AI summary by TwinMind
TSMC and AWS: Comparing Two Infrastructure Giants
Asianometry
15m 2s
Company Introductions and Business Models
- Both AWS and TSMC serve as platform enablers for other companies' success
- AWS: Cloud computing division of Amazon, dominant western cloud platform
- 33% market share, competing with Google and Microsoft
- $35 billion revenue in 2019
- B2B company serving other businesses, not end users directly
- Contributes over half of Amazon's total profit
- When AWS broke out as individual line item, caused Amazon market cap to rise $25 billion to $207 billion
- TSMC: Taiwanese state-backed semiconductor manufacturer
- Pure-play foundry model focused on manufacturing only
- Commitment not to compete with customers
- Spawned generation of "fabulous" semiconductor companies
Historical Evolution and Founding Stories
- Amazon evolution: Started as bookstore → expanded to small vendor services → "picks and shovels" approach across multiple industries
- Fulfilled by Amazon program (logistics)
- Amazon Web Services (internet services)
- Strategy embeds Amazon into both buyer and merchant ecosystems
- TSMC founding by Morris Chang:
- Part of Taiwan government semiconductor development initiative
- Chang worked decades at Texas Instruments in US
- Taiwan's 3-micrometer process was two generations behind leading edge
- UMC earlier attempt failed - could only get chips into toys
- Pure-play foundry model created out of necessity
- Focused on Taiwan's manufacturing strengths, avoided design/marketing
Technical Interfaces and Abstraction
- Both companies abstract complexity through standardized interfaces
- AWS API system:
- Simple, easily understood method to abstract complex tasks
- Uses open web standards (XML, JSON) for communication
- Example: Amazon Recognition service for image object detection
- Application sends image via API, receives results without knowing internal workings
- TSMC Process Design Kit (PDK):
- Essentially an API for making semiconductors
- Set of design files with library of blocks
- Customizable per documented rules
- Allows designers to lay out and test designs using emulator
- CMOS technology fits building blocks model well
- Abstracts complexity of 5-nanometer chip manufacturing
- PDK is proprietary and exposes trade secrets (unlike AWS's open standards)
Customer Base and Market Dynamics
- TSMC customers: Fabulous semiconductor firms
- Medium-sized companies (hundreds vs thousands of employees)
- Serve specific niches different from Intel/established leaders
- Examples: Altera, Linux, Nvidia, Qualcomm
- Existed since 1980s but were small before TSMC
- Previously faced harsh terms from large manufacturers:
- Required IP transfer
- Manufacturer could make competing versions
- No guaranteed foundry capacity
- Nvidia example: Originally partnered with SGS Thompson, Jensen Huang wrote letters to become TSMC customer
- AWS customers: Primarily small/medium businesses
- Million+ customers, vast majority SMBs
- Developer shops appreciating pay-as-you-go flexibility
- Some mega-corporations: Airbnb, GE, Netflix
- Both platforms grew with their customers (Nvidia, AMD now multi-billion dollar giants)
Competitive Dynamics and Customer Relations
- TSMC: Explicit pledge not to compete with customers
- AWS: No such pledge, known for competing with platform users
- Jeff Bezos quote: "There might be categories... where we see it's an important product for customers and we make our own product offering"
- Uses customer data to compete with them
- Walmart doesn't want vendors using AWS for this reason
- Developers continue despite reputation due to pricing/lack of alternatives
- Pledge is implicit rather than explicit
Capital Expenditure and Scale Economics
- Massive capital requirements for both companies:
- TSMC 2021 capex: $25-28 billion
- AWS 2021 capex: $32 billion
- AWS: Building data centers in 24 geographic regions worldwide
- Both require land, buildings, expensive equipment, experienced personnel
- Economics favor outsourcing at low scale:
- Facilities cost billions but useful only few years
- Equipment depreciates rapidly like new cars
- Both companies aggregate demand to maximize facility utilization
- Accounting treatment: expenses become balance sheet assets, not income statement hits
Design Costs and Complexity
- Advanced semiconductor design extremely expensive:
- N5 and N3 cutting-edge nodes prohibitively expensive
- N3 design costs: $500 million to $1.5 billion
- Learning curve steep even with TSMC tools
Business Model Success and Risk Analysis
- Success factors:
- Focus on servicing other businesses
- Abstracting complexity of inherently complex systems
- Demand aggregation
- "Selling pickaxes and shovels during gold rush"
- Market perception: Companies without fabs/data centers seen as having better business models
- Price multiple comparisons: Intel vs AMD, Walmart vs Amazon
- Infrastructure dependency risks:
- Constant nervous tension for customers
- Business depends on external party for critical services
- Risk of sudden service termination (bad email scenario)
- Examples: AWS network boot, government bans on TSMC technology, Stripe account closures
- Everything can be "pulled out from under you"
