JPMorgan maintains TSMC at Overweight and remains bullish on demand for AI advanced packaging and leading-edge processes

Overview

This report focuses on updates to TSMC's advanced back-end packaging, with emphasis on CoWoS, SoIC, CoPoS, CoWoP, and demand changes from AI ASIC, GPU, and CPU customers. JPMorgan believes AI compute demand will continue to lift utilization at TSMC's leading-edge processes and advanced packaging, with N3/N5 utilization above 100% and N3/N2 supply tightness potentially lasting through 2027. The report maintains an Overweight rating on TSMC and sets a Dec-26 target price of NT$2400.0.

Core views

The core views are: first, TSMC's CoWoS and SoIC capacity expectations are raised, and AI accelerators and CPO adoption will be new demand sources; second, NVDA's 2026 CoWoS estimate is slightly cut because Blackwell is stronger and Rubin ramps more slowly, but the 2027 estimate is raised and may still be conservative; third, Broadcom's CoWoS forecast is materially raised, mainly due to higher demand for TPU v7/v8; fourth, Trainium3 lifetime demand is stronger, but part of the ramp is delayed to 2027; fifth, AMD's 2026 CoWoS estimate is slightly lowered while 2027 is unchanged, and a smooth MI450 ramp could create upside; sixth, OSAT participation is increasing, and ASE and Amkor may take on more CoWoS-R or similar capacity.

Analysis framework

The report combines supply chain research, customer program tracking, advanced packaging capacity estimation, AI accelerator shipment forecasting, and valuation multiple analysis. On the supply side, it tracks the availability of TSMC and non-TSMC CoWoS, SoIC, CoPoS, CoWoP, and EMIB; on the demand side, it tracks shipment volumes and technology paths for AI compute programs at NVDA, Broadcom, MediaTek, AWS/Alchip, AMD, Microsoft, and Meta. On valuation, the target price is based on roughly 20x 12-month forward earnings.

Methodology notes

• capacity_forecastCoWoS/SoIC capacity forecast

  Estimate advanced packaging supply and demand using year-end wfpm and customer program mapping

  The report models TSMC CoWoS capacity, non-TSMC CoWoS capacity, and SoIC capacity separately, and adjusts demand assumptions using TPU, Trainium, NVDA GPU, AMD MI450, MTIA/MAIA, and similar programs.

• supply_chain_researchSupply chain research

  Use supply chain checks to judge program ramps, technical bottlenecks, and outsourcing paths

  The report repeatedly cites supply chain research to assess how factors such as HBM4, N3P, 2nm, EMIB, RDL layer count, and CoPoS engineering challenges affect capacity allocation and mass-production timing.

• Valuation methodsForward P/E

  Estimate target price using 12-month forward P/E

  The Dec-26 target price of NT$2400.0 for TSMC is based on about 20x 12-month forward P/E, above TSMC's five-year historical average multiple, reflecting AI demand and fundamental drivers.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• TSMC / 2330.TW

  Core coverage name and advanced packaging supplier

  Strengths

  Leading-edge process and CoWoS/SoIC capacity leadership, high N3/N5 utilization, strong AI compute demand, and a high FY26/FY27 revenue growth outlook.

  Weaknesses

  Client PC and Android demand remain weak, and some advanced packaging new technologies are still in the engineering or pilot stage.

  Comparison

  Compared with non-TSMC CoWoS capacity, TSMC remains the dominant supplier; non-TSMC capacity is expected to be only about 25K/35K wfpm by the end of 2027/2028.

  Risks

  Debate over the sustainability of the AI capex cycle, weak 2H26 PC/smartphone demand, and slower-than-expected advanced packaging ramp.

• NVDA

  An important customer for TSMC CoWoS/SoIC demand and a driver of technology roadmaps

  Strengths

  Blackwell shipments are revised up, the 2027 CoWoS forecast is raised, and the high-end Feynman version may use A16 and 3D SoIC.

  Weaknesses

  Rubin ramps more slowly, and HBM4 availability, N3P wafers, and other bottleneck components may limit upside.

  Comparison

  Compared with other AI accelerator programs, NVDA remains a key source of demand in CoWoS allocation.

  Risks

  HBM4 supply shortages, changes to Rubin/Rubin Ultra architecture and packaging choices, and uncertainty around CoWoP mass-production timing.

• Broadcom / TPU

  TPU v7/v8 demand revisions up drive a higher CoWoS forecast

  Strengths

  TPU shipment outlook is significantly raised, and Broadcom's Pumafish design has relatively lower risk in the TPUv9 competition.

  Weaknesses

  TPUv9 still has uncertainty around project scale allocation.

  Comparison

  The report believes MediaTek's Humufish carries higher risk than Broadcom's Pumafish, because it involves more leading-edge technologies and Intel EMIB execution challenges.

  Risks

  400G Serdes, 3D SoIC, CoWoS-L, and changes in customer demand timing.

• MediaTek / Humufish / Zebrafish

  Demand source for AI ASIC and TPU-related advanced packaging

  Strengths

  TPU v8x Zebrafish still has shipment expectations for 2026/2027, and Humufish is participating in TPUv9 competition.

  Weaknesses

  Zebrafish design is still seeing minor changes, and Humufish involves multiple leading-edge technologies such as high-spec Serdes, Intel EMIB, and 3D SoIC.

  Comparison

  Compared with Broadcom's solution, the report sees MediaTek's TPUv9 project as riskier.

  Risks

  Design issues not fully resolved before 1H26, Intel EMIB execution risk, and multi-die tape-out risk within a short window.

• AWS / Alchip Trainium3

  AI ASIC demand and a source of CoWoS allocation

  Strengths

  Trainium3 lifecycle volume expectation is raised to 4.4M units, driven by Anthropic and AWS internal workloads.

  Weaknesses

  Part of the production ramp is delayed to 2027, and the 2026 CoWoS forecast is lowered.

  Comparison

  The report believes nearly all CoWoS wafers for Trainium3 are handled directly by AWS or Alchip, with no third fabless participant at present.

  Risks

  TSMC reluctance to expand CoWoS-R capacity, and whether ASE's high-RDL-layer CoWoS ramp can succeed.

• AMD

  Demand source for MI450, Venice CPU, and 2.5D packaging

  Strengths

  If MI450 ramps smoothly in 2027, there could be upside, and Venice CPU and data center CPU demand may increase CoWoS consumption.

  Weaknesses

  For 2026 MI450 there are 2nm top-die re-tape-out and HBM4 characterization challenges, so the CoWoS estimate is lowered.

  Comparison

  MI450 relies on the N2 node and may obtain more capacity flexibility than competitors beyond limited N3P capacity.

  Risks

  MI450 ramp delays, HBM4 challenges, and uncertainty over ASE/PTI taking on packaging projects.

• ASE / Amkor / OSATs

  Outsourcing partners for CoWoS-R or similar advanced packaging

  Strengths

  TSMC capacity tightness is pushing greater OSAT participation; ASE may support AMD Venice, NVDA Vera CPU, Trn3 and TPU, while Amkor is involved early in NVDA GB10, Windows ARM CPU, Vera CPU, and Broadcom Tomahawk 6 Switch IC.

  Weaknesses

  Some projects still need qualification, and ASE's high-RDL-layer capability and Amkor's early project mass-production pace still need validation.

  Comparison

  OSAT capacity is a supplement when TSMC CoWoS supply is insufficient, but its scale and maturity remain below TSMC.

  Risks

  Qualification failures, RDL-layer and yield challenges, and a slower-than-expected customer program transition pace.

Key data

• TSMC current priceNT$1950.0Price date is 2026-04-08.
• TSMC target priceNT$2400.0Target price date is Dec-26, implying about 23.1% upside.
• TSMC CoWoS capacity expectation115K/155K/175K wfpmCorresponds to the end of 2026/2027/2028, respectively.
• Non-TSMC CoWoS capacity expectationabout 25K/35K wfpmCorresponds to the end of 2027/2028, respectively, with a slow ramp expected.
• TSMC SoIC capacity expectation35K/65K wfpmCorresponds to the end of 2027/2028, driven by 2nm-class AI ASIC and CPO/Silicon Photonics demand.
• NVDA 2026 CoWoS estimate675K wafers per yearDown 4% from before, but GPU shipments are expected to reach 8.3M units, up 34% YoY.
• Broadcom CoWoS forecast250K/400K in 2026/2027Mainly due to higher unit demand from TPU v7/v8.
• TPU shipment outlook4.3M/6.9M units in 2026/2027The report sees higher external demand for Ironwood and Sunfish.
• MediaTek CoWoS forecast18K/60K wafers in 2026/2027Corresponds to about 0.4M/1.5M TPU v8x Zebrafish units.
• Trainium3 lifecycle volume4.4M unitsDemand is stronger, but part of the production ramp is delayed to 2027.
• AMD 2026 CoWoS estimate77KPreviously 90K, lowered due to MI450 delays and HBM4 characterization challenges.
• TSMC FY26/FY27 U.S. dollar revenue growth outlookabout 35%/30%Driven by N3 growth and strong demand for advanced packaging such as CoWoS/SoIC.

Impact & implications

If the report's view proves correct, TSMC will continue to benefit from the AI compute cycle, tight leading-edge process supply, and advanced packaging capacity expansion. In the near to medium term, profitability may be supported by high utilization, rush wafer demand, ASP premiums, and a weaker Taiwan dollar. At the industry level, OSATs such as ASE and Amkor may gain more outsourcing opportunities; Broadcom TPU, AWS Trainium, and NVDA's future GPU programs will continue to dominate advanced packaging allocation; and if CoPoS and CoWoP make smooth engineering progress, they could reshape the packaging architecture for large AI accelerators after 2028.

Risks

• There is debate over how long the AI capex growth cycle can last.
• Weak PC and smartphone demand in 2H26 could weigh on TSMC.
• Bottlenecks such as HBM4, N3P wafers, the 2nm node, Serdes, and EMIB may constrain AI accelerator ramps.
• CoPoS still faces engineering challenges such as warpage, equipment compatibility, and material selection, leaving mass-production timing uncertain.
• CoWoP has made engineering progress, but its timing for mass production remains uncertain.
• Design changes or delayed ramps at customer programs such as Rubin, MI450, Zebrafish, Humufish, and Trainium3 may affect CoWoS allocation.

What to watch

• TSMC's actual CoWoS and SoIC expansion progress in 2026-2028.
• N3/N5 utilization, the tightness of N2/N3P supply and demand, and rush wafer ASP premiums.
• NVDA Rubin/Rubin Ultra, Feynman, and small-batch CoWoP pilot results.
• Confirmation of demand for Broadcom TPU v7/v8 and TPUv9 Pumafish.
• MediaTek Zebrafish design convergence and Humufish EMIB/3D SoIC execution risk.
• AWS/Alchip Trainium3 volume replenishment in 2027 and progress on ASE outsourcing qualification.
• AMD MI450, Venice CPU, and other 2.5D packaging program ramps.
• Yield, RDL-layer count, and customer onboarding progress for ASE, Amkor, and other OSATs in CoWoS-R or equivalent capacity.