台积电否认3D DRAM计划,英伟达推动3nm HBM底片合作
台积电否认3D DRAM计划,英伟达推动3nm HBM底片合作
Conclusion: TSMC is reportedly drafting plans to make 3D DRAM at Nvidia's request, but this is an early, unconfirmed option. TSMC has denied it, saying it "focuses on our core business at all times, which is dedicated IC foundry" and plans to remain a logic foundry. The better-documented development is narrower: Nvidia has asked memory makers to build the base dies for its custom HBM on TSMC's 3nm process. That makes TSMC more central to HBM without TSMC making memory itself. If TSMC did enter DRAM, it would pull capacity and capital away from logic and packaging, which are currently its tightest constraints. Treat the DRAM idea as a low-probability, high-impact scenario, not a planned investment.
Data as of 2026-11-08. The newest reports are from 6–7 November 2026. I did not retrieve TSMC's or Nvidia's own filings, and I did not pull market quotes because the question is about strategy rather than prices.
What is known
- The DRAM report. The Information reported that TSMC has begun drafting plans for 3D DRAM at Nvidia's request, and that the effort is led by Hsu Kuo-chin, a TSMC vice president who oversees advanced packaging. Options reportedly include building new fabs or buying and converting an existing plant. Reuters, citing The Information, is the source for the buy-or-build detail. TheNextWeb, 6 Nov 2026; Tom's Hardware, 7 Nov 2026
- The denial. TSMC's spokesperson's statement, as quoted by Tom's Hardware, reaffirms its logic-foundry focus. Tom's Hardware also reports that a DRAM entry could come by acquiring a small specialty memory maker, but that TSMC would still need billions of dollars and years of R&D to reach leading-edge DRAM. Tom's Hardware, 7 Nov 2026
- The HBM base-die request. Reuters, citing the Seoul Economic Daily (18 Oct 2026), reports that Nvidia asked Samsung, SK Hynix and Micron to use TSMC's process for the base dies of its custom HBM (NVHBM), and sixth-generation DRAM for the core dies. Samsung reportedly accepted and plans to make base dies on TSMC's 3nm process, and is assessing whether its own foundry could do so if TSMC lacks capacity. Reuters via GoldSea, 18 Oct 2026; World Online, 18 Oct 2026
- Earlier signals. TrendForce, citing Commercial Times (Aug 2025), reported that Nvidia was developing its own HBM base die on 3nm, with small-batch trial production planned for the second half of 2027. These were rumors at the time. TrendForce, 18 Aug 2025
- Market backdrop. Micron expects the memory shortage to last into 2028, according to TheNextWeb's summary. Reuters (6 Nov 2026) reports that AI demand is outrunning advanced packaging capacity. TheNextWeb; Reuters, 6 Nov 2026
Why Nvidia might want this
- Memory is now a supply constraint, not just a cost. Reports say memory shortages can decide whether a product ships at all. A supplier that can make DRAM or HBM components under Nvidia's design would reduce its dependence on three suppliers that are already sold out or pricing up.
- Control of HBM co-design. Nvidia wants a custom controller inside the HBM base die, which is why it is pushing the logic part onto TSMC's advanced nodes. Putting the logic and memory processes closer together under one foundry would tighten that integration.
- Leverage over memory makers. Nvidia's request forces Samsung, SK Hynix and Micron to use TSMC's logic process, which strengthens Nvidia's position in negotiations. The DRAM idea would add a second, more direct option if the memory makers do not deliver.
- Speed and qualification. Samsung's reported decision suggests memory suppliers are accepting TSMC integration to win Nvidia qualification. For Nvidia, TSMC is the common manufacturing partner it already trusts.
What it could mean for TSMC's manufacturing priorities
- Packaging and wafers are already tight. Reports put Nvidia's 2026 CoWoS allocation at roughly 800,000–850,000 wafers, more than half of TSMC's capacity, and TSMC's monthly CoWoS capacity is projected to rise from about 70,000 wafers in 2025 to 130,000–140,000 by end-2026, with a supply-demand gap still near 20%. TSMC is also outsourcing more CoW steps to OSATs such as ASE. TrendForce, 5 Aug 2026; Silicon Analysts, 7 Sep 2026. Morgan Stanley's field research projected 125,000 wafers per month in 2026. Estimates vary across these sources, so treat the exact figures as reported ranges.
- Cleanroom space is the binding constraint. Morgan Stanley's note says TSMC management has said there is not enough room for new 3nm capacity because new cleanroom space is going to 2nm. A DRAM line would compete for that same space, so it would likely displace logic or packaging capacity, or require new fabs. Boardor summary of Morgan Stanley field research, 3 Aug 2026
- Capital intensity. Morgan Stanley estimates TSMC's 2026 capex could rise to $48–50 billion if 3nm expansion and fab relocations proceed. A commodity DRAM business would add to that load and bring cyclical risk, since memory shortages usually lead to oversupply and price collapses. Boardor summary, 3 Aug 2026; Tom's Hardware, 7 Nov 2026
- Customer conflict. SK Hynix already partners with TSMC on HBM base dies. A TSMC DRAM business could compete with its largest memory partners, and the denial reads partly as a signal to them.
My assessment: TSMC's priority is to protect its logic and packaging franchise, which is more profitable and less cyclical than DRAM. Most likely, it will serve Nvidia through base dies, foundry, and packaging, and it will keep DRAM plans as a contingency. A DRAM build-out would only make sense if memory shortages persist and Nvidia offers firm volume commitments.
Implications for the broader AI chip supply chain
- Memory makers keep leverage but share the stack. Samsung, SK Hynix and Micron may keep the DRAM cores, but TSMC is becoming the logic layer for their HBM. This shifts value toward TSMC even if memory makers keep their current share of bits.
- OSATs gain share. TSMC's outsourcing of CoW and the rise of ASE and SPIL capacity suggest that packaging will be spread across more firms. Digitimes reports Morgan Stanley's view that AMD's CoWoS share could rise in 2027 at Nvidia's expense. Digitimes, 26 Aug 2026
- Custom silicon competition. Custom HBM and base dies are part of a broader move toward custom chips. Google's TPU and Broadcom's CoWoS allocation are also growing, which tightens TSMC capacity further. Boardor summary, 3 Aug 2026
- Concentration risk. If TSMC takes on DRAM and HBM logic as well as packaging, one manufacturer would hold more of the AI supply chain. That raises geopolitical and single-supplier risk for Nvidia and its customers.
What would change this view
- Confirmation from TSMC or Nvidia, or a named acquisition target or fab conversion, would move the DRAM idea from rumor to a strategic decision.
- Evidence that TSMC's 2027 cleanroom and CoWoS expansion are on track would reduce the case for DRAM, since logic and packaging would have enough room.
- A memory price collapse would undermine the economics of any DRAM entry.
- Confirmation that Samsung, SK Hynix and Micron are qualified on TSMC's 3nm base-die process would indicate that the HBM integration plan is working without a TSMC DRAM business.
Sources
- TheNextWeb, "TSMC is weighing a move into AI memory chips at Nvidia's request," 6 Nov 2026. https://thenextweb.com/news/tsmc-3d-dram-memory-chips-nvidia
- Tom's Hardware, "TSMC denies exploring 3D DRAM at Nvidia's request," 7 Nov 2026. https://www.tomshardware.com/pc-components/dram/tsmc-denies-exploring-3d-dram-at-nvidias-request-advanced-memory-tech-is-still-years-from-market-viability
- Reuters (via GoldSea), "Samsung to use TSMC technology for Nvidia's next-generation memory chips," 18 Oct 2026. https://goldsea.com/article_details/samsung-to-use-tsmc-technology-for-nvidias-next-generation-memory-chips
- World Online (Seoul Economic Daily report), 18 Oct 2026. https://www.wol.com/2026/10/18/samsung-use-tsmc-technology-nvidias-next-generation-memory-chips-seoul-newspaper-reports-6469576/
- Chip Dispatch (citing GuruFocus), 19 Oct 2026. https://chipdispatch.com/articles/samsung-reportedly-turns-to-tsmc-to-build-nvidia-s-next-ai-memory-d02e43f6
- Reuters, "AI boom turns chip packaging into next battleground," 6 Nov 2026. https://www.reuters.com/5f374b32b07b/world/asia-pacific/ai-boom-turns-chip-packaging-into-next-battleground-semiconductors-2026-11-06/
- TrendForce, "TSMC reportedly expands outsourcing of CoWoS front-end step," 5 Aug 2026. https://www.trendforce.com/news/2026/08/05/news-tsmc-reportedly-expands-outsourcing-of-key-cowos-front-end-step-to-osats-amid-rising-nvidia-asic-demand/
- Silicon Analysts, CoWoS and HBM newsletter, 7 Sep 2026. https://siliconanalysts.com/newsletter/qual-watch/2026-09-07
- Boardor (summary of Morgan Stanley AI supply chain field research), 3 Aug 2026. https://boardor.com/blog/morgan-stanley-ai-supply-chain-field-research
- Digitimes, "TSMC packaging shift could boost AMD CoWoS share," 26 Aug 2026. https://www.digitimes.com/news/a20260826PD243/amd-cowos-tsmc-nvidia-demand.html
- Domino Theory, "Taiwanese packaging ecosystem adjusts to AI boom," 3 Sep 2026. https://dominotheory.com/taiwanese-packaging-ecosystem-adjusts-to-ai-boom/
- TrendForce, "Nvidia reportedly eyes small-scale HBM base die production in 2027," 18 Aug 2025. https://www.trendforce.com/news/2025/08/18/news-nvidia-reportedly-eyes-small-scale-hbm-base-die-production-in-2027-rattling-memory-chip-markets/