台积电进军HBM基础芯片的战略评估
台积电进军HBM基础芯片的战略评估
Conclusion: The most plausible fit is TSMC making HBM base dies and related packaging work, not mainstream DRAM. I could not verify the report that TSMC is considering advanced memory chips at Nvidia's request, so treat it as an unconfirmed option. Even on the generous reading, it is a strategic option rather than a growth line to model yet. TSMC's core leading-edge business is growing much faster, and the margin, capex, and competitive questions below are unresolved.
Data as of: TSMC's Q3 2026 results (Oct 15, 2026) and news items through Nov 4, 2026. Market context is the Nov 6, 2026 US close (Friday).
Evidence and reasoning
-
The report itself is unverified. My searches surfaced adjacent reporting, not a primary report or a TSMC statement. Digitimes (Sep 2, 2026) reported that Nvidia is pushing into HBM base dies as competition in HBM moves below the DRAM stack (Digitimes). I only retrieved the lead paragraph, so I cannot confirm details.
-
Base dies fit TSMC better than DRAM does. A base die is a logic chip, which is TSMC's core competence. Making DRAM cells is a different business, and the incumbents (SK Hynix, Samsung, Micron) hold the process know-how and supply contracts. This distinction is my analysis, not a sourced claim.
-
TSMC's existing business is strong and the core is still the growth engine. Q3 2026 revenue was , or , up year over year. Gross margin was , and advanced nodes (7nm and below) were of wafer revenue. Q4 guidance is US$52.3 billion to US$54.3 billion, with gross margin of . Against that base, a memory-adjacent line would need to be very large to move group growth in the near term. I did not find any source giving memory-related revenue or content value, so I have not estimated one.
-
Capex is the funding question. CNBC reported that TSMC expects 2026–2028 capex to be significantly higher than in the previous three years to support customer demand (CNBC). I could not verify the exact capex figure. Any memory-related capacity would compete with leading-edge wafers and advanced packaging for the same constrained capacity. Customer prepayments or co-investment from Nvidia would be a reasonable structure to watch for, but that is speculation.
-
Why Nvidia would ask. Nvidia gains supply diversification and bargaining leverage against HBM suppliers, plus tighter integration of the base die with its GPUs. This is my inference, not sourced.
-
The memory market is tight, but incumbents have locked in supply. SK Hynix says customers want more than twice the HBM volume it can supply (Digitimes). Samsung is reportedly entering 2027 with of planned memory production under long-term contracts. That leaves a narrow window for a new entrant. Samsung and SK Hynix target 2Q27 volume shipments of 8-layer HBM4 to Nvidia, and Nvidia is reportedly trimming HBM4 bandwidth per GPU for Rubin. These spec changes can alter base-die content.
-
Competitive threats. Samsung is reportedly planning HBM base dies for AMD and Nvidia at its Taylor, Texas plant, though Samsung has not confirmed either customer (Digitimes). High-bandwidth flash could cost roughly 90% less per GB than HBM4 but may consume far more wafer capacity than conventional NAND, which is another capacity competitor.
What investors should track before assuming meaningful growth
- Official scope from TSMC: whether this means base dies, packaging, or DRAM, plus capacity, timing, and whether any customer commitment exists.
- Pricing model: foundry-style pricing would be margin-accretive; commodity-memory economics would dilute the 66% gross margin.
- Capex guidance changes: how much is allocated to this versus leading-edge logic and CoWoS.
- Yield and qualification: SK Hynix's HBM4 yield is reported in the mid-60% range, a benchmark a new entrant would need to match.
- Nvidia's spec path: Rubin bandwidth changes and the 8-layer shift could reduce or change base-die content.
- Samsung and SK Hynix base-die progress: whether Nvidia actually dual-sources and how much volume TSMC could win.
- The next quarterly results and call: whether management quantifies any memory contribution.
Bottom line: This is a credible strategic option that fits TSMC's logic and packaging strengths better than DRAM does. It does not yet justify raising growth estimates. Wait for official confirmation of scope, customer commitments, and capex allocation before counting it as meaningful revenue.
Sources
- TSMC, "TSMC Reports Third Quarter EPS of NT$29.43," Oct 15, 2026 — https://pr.tsmc.com/english/news/3346 (primary)
- CNBC, "TSMC posts record profit for a sixth straight quarter," Oct 15, 2026 — https://www.cnbc.com/2026/10/15/tsmc-third-quarter-profit-global-ai-.html
- Digitimes, "Nvidia pushes into HBM base dies as NVHBM battle heats up," Sep 2, 2026 — https://www.digitimes.com/news/a20260901PD207/hbm-competition-dram-samsung-bandwidth.html
- Digitimes HBM4 topic page (headlines through Nov 4, 2026) — https://www.digitimes.com/tag/hbm4/00113922.html
- Driven Market Data, TSM and NVDA quotes, Nov 6, 2026 close (retrieved for context; not used in the valuation conclusion)