Working exhibit / Taiwan AI infrastructure wedge
NVIDIA GPU stack: where supply actually constrains deployment
A GPU is the visible object. The scarce product is the powered, cooled, networked rack delivered into a permitted site. The constraint map below separates the layers Lawrence and Andrew can influence from the layers they must route around.
Site and facility
01
Land, zoning, and permits
Plain English: Can we legally and physically build the thing here?
Think
- Industrial zoning
- Owner approval
- Fire code and seismic rules
- Floor loading and neighbors
Why it matters
AI data centers now look like industrial power assets. A normal office or warehouse cannot quietly become a GB200 site overnight.
Constraint
Land is scarce in Taiwan, permitting is local, and seismic/fire rules can kill a site before the GPU discussion starts.
site + permit
hard gate
02
Grid energy and power allocation
Plain English: Can the grid actually feed this monster?
Think
- Taipower allocation
- Nearby substations
- Transformers
- Backup generation and PPAs
Why it matters
AI clusters consume electricity like factories, steel mills, or small cities. Power became strategy, not a utility line item.
Constraint
No electricity means no project. This is why Taiwan's northern grid limits matter and why hyperscalers buy land directly.
grid feed
hard gate
03
Data center shell and operations
Plain English: The building and the people keeping it alive.
Think
- Security and guards
- Network rooms
- Maintenance staff
- Uptime procedures
Why it matters
If the facility goes down, AI training stops. The GPU is only valuable when the building operates reliably.
Constraint
Important, but less scarce than grid power, CoWoS, HBM, or liquid-cooling retrofit expertise.
shell + ops
watch item
04
Facility power train
Plain English: How electricity safely moves through the building.
Path
- Grid to transformer
- Switchgear to UPS
- PDU to rack busbar
- Grounding and protection
Why it matters
Older data centers handled 5-15 kW per rack. Modern AI racks can exceed 100 kW. That changes everything.
Constraint
Transformers, switchgear, UPS, busbars, and electrical contractors have brutal lead-time pressure.
power train
tight capacity
05
Facility cooling train
Plain English: How we stop the GPUs from melting.
Think
- Chillers and dry coolers
- Cooling towers
- CDUs and liquid loops
- Water treatment
Why it matters
New AI racks are too dense for old air-cooling assumptions. Liquid cooling is becoming the baseline.
Constraint
Without cooling, chips throttle or fail. Cooling feasibility is now part of site selection, not a later engineering chore.
liquid cooling
hard gate
Rack and system
06
NVIDIA rack system
Plain English: The AI supercomputer rack itself.
Contains
- Grace CPUs
- Blackwell GPUs
- NVLink switching
- Rack power and liquid manifolds
Why it matters
This is not buying graphics cards. It is buying an AI factory module with compute, power, cooling, networking, and software.
Constraint
NVIDIA increasingly sells full systems. Allocation is tied to the whole rack ecosystem, not just GPU chips.
rack system
tight capacity
07
ODM rack integration
Plain English: The companies physically assembling and validating AI systems.
Work
- Assemble racks
- Integrate cooling
- Wire and test systems
- Manufacture at scale
Why it matters
Taiwan dominates this layer. That is why Taiwan is critical even when the buyer is American, European, or sovereign.
Constraint
ODM factory slots, liquid-cooling readiness, and validation capacity can become practical deployment bottlenecks.
rack assembly
tight capacity
08
Network, fiber, and optics
Plain English: How the AI systems communicate.
Think
- Switches and NICs
- Optical transceivers
- Cross-connects
- Submarine cables
Why it matters
Thousands of GPUs need to talk constantly. If networking is slow, GPUs sit idle and money burns.
Constraint
Especially important for APAC latency, sovereign AI, multi-region deployments, and carrier diversity.
fiber network
watch item
09
Board power delivery
Plain English: How electricity is managed inside the AI server.
Think
- Voltage regulation
- High-current connectors
- Firmware and telemetry
- Thermal safety
Why it matters
Power density exploded. Tiny failures can burn connectors, crash systems, or create thermal faults.
Constraint
Power electronics are no longer boring. They are part of the AI rack's reliability envelope.
board power
tight capacity
Silicon and package
10
GPU package
Plain English: The assembled AI chip package, not just the GPU die.
Contains
- GPU logic dies
- HBM memory
- Interposer
- Package substrate
Why it matters
HBM sits extremely close to the GPU for speed. This package is one of the hardest manufactured objects in AI.
Constraint
Yield problems here are catastrophic because multiple scarce parts are combined into one finished component.
GPU package
hard gate
11
HBM memory supply
Plain English: The stacked memory sitting next to the GPU so data can move fast enough.
Think
- HBM3E today
- HBM4 next
- Memory stacks
- Test and yield
Why it matters
Modern AI chips are starved without memory bandwidth. The GPU die alone is not useful enough.
Constraint
Even if GPU dies exist, finished package output is capped if HBM supply or quality is short.
HBM stacks
hard gate
12
Chip architecture, design, and EDA/IP
Plain English: The blueprint layer before TSMC can manufacture anything.
Think
- NVIDIA architecture
- EDA tools
- IP blocks
- Verification and physical design
Why it matters
NVIDIA's dominance starts here: architecture, CUDA/software coupling, verification, and designs TSMC can build.
Constraint
This is talent/tool/IP constrained. It is not the Taiwan real-estate wedge, but it explains why NVIDIA orchestrates the stack.
chip design
hard gate
13
TSMC advanced packaging
Plain English: The magic glue layer enabling modern AI.
Think
- CoWoS
- 2.5D integration
- Interposers
- GPU plus HBM
Why it matters
CoWoS lets GPU logic and HBM function together efficiently. You cannot instantly build more of this capacity.
Constraint
This became one of the world's central AI bottlenecks and made TSMC even more strategic.
CoWoS
hard gate
14
TSMC wafer fabrication
Plain English: Where the GPU dies are born.
Requires
- EUV lithography
- Advanced cleanrooms
- Gigantic capex
- Elite process engineering
Why it matters
Only a handful of companies can manufacture frontier logic at scale. For NVIDIA, TSMC is the key manufacturing partner.
Constraint
Without wafers, nothing else exists. TSMC process capacity remains a hard upstream gate.
wafer fab
hard gate
The big realization
AI stopped being just software. It became energy, industrial systems, semiconductors, infrastructure, and supply-chain orchestration.
Where Lawrence and Andrew can play
| Layer | Role in wedge |
|---|---|
| Land and site control | Andrew can surface Taiwan-side options and ownership reality. |
| Power path | The first diligence question: existing MW, expandable MW, substation distance, and approval path. |
| Cooling feasibility | Filter out sites that cannot handle liquid-cooled rack density. |
| Integration slate | Map ODM, cooling, electrical, network, freight, and operations counterparties. |
| Buyer mandate | Lawrence packages the options into a neutral Western-facing work product. |