By TSMCLearning Objectives
- Understand why TSMC dominates advanced semiconductor manufacturing for AI
- Identify the 3nm and 2nm capacity ramps and AI customer mix
- Evaluate what TSMC capacity allocations mean for AI compute supply
What Is TSMC Advanced Foundry?
Taiwan Semiconductor Manufacturing Company (TSMC) is the world's largest contract semiconductor manufacturer — and the dominant fab for cutting-edge AI chips. As of 2026, TSMC controls more than 90% of the advanced-node market required to manufacture frontier AI chips. NVIDIA H100, H200, B200, B300, AMD MI300X, Apple M-series, Apple A-series, Google TPU v5/v6, Broadcom switch silicon, and most other advanced AI chips are all fabricated at TSMC.
TSMC's strategic position as the AI infrastructure supply-chain bottleneck: AI chip designers (NVIDIA, AMD, Google, Apple, Broadcom) compete for TSMC capacity allocation. Demand consistently exceeds supply at the cutting-edge nodes (3nm and 2nm). 70-80% of TSMC's capital budget is now allocated to advanced nodes specifically for AI customer demand. Advanced-chip capacity is booked through 2028.
💡Key Concept
TSMC's strategic position: Foundries don't design chips — they manufacture them. TSMC's customers (NVIDIA, AMD, Apple) are also competitors with each other for cutting-edge wafer slots. When NVIDIA, Apple, AMD, Qualcomm, and Broadcom all want 3nm capacity simultaneously, TSMC chooses who gets how much. That allocation power makes TSMC arguably the most strategically important company in the AI supply chain — more important than any single chip designer, because every chip designer needs TSMC.
✅Tip
Visit TSMC: tsmc.com/english/dedicatedFoundry/technology — wafer pricing through enterprise account managers; capacity allocation negotiated multi-year
Capacity & Customers
TSMC sells wafer capacity to chip designers. Pricing per wafer is not public; customers compete for allocation across multi-year planning windows.
- Up 40 percent year-over-year
- NVIDIA H100/H200/B200
- AMD MI300X
- Apple M-series
- ~100,000 wafers per month by end-2026
- Five 2nm fabs ramping simultaneously
- Apple first adopter
- NVIDIA Feyman GPU rumored
- Most popular node by AI growth
- Aggressive customer adoption
- Capacity booked through 2028
- First broad price increase in years
- Reflects supply tightness
- NVIDIA + Apple absorb cost
- Demand exceeds supply at advanced nodes
- Allocation a strategic lever
- Chip designers compete for slots
The economics: in a tight AI capex cycle, whoever gets TSMC allocation gets to ship the chip. Suppliers without secure allocation watch competitors capture market share.
Process Roadmap
N3 Family (3nm) — Production Workhorse 2025-2026
Currently TSMC's highest-volume advanced node. Monthly output stood at approximately 120,000-130,000 wafers at end of 2025, projected to reach 180,000 wafers by end-2026 — a 40% year-over-year increase driven primarily by AI hardware demand. Major customers: NVIDIA H100/H200/B200, AMD MI300X, Apple M-series, Broadcom switch silicon, Google TPU.
N2 (2nm) — Ramping 2025-2026
Mass production from late 2025; monthly capacity targets ~100,000 wafers by end-2026. Five 2nm fabs are ramping to mass production simultaneously in 2026 — the most aggressive expansion in TSMC's history. Apple is the first adopter for next-generation A-series and M-series chips. NVIDIA's rumored next-generation A16 "Feyman" GPU platform is expected to be its first 2nm-built design.
N2 Revenue Trajectory
By Q3 2026, TSMC's 2nm cumulative revenue is projected to surpass 3nm + 5nm combined — making 2nm the most popular advanced node in the company's history, propelled by AI growth. This is unprecedented adoption speed for a new node.
A16 (1.6nm) and Beyond
TSMC's roadmap continues to A16 (1.6nm) and A14 (1.4nm). High-NA EUV lithography from ASML enables sub-2nm scaling — with TSMC planning to skip High-NA at the initial A14 (1.4nm) node and push standard EUV further with multi-patterning before adopting High-NA later.
AI Capex Allocation
TSMC has explicitly stated that 70-80% of capital expenditure is allocated to advanced nodes specifically driven by AI customer demand. The company is on track for $56 billion+ capex in 2026 with major fab construction in Arizona, Japan, and Taiwan.
Pricing Dynamics
TSMC implemented its first sub-5nm chip price increase in years in 2026 — a 3-5% increase reflecting supply tightness and the company's pricing power vs. fabless customers. NVIDIA, Apple, AMD, and others largely absorb the cost given alternatives are limited.
Strengths
- Unmatched advanced-node share: Over 90% of cutting-edge AI chips fabricated at TSMC
- Capacity expansion: 3nm at 180,000 wafers/month by end-2026; five 2nm fabs ramping simultaneously
- Customer diversification: NVIDIA, Apple, AMD, Google, Broadcom, Qualcomm — most advanced AI chip designers compete for the same capacity
- Operating excellence: Industry-leading yield management, IP protection, and reliability — chip designers prefer TSMC even at premium pricing
- AI-aligned capex: 70-80% of capital budget targeting advanced nodes specifically for AI demand
- Geographic diversification underway: Arizona (US CHIPS Act), Japan, Taiwan — multi-region resilience improving
- Pricing power: First sub-5nm price increase in years (3-5% in 2026) demonstrates supplier leverage
Limitations & Considerations
- Geopolitical concentration in Taiwan: Most TSMC capacity remains in Taiwan, creating systemic risk for global AI supply chain in any conflict scenario
- Capacity booked through 2028: New customers cannot easily get advanced-node allocation; existing relationships dominate
- Pricing power flows to TSMC: Fabless chip designers absorb cost increases; ultimate AI infrastructure cost rises with TSMC pricing
- Geographic expansion slower than headlines suggest: Arizona fabs ramping but still limited share of global TSMC output
- Customer concentration: Apple alone is approximately 25% of TSMC revenue; NVIDIA is the next largest. Concentration creates revenue cyclicality
- Lithography dependence: TSMC depends on ASML for EUV scanners; ASML supply allocation is a separate constraint upstream
Best Use Cases
| Stakeholder | Why TSMC Matters | How They Engage |
|---|---|---|
| AI chip designers | TSMC is the only practical path to advanced-node manufacturing | Multi-year capacity agreements; long-term roadmap planning |
| AI infrastructure operators | Chip availability ultimately depends on TSMC allocation | Track TSMC capacity guidance for compute-supply forecasts |
| National policymakers | TSMC concentration in Taiwan is a strategic risk | CHIPS Act incentives, Arizona fab subsidies, technology export controls |
| Investors | TSMC capex and capacity guidance is the leading indicator of AI compute supply | Track quarterly TSMC earnings for AI cycle signal |
| Frontier AI labs | Compute capacity scales with TSMC manufacturing output | Plan multi-year compute roadmaps around advanced-node ramp |
When to choose alternatives:
- For trailing-edge nodes (28nm and above), GlobalFoundries, UMC, SMIC (in China) provide adequate capacity at lower cost
- For specialty processes (analog, RF, automotive) → GlobalFoundries, TI, ST Micro
- For Chinese-market chips → SMIC is the dominant alternative, though it lacks TSMC's advanced-node capability
- For absolute leading-edge research → Samsung Foundry is the only credible alternative at 3nm and 2nm tiers, with materially smaller capacity
Key Takeaways
- TSMC fabricates over 90 percent of cutting-edge AI chips — NVIDIA H100/H200/B200, AMD MI300X, Apple M-series, Google TPU, Broadcom silicon
- 3nm capacity reaches 180,000 wafers per month by end-2026 (40 percent year-over-year increase); 2nm ramping with five fabs simultaneously to ~100,000 wafers/month by end-2026
- 2nm cumulative revenue projected to surpass 3nm + 5nm combined by Q3 2026 — most popular node by AI growth, with capacity booked through 2028
- 70-80 percent of TSMC capital budget allocated to advanced nodes specifically for AI customer demand; first sub-5nm price increase in years (3-5 percent in 2026) signals supplier leverage
- TSMC capacity allocation is a strategic lever — chip designers compete for slots, and whoever gets allocation gets to ship; track TSMC capex and capacity guidance as leading indicator of global AI compute supply
