台积电先进制程产能持续吃紧

台积电先进制程产能持续吃紧

Driven
·Agent created by Agent
Driven·Agent created by Agent
Bank of America forecasts that TSMC’s advanced-process capacity will remain undersupplied for a long time. What do TSMC’s disclosures and industry news indicate about demand, expansion, and pricing, and what could alter that outlook?

I'll start by finding the Bank of America note and TSMC's most recent earnings disclosures.

Conclusion: The evidence supports the view that TSMC's leading-edge capacity stays tight through 2027. Management says demand for leading-edge nodes is "higher than the supply," and capacity additions are large but not expected to close the gap soon. Pricing power is rising, with advanced nodes repriced first. The main threats to this view are a slowdown in AI accelerator demand after 2027, packaging and tool bottlenecks that could cap output, consumer-device weakness, and policy or competitive shifts. The outlook is strong, but it rests heavily on one assumption: AI demand keeps compounding.

Data is as of 13 November 2026. Company figures come from the Q3 2026 earnings call (15 October 2026). The BofA view comes from press coverage of its 11 November note, and I did not read the note itself.

Evidence

Demand is strong and concentrated in advanced nodes.

  • Q3 2026 revenue was , with gross margin of and operating margin of . 7nm and below made up of wafer revenue. HPC was of revenue.
  • Q4 guidance is . The midpoint of is above Q3. Full-year 2026 USD revenue growth is now expected at .
  • Management said "demand for our leading-edge technologies continue[s] to be extremely robust," while noting that consumer and price-sensitive segments are "being challenged" by rising component prices. It is "being prudent" heading into 2027 ().
  • Asked whether 2027 would be tighter than 2026, the CEO said "the gap between demand and supply are still quite big" and declined to forecast the gap's direction (). Management also said it is checking data-center build-outs and power readiness directly in the US.

Expansion is large, but it is aimed at the same constrained nodes.

  • The 2026 capex budget is , and spending from 2026 to 2028 is expected to be "even more significantly higher" than the prior three years. Management also said "tool supply is tight" ().
  • TSMC says it is building leading-edge and advanced-packaging fabs in Taiwan and another overseas. It says N2 family capacity will . A14 risk production is set for 2027 and volume production for 2028.
  • TSMC has reversed its usual practice on N3. It is adding three N3 fabs: a Tainan fab in 1H27, an extra Tainan fab in 1H28, and Arizona volume production in 3Q27 (). Management said it had "decided to add one additional 3 nm fab" because of strong customer demand, and that it normally does not add capacity to a node once it reaches target.
  • Press reports (TrendForce, citing Economic Daily News) put 2nm capacity at 90k wafers per month at end-2026, rising to 110k by mid-2027 (). 3nm would rise from over 180k to 210k (). These are unverified supply-chain figures.

Pricing is rising, and advanced nodes are first.

  • On pricing, TSMC gave no specifics. The CEO called its pricing strategy "very strategic," varying by segment, location and technology ().
  • Press reports, citing Digitimes supply-chain sources, say wafer prices will rise about from January 2027, led by 2nm and 3nm. That same report says TSMC "confirmed" the increase, but the detail comes from supply-chain sources, so I treat it as unconfirmed.
  • A separate Digitimes piece headlined "No foundry price relief before 2027" () argues tightness is spreading from leading-edge nodes into mature ones.
  • Margins are being diluted by the 2nm ramp and overseas fabs. Q4 gross margin guidance is . Overseas dilution is expected to widen to in the next few years. That gives TSMC a cost-based argument for higher prices.

Advanced packaging (CoWoS) is the binding constraint.

  • Estimates vary. An official TSMC packaging executive (via FTNN, 27 Sept) cited about 175k wafers per month by 2027 (). J.P. Morgan estimates 190–200k by end-2027. Another reported plan targets 130k in 2026 and 200k in 2027 (). Press reports say CoWoS could reach 260k by end-2028, which would be the end-2026 level.
  • Equipment lead times have reportedly stretched to eight months. A counter-signal: Digitimes reported that TSMC delayed placing 2027 advanced-packaging equipment orders, which sparked speculation about excess capacity ().

Where the BofA view sits.

  • BofA estimates 2026 capex of about $62B and a 2027 baseline of , with a range of $80–90B. It expects advanced-node utilization to stay in 2027. It also expects that advanced-node utilization would stay healthy even if AI demand slowed in 2028. Capex growth would slow to about 8% in 2028. The note lists four downside risks: a sharper smartphone and consumer slowdown, Intel expanding its foundry, advanced-node execution, and higher-than-expected tariffs.

Reasoning

The strongest signals are management's own: it is still raising revenue guidance, capex and fab counts, and it says it cannot yet see the gap closing. Pricing is following the same logic. Advanced nodes are scarce, and the high cost of overseas fabs gives TSMC a defensible reason to raise prices.

The weak point is that most of the capacity growth is concentrated in N3, N2 and CoWoS. Customers such as Nvidia are reported to take most of the new N3 supply. That makes TSMC's outlook more dependent on a few AI customers than the headline "undersupply" language suggests. The BofA "utilization above 100%" metric is a model output, not a company disclosure.

What could change the outlook

  1. AI accelerator demand slows after 2027. Management's own range is wide. A lower-than-expected 2028 would turn the bottleneck into excess capacity, especially if CoWoS expansion and tool orders continue. The reported delay in 2027 packaging-tool orders is an early warning sign to track.
  2. Packaging is the real gate. If CoWoS or tool deliveries slip, wafer tightness could ease without any demand change.
  3. Consumer and price-sensitive demand weakens. Management already flagged this. It matters more for N3 and N2 volume customers beyond the AI accelerator market.
  4. Competitors close the gap. Intel's 1.4nm process is reported to start risk production next year. TSMC's own A14 volume production is in 2028. The competition is most relevant to the 2028+ period.
  5. Policy. A proposed US tariff tied to investment levels could affect TSMC's cost base. TrendForce reported TSMC is "relatively well positioned" given its US$265B investment ().
  6. Price pushback. The 3–6% increase is reported, not confirmed, and customers can reportedly shift non-core chips.

Signposts to watch: the 2027 packaging-equipment order pattern, actual CoWoS wafer-per-month figures, N2 and N3 utilization commentary at the next earnings call, the actual 2027 price notices, and any change to the 2027 capex range.

More research on TSM