AI summary by TwinMind
The Superior Economics of TSMC’s Giga-Fabs
Asianometry
17m 22s
TSMC Fab 18 Overview
- 2018: TSMC broke ground on Fab 18 near Tainan City, southern Taiwan
- Massive scale: 103 acres, 950,000 square meters total floor space
- 3x size of AT&T Stadium (Dallas Cowboys)
- Total cost: nearly $20 billion across all phases
- More expensive than USS Gerald R. Ford aircraft carrier
- Focus: understanding why TSMC's fabs getting bigger/more expensive makes economic sense
Fab Classification System
- Mini fabs: 10,000-30,000 wafer starts/month
- Example: Fab 21 (Arizona) - $12 billion, 5nm, 1,100 acres
- Phase 1 target: 20,000 wafers by Q1 2024
- Mega fabs: 30,000-100,000 wafer starts/month
- Gigafabs: 100,000+ wafer starts/month (TSMC trademarked term)
Different Fab Types & Output Requirements
- Logic fabs: target 80,000 wafers/month (microprocessors)
- Memory fabs: exceed 120,000 wafers/month (DRAM, NAND)
- Samsung, Yangtze Memory, SK Hynix
- More generalized product → brutal overproduction cycles
- Specialty fabs (smaller volumes):
- Analog chips (physical-to-digital signal conversion)
- CMOS image sensors (digital cameras)
- MEMS (micro-electromechanical systems)
- TFT LCD panels
Fab Scaling Philosophy Evolution
- Intel's "Copy Exactly": technology transfer without modifications
- TSMC's early "Copy Exact": similar to Intel approach
- Current "Fab Matching": allows flexibility
- Process engineers monitor all lines across all fabs
- Successful tweaks propagated to other production lines
- Continuous optimization across facilities
Rising Fab Costs Over Time
- 1983: UMC spent $200M for 1.2 micron fab
- 1999: UMC spent $1.8B for 0.35 micron fab
- 2007: 65nm fab cost $10B
- 2010: TSMC invested $9.4B total across 4 phases for 28nm Gigafab (Fab 15, Taichung)
- 2018: 5nm Fab 18 costs nearly double Fab 15
Equipment Cost Drivers
- Photolithography equipment: 70% of total manufacturing costs, 20% of fab's entire budget
- Cutting-edge systems: EUV and high-NA EUV pushing $300M per system
- Physical limitations: driving exponential price increases
- Size/complexity: requiring larger, more expensive clean rooms
Clean Room Requirements & Costs
- Purity standards: Class 1 to Class 10 and beyond
- Size evolution:
- Fab 14: 31,000 sq meters (5.5 soccer fields)
- Fab 15: 104,000 sq meters (14.5 soccer fields)
- Fab 18: 160,000 sq meters (22.4 soccer fields)
- Early EUV challenges: lacked pellicle (1 micrometer polymer film)
- Ultra-strict standards: <1 particle over 52nm wide per week
- ASML eventually developed suitable pellicle
Engineering Challenges
- Earthquake/vibration issues: Taiwan location near high-speed rail
- Clean rooms elevated on dampeners
- EUV infrastructure: CO2 lasers generate EUV light
- Machinery infrastructure must go underneath clean room floor
- Additional complexity: from vibration prevention systems
TSMC Financial Context (2020)
- Revenue: $45 billion
- Cash on books: $23 billion
- Four gigafabs costing billions to run
- Fab 18 cost: ~25% of annual revenue
TSMC's Strategic Approach
- Priority hierarchy: speed > quality > cost ("pick two" philosophy)
- Not low-cost leader: Samsung frequently offers lower prices
- Value-focused: customers have good margins, can pay premium for on-time delivery
Speed-First Examples
- Chinese New Year: paid extra $140/day per worker during holiday
- Nighthawk Army: overnight R&D engineers
- 30% higher salaries, 50% higher bonuses
- "Liver buster" shifts (cultural reference to overnight work)
Customer Service Innovations
- 1997: Morris Chang pushed service-oriented enterprise transformation
- E-foundry initiative: better information flow management
- Customers can track order status in development cycle
- Cyber Shuttle Service: cost-sharing for prototyping
- New Tape Out (NTO) masks cost millions
- Multiple customers share single wafer costs
Cycle Time Optimization
- Critical metric: Days Per Mask Layer (DPML) - lower is better
- Layer count evolution:
- Late 1990s 180nm: 25 layers, 2 days/layer = 2 months total
- N14/N16: 60 layers
- N10: 78 layers
- N7: 87 layers
- N5: 115 layers (reduced to 81 with EUV replacing some steps)
- Without improvements: N5 wafer would take 8 months (unacceptable)
Gigafab Scale Benefits
- Redundancy advantage: 4 tools vs 1 reduces downtime risk
- Equipment reliability: ion implanters down 30-40% of cycle
- Automatic routing: wafer batches rerouted when machines down
- 2006 simulations: 8x fab size can reduce cycle times by 50%
- Sophisticated planning: queue processes ensure optimal wafer batch timing
Customer Base Advantage
- Premium customers: Apple, AMD, Qualcomm, NVIDIA, MediaTek
- Unique combination: afford higher prices + buy high volume
- Apple relationship: speculation about upfront facility funding
- No confirmed evidence of direct funding
- Apple's buying power enables better demand forecasting
- Capacity justification: reliable lead customers justify gigafab investments
- Customer retention: prevents profitable business migration to competitors
Competitive Positioning
- Smaller foundries: GlobalFoundries stuck with lower-profit customers
- Lacking scale, makes negative gross margins
- "Apple of foundry business": captures most industry profits
- Aggressive cannibalization strategy: "eat their own children"
Node Process Strategy
- Three tiers: leading edge, medium, mature processes
- Development cycle:
- Spend billions on leading edge for premium customers (Apple)
- High prices but low yields, high depreciation
- Rapidly apply benefits to medium processes
- Explicit goal: cannibalize medium node users
N7 to N6 Transition Example
- 2018: N7 process debut
- 2019: N6 evolutionary step introduced
- Rapid replacement: N6 designed to quickly replace N7
- Market penetration: 0% to 50% in single year
- Profitability: N6 suspected very profitable for TSMC
Information Security
- Fab secrecy: very secretive about fab operations
- Open elsewhere: transparent about products, customer service, financials
- Research trail: paper trail goes cold around 2010
- Contrast with ASML: European company very open about engineering research
2009 SMIC Lawsuit Impact
- Trade secret theft: TSMC won lawsuit against Shanghai-based SMIC
- Stolen recipes: SMIC quickly gained China market share using secrets
- Security response: TSMC likely shut down content about fab operations post-lawsuit
- Access difficulty: makes understanding their operations much harder
