Every AI chip passes through
seven layers and a handful of hands.
Thirty facilities. Seven supply-chain layers. One map. This page traces the physical and corporate geography of the semiconductor industry, from rare-earth mines in Mongolia to EUV lithography in the Netherlands to advanced packaging in Taiwan. The data is drawn from public filings, industry associations, and open government reports.
Explore three interactive visualizations: a world map of facilities, a layer-by-layer flow diagram, and a chokepoint dashboard showing where concentration risk is highest.
TEST. 02 · Created with Kimi K2.6 · Language models can hallucinate. Verify primary sources.
AI runs on silicon. But silicon does not appear by magic.
A single leading-edge GPU passes through a global supply chain that spans rare-earth mining, polysilicon refinement, lithography equipment, wafer fabrication, memory stacking, advanced packaging, and final assembly. Each step is concentrated in a handful of companies and, often, a handful of countries. The result is a network that looks robust on paper but is fragile in practice.
This page maps that network. It shows where the facilities are, what flows between them, and where the chokepoints lie. The question underneath is simple: if one node fails, what stops?
Thirty facilities positioned by latitude and longitude. Toggle layers to isolate parts of the chain. Turn on flows to see the directed connections between facilities.
What this means
Each dot is a real facility with a known latitude and longitude. The colors encode the supply-chain layer. The map uses an equirectangular projection -- straightforward, but accurate enough to show the geographic concentration. Notice the clustering: design in California, fabrication in East Asia, equipment in the Netherlands and Japan, rare earths in China and Australia.
A Sankey-style view of the supply chain. Materials enter on the left; packaged chips exit on the right. Toggle between all layers and critical paths only to see where the bottlenecks are.
What this means
The width of each connection is proportional to the number of edges in the dataset, not dollar value or physical volume. Even so, the pattern is clear: TSMC sits at the center of the network, receiving inputs from nearly every upstream layer and shipping wafers to packaging and design. Remove TSMC and the graph falls apart.
Six metrics where market concentration exceeds 80%. Below: the Herfindahl- Hirschman Index (HHI) for each layer. An HHI above 2,500 indicates high concentration; above 5,000 indicates near-monopoly conditions.
What this means
The HHI is a standard antitrust metric: sum the squared market shares of all players. It ranges from near zero (perfect competition) to 10,000 (pure monopoly). Equipment and fabrication both exceed 2,500; fabrication hits 7,200, driven by TSMC's ~90% share of leading-edge logic. These numbers matter because they predict vulnerability: the higher the HHI, the more damage a single disruption can do.
The semiconductor supply chain is one of the most concentrated industrial networks on earth. ASML has a 100% monopoly on EUV lithography. TSMC makes ~90% of leading-edge logic. SK Hynix and Samsung together control ~95% of HBM, the memory type that determines AI training speed. NVIDIA holds ~90% of datacenter GPUs.
These are not ordinary market shares. They are structural chokepoints created by decades of cumulative technical expertise, patent moats, and capital intensity. A new competitor cannot simply enter. The barriers are too high and the learning curves too steep.
The honest conclusion: AI hardware progress is bottlenecked by a dozen companies and half a dozen countries. If any of them stumble, the whole stack slows.
Concentration is not the same as fragility. TSMC has survived earthquakes, typhoons, and geopolitical tension for decades. ASML's single-factory monopoly is a deliberate strategy: one perfect machine beats ten adequate ones. And every major player is diversifying.
Intel is spending $100B to return to leading edge. Samsung has poured $40B into new fabs. The U.S. CHIPS Act, the EU Chips Act, and Japan's subsidies are all aimed at reshoring capacity. By 2030 the geographic map may look very different.
The counter-argument is that diversification takes time. A fab takes 3--5 years to build and costs $20B. In the meantime, the current concentration is the only game in town.
Both stories are true. The supply chain is dangerously concentrated and the industry knows it and is spending unprecedented sums to fix it. The single question that splits them is timing.
Will diversification arrive before the next disruption? A Taiwan contingency, an export-control escalation, or a raw-material shortage could all stress-test the network before the new fabs come online.
Until then, the honest read is this: the map shows where the risk is. It does not predict when, or whether, that risk turns into a crisis.
Every chart is drawn with plain JavaScript and SVG. No frameworks, no build step, no analytics trackers. The data lives as simple JSON files right next to the page. Once loaded, it works offline.
Every number comes from a public dataset you can check yourself. Primary sources: SIA 2024 Factbook, CSIS Silicon Boxed report, Boston Consulting Group, CHIPS Act Tracker, and company SEC filings. Last updated April 2026.
SIA · 2024 Factbook ↗
CSIS · Technology Policy ↗
BCG · Turning the Tide ↗
CHIPS Act Tracker ↗
Market-share figures are approximate and move quarterly. The dataset covers only the most prominent facilities; many smaller players are omitted. Edge weights in the flow diagram are uniform, not proportional to real transaction volumes. The map projection distorts high-latitude regions.
This page was generated by a large language model (Kimi K2.6). While every dataset is anchored to a real public source, LLMs can hallucinate figures, misattribute findings, and confabulate citations. Do not treat any number here as verified fact without checking the primary source. The links in the reading list are the ground truth.