Taiwan Semiconductor -- How the Business Works

TSMC is the world dominant semiconductor foundry -- a pure-play contract manufacturer that fabricates chips designed by other companies. Unlike integrated device manufacturers (IDMs) such as Intel or Samsung, TSMC does not design or sell its own chips. This pure-play model eliminates customer conflict and has made TSMC the trusted manufacturing partner for virtually every major chip designer in the world. TSMC commands ~72% of the global foundry market and ~90%+ at advanced nodes below 7nm. At 3nm and below, TSMC is effectively the sole supplier at scale. FY2025 revenue reached $122.4B with HPC/AI representing 55-60% of revenue.

FY2025 Revenue

$122.4B

+36% YoY, accelerating growth

Global Foundry Share

~72%

Record high in Q3 2025, per TrendForce

Advanced Node Share (<7nm)

~90%+

Samsung distant #2 at ~7%

HPC/AI Revenue Share

55-60%

Up from 46% in Q1 2024

Revenue by platform -- HPC/AI is now the dominant driver

Revenue by Platform -- Q4 2025 (% of Revenue)

HPC 55%

Smartphone 32%

IoT/Auto/DCE 13%

Q1 2024 HPC

46%

Pre-AI acceleration

Q2 2025 HPC

60%

Peak HPC share

Q4 2025 HPC

55%

Still dominant platform

AI Accel CAGR

Mid-40s%

Through 2029

Platform mix from TSMC quarterly earnings releases via Daloopa.

Technology node mix -- 3nm tripled in 8 quarters, advanced nodes at 77%

Wafer Revenue by Node -- Q4 2025

3nm 28%

5nm 35%

7nm 14%

Mature (>7nm) 23%

Q1 2024 3nm

Early ramp

Q4 2024 3nm

26%

Nearly 3x in 3 qtrs

Q4 2025 3nm

28%

3x from Q1 2024

Advanced (<=7nm)

77%

Up from 65% in Q1 2024

Node mix from TSMC quarterly earnings releases via Daloopa.

Pure-play foundry model -- why it works

No Customer Conflict -- TSMC Does Not Compete With Its Customers

The pure-play foundry model is the foundation of TSMC moat. Unlike Samsung (which sells its own Exynos chips and competes with customers like Qualcomm) or Intel (which designs its own processors), TSMC has zero customer conflict. This trust enables TSMC to receive chip designs containing the most sensitive intellectual property from every major tech company -- including fierce competitors like NVIDIA and AMD, or Apple and Qualcomm -- simultaneously. No other foundry can offer this neutrality at scale.

Customer portfolio -- every major AI chip designer is a TSMC customer

Largest Customer

NVIDIA ~19%

Overtook Apple in 2025 (~$23.4B est.)

Second Largest

Apple ~17%

~$20.5B estimated revenue

Top 2 Concentration

~36%

Concentrated but diversified AI + consumer

Total Customers

500+

AMD, Broadcom, Qualcomm, Google, Amazon...

A major structural shift occurred in 2025: NVIDIA overtook Apple as TSMC largest customer, with estimated ~19% of revenue (~$23.4B) versus Apple at ~17% (~$20.5B). This reflects the explosive AI/data center demand that now defines the TSMC cycle. Rising ASIC demand from hyperscalers (Google TPU, Amazon Trainium, Meta custom silicon) provides further diversification within HPC. C.C. Wei explicitly named Apple, NVIDIA, AMD, Qualcomm, and Broadcom as key U.S. customers driving Arizona expansion.

CoWoS advanced packaging -- the critical AI bottleneck

CoWoS Capacity Nearly Quadrupling by Late 2026

CoWoS (Chip-on-Wafer-on-Substrate) is TSMC advanced packaging technology that integrates multiple chiplets and HBM memory onto a single interposer. It is essential for AI accelerators like NVIDIA H100/B100/B200 and AMD MI300X. CoWoS remains the critical capacity constraint for AI chip production. TSMC doubled capacity in 2025 and plans to reach ~130,000 wafers/month by late 2026 (~4x late 2024 levels). NVIDIA has reportedly booked over half of CoWoS capacity for 2026-27. Two advanced packaging fabs planned for Arizona (AP1, AP2) beginning construction in 2026.

CoWoS Capacity Target

~130K wpm

By late 2026, ~4x late 2024

NVIDIA CoWoS Booking

>50%

Of 2026-27 CoWoS capacity booked

Arizona Packaging Fabs

AP1 + AP2

Construction starting 2026

Arizona expansion -- $165B commitment, from liability to strategic asset

TSMC has committed $165B to build a giga fab cluster in Arizona, with long-term projections toward a $465B 12-fab cluster. The strategic rationale is geographic diversification of advanced manufacturing:

Fab 1: Volume production achieved with yield comparable to Taiwan fabs
30% of N2+ capacity planned for Arizona, providing meaningful geographic diversification
Premium pricing: Customers accept higher prices for geographic flexibility
Margin dilution: Starts at 2-3% of gross margin, narrowing from 2-3% to closer to 2% in H2 2025 -- ahead of schedule. Expected to widen to 3-4% in later years as more fabs come online
AP7 advanced packaging fab targeting 2026 output; P6 fab eyed for US packaging hub

C.C. Wei: "Yes, we did charge a little bit higher... bearing USA is a premium product."

Quality gate -- PASS (3/3)

Criterion	Assessment
Oligopoly/Monopoly?	YES -- Undisputed foundry monopoly. ~72% global share, ~90%+ at advanced nodes. Samsung distant #2 at ~7%. No other competitor at 3nm or below at scale.
FCF Generation?	YES -- FY2025 FCF of NT$1,003B (~$31B USD), up 15% YoY despite massive $40.9B capex. FCF positive every year even during peak investment cycles.
Management Track Record?	YES -- Revenue 20% CAGR since 2021, N3 yield ramp ahead of schedule, Arizona Fab 1 at Taiwan-comparable yield, revenue exceeded guidance midpoint in 7 of 8 quarters.

Data sourced from Daloopa, earnings transcripts, and TrendForce. All financials in USD unless noted as NTD.