Mapping the physical movement of materials, components, and systems across 6 layers and 12+ countries | April 2026
Click any layer to see detailed physical flows, companies, and freight characteristics.
Wafers, chemicals, gases, rare earths, EDA
EUV lithography, etch, deposition, inspection
Foundries, HBM memory, advanced packaging
GPUs, ASICs, CPUs, AI servers
Power, cooling, networking, cabling, DCs
Models, APIs, training, inference
| Origin | Company | Cargo | Destination | Mode |
|---|---|---|---|---|
| Japan | Shin-Etsu, SUMCO | 300mm silicon wafers | ASML (NL), Applied Materials (US) | Air / Sea |
| Japan | TOK, JSR, Fujifilm | EUV photoresists | TSMC (Taiwan), Samsung (Korea) | Temp-controlled air |
| US/France/Germany | Linde, Air Liquide, Air Products | Specialty gases (xenon, neon) | All major fabs globally | Hazmat certified |
| US/Australia | MP Materials, Lynas | Rare earth minerals | Magnet / component mfrs | Bulk sea + air |
| Japan/Taiwan | Ajinomoto, Unimicron, Ibiden | ABF substrates | TSMC, ASE Group (Taiwan) | Air freight |
| Origin | Company | Cargo | Destination | Mode |
|---|---|---|---|---|
| Netherlands | ASML | EUV machines (180t, 40 containers each) | TSMC Taiwan, TSMC Arizona, Samsung Korea, Intel Arizona | White-glove oversized |
| California, US | Applied Materials, Lam Research, KLA | Deposition, etch, inspection tools + spare parts | TSMC, Samsung, SK Hynix, Micron, Intel fabs | Air freight + AOG express |
| Japan | Tokyo Electron | Deposition, cleaning tools | All major fabs | Sea + air |
| Origin | Company | Cargo | Destination | Mode |
|---|---|---|---|---|
| Taiwan | TSMC | Finished wafers / packaged dies (B200, MI350X, TPU, Trainium) | Nvidia, AMD, Broadcom, Google, Amazon (California/Seattle) | High-value air freight |
| South Korea | SK Hynix | HBM3e memory stacks | TSMC Taiwan (for CoWoS packaging) | Temp-controlled air |
| South Korea | Samsung Memory | HBM3e | TSMC, AMD, Google | Air freight |
| US/Japan | Micron | HBM3e | TSMC Taiwan | Air freight |
| Taiwan/Korea | ASE Group, Amkor | Final packaged AI chips | Nvidia, AMD, Broadcom | Secure air freight |
| Origin | Company | Cargo | Destination | Mode |
|---|---|---|---|---|
| California | Nvidia | GPU modules (B200/GB200) | Dell (TX), Supermicro (CA), HPE (TX), direct to hyperscalers | Secure express |
| California | AMD | MI350X accelerator modules | Dell, HPE, Supermicro, Microsoft, Meta | Express air |
| Texas / California | Dell, Supermicro | Assembled AI server racks (multi-ton) | AWS (Virginia, Oregon), Google (Iowa), Meta (Louisiana, TX) | White-glove trucking |
| Category | Company | Cargo | Origin | Mode |
|---|---|---|---|---|
| Power | Eaton, Schneider, Vertiv, ABB | UPS systems, PDUs, switchgear | US, France, Switzerland | Heavy freight / trucking |
| Generators | Caterpillar, Cummins | On-site generators (multi-MW) | Illinois, Indiana (US) | Oversized freight |
| Cooling | Vertiv, Schneider, Stulz, LiquidStack | CDUs, liquid cooling, immersion tanks | US, Germany, Netherlands | Heavy freight |
| Networking | Arista, Cisco, Nvidia/Mellanox | Ethernet/InfiniBand switches | California, Israel | Express air |
| Cabling | TE Connectivity, Amphenol, Corning | DAC cables, fiber optic | Switzerland, Connecticut, New York | Standard freight |
| Optics | Coherent, Lumentum, Ciena | 800G transceivers, coherent modules | Pennsylvania, Maryland | Express air |
This flow is primarily digital (API calls, model weights over network). Physical flows are limited to initial GPU server deployments covered in L4 to L5. Key relationships: AWS serves OpenAI + Anthropic | Google Cloud runs Gemini on TPU pods | Microsoft Azure hosts OpenAI exclusively | Meta runs LLaMA on private infrastructure | CoreWeave provides overflow GPU capacity.
16 steps from silicon sand to running GPT-5 inference.
High-purity quartz sand refined for semiconductor use
Sand refined into ultra-pure silicon ingots, sliced into 300mm wafers
Temperature and vibration controlled transport to TSMC
Light-sensitive chemical coating enables circuit patterning at 3nm/2nm
Billions of transistors printed at nanometer scale
Material layers added and removed to form 3D circuit structures
High Bandwidth Memory stacks shipped for co-packaging with GPU die
GPU die + HBM stacked together on silicon interposer (Chip-on-Wafer-on-Substrate)
Completed B200 GPU modules shipped under tight security
GPU mounted on PCB with power delivery and thermal solution
8x GPU modules per server tray for DGX / HGX systems
Complete rack: GPUs + CPUs + networking + power + cooling integration
Multi-ton racks transported with vibration protection
Eaton UPS, Vertiv CDU liquid cooling, Arista Ethernet switch connected
GPU cluster linked to storage and other racks at 800Gbps per link
From silicon sand to answering user queries in ~4 to 6 months elapsed time
Dominates L1 (wafers 55%, photoresists 70%) and L2 (TEL 15% equipment)
L3 hub: 66% foundry share, 90%+ advanced AI chips, CoWoS packaging
HBM memory capital: SK Hynix ~57-62% share, Samsung ~22%
Single point of failure: EUV lithography monopoly, ~19% equipment market
Chip design, EDA, equipment, networking HQ hub
Data center operations, power/cooling manufacturing, fiber optics
Strong in equipment (ASML), power/cooling, IP (ARM), and emerging AI (Mistral)
InfiniBand networking R&D and manufacturing hub for AI cluster interconnects