yield better products than existing DRAM without EUV. No such product has been made, and the difficulty shifts to deep etching and epitaxial growth equipment.
Is TSMC really going to make DRAM? What is confirmed is a report that it began drafting plans and TSMC's statement that it is focused on its core business. There is no investment decision and no production date. The memory-related work TSMC already does is HBM base dies and packaging.
6. What remains open or unconfirmed
| Item | Confirmed | Not confirmed |
|---|---|---|
| TSMC's plan | The Information reported drafting, citing two people | Whether an internal study exists and how far it has gone; the original report was not opened |
| TSMC's position | Focused on its core foundry business, November 7, 2026 | The original statement text; whether it explicitly rules out a DRAM study |
| Nvidia's requests | Several outlets report both the 3D DRAM request and the 3-nanometer base-die request for NVHBM | Official confirmation from Nvidia or the three memory makers |
| Samsung's schedule | Said in 2024 that commercial 3D DRAM would not come before 2030 | Any change since; its current layer count |
| SK hynix's schedule | The 2029-2031 window in its November 2025 roadmap | A specific year; progress since the 5-layer result of 2024 |
| Micron's schedule | 1-delta in the second half of 2027; wafer bonding at 1-epsilon | When 3D DRAM enters production |
| Chinese makers | CXMT's end-of-2026 target for 4F² | Their 3D DRAM schedule and level |
| Layers needed | Differing estimates of 32 to 192 and 200 to 300 | The layer count at which cost beats planar DRAM |
| Structure | Cells laid horizontally and silicon/silicon-germanium stacks are the leading research direction | The cell structure each maker will finally choose; capacitorless designs are also being studied |
Shrinking the cell within one layer is covered in our explainer on 4F² DRAM (October 13, 2026). Stacked-chip packaging is in our explainers on HBM and on HBM4 versus HBM3E (August 1, 2026). TSMC's business model is in our explainer on foundries (October 15, 2026), and the bottleneck in mounting HBM and processors on one substrate is in our explainer on silicon interposers (October 11, 2026). The legal fight over memory pricing is in our report on the DRAM price-fixing lawsuit in this same November 20, 2026 edition.
Sources
- TrendForce — TSMC Pushes Back on 3D DRAM Rumors: Three Memory Plays Where It Still Has a Say (2026-11-13)
- Seoul Economic Daily (English) — TSMC Weighs 3D DRAM Push at Nvidia's Request (2026-11-07)
- The Next Web — TSMC 3D DRAM memory chips Nvidia (2026-11-06)
- Korea JoongAng Daily — 'No permanent friends': Nvidia's two-track play may reduce Korean chipmakers to subcontractors (2026-11-10)
- Kyunghyang Shinmun (English) — report on TSMC's statement (2026-11-08)
- eeNews Europe — TSMC 3D DRAM Nvidia report response (2026-11-10)
- XenoSpectrum — TSMC 3D DRAM statement not denial (2026-11-08)
- Futurum — Micron Accelerates 3D DRAM From a Fifth Consecutive Node of Process Leadership (2026-10-26)
- NewsTomato (Korean) — Samsung, SK and Micron race on next-generation 3D memory (2026-10-22)
- Korea JoongAng Daily — SK hynix bets on 3-D stacking as next frontier in AI memory (2026-09-09)
- TrendForce — CXMT Accelerates 4F² DRAM Push, Eyes Year-End Server DDR5 RDIMM Launch (2026-10-12)
- Wccftech — SK hynix's Roadmap Positions HBM5/HBM5E, GDDR7-Next, DDR6 & 400+ Layer 4D NAND In 2029-2031 (2025-11-03)
- TweakTown — SK hynix unveils DRAM memory chip roadmap for the next 30 years: 4F2 VG tech, 3D DRAM, and more (2025-06-10)
- TrendForce — SK hynix's 5-layer 3D DRAM Yield Reportedly Hits 56.1% (2024-06-24)
- Tom's Hardware — Samsung reveals 16-layer 3D DRAM plans with VCT DRAM as a stepping stone (2024-05-21)
- Semiconductor Engineering — Baby Steps Toward 3D DRAM (2024-12-17)
- Semiconductor Engineering — Will Monolithic 3D DRAM Happen? (2021-09-09)
- Semiconductor Engineering — Epitaxial growth of up to 120 Si/SiGe bilayers in view of 3D DRAM applications, imec and Ghent University paper (2025-08-05)
- ASM — Scaling DRAM in the 3D era