The AI Chip Supply Chain, Explained in One Map

The hardware artificial intelligence runs on passes through a supply chain with extraordinary geographic concentration. Understanding it explains most of the geopolitics currently shaping the industry.

Design:

AI accelerators are designed principally in the United States, with the dominant designer headquartered in California and significant efforts at other American firms and at cloud providers designing their own silicon.

Chip design also depends on electronic design automation software, supplied by a small number of firms, overwhelmingly American. Processor architecture licensing adds another chokepoint, with a British-headquartered firm central to it.

Manufacturing equipment:

Turning designs into chips requires equipment of extreme complexity. Extreme ultraviolet lithography systems, essential for the most advanced nodes, come from a single Dutch company. Other critical equipment comes from American and Japanese suppliers.

Japan is also central in materials: photoresists, specialty chemicals and silicon wafers come disproportionately from Japanese firms.

Fabrication:

The overwhelming majority of leading-edge AI chips are manufactured in Taiwan, principally by one company. South Korea has significant advanced capacity. The United States, Japan and Europe are building more with substantial public subsidy, and new facilities take years to reach comparable yields.

Fabrication is where the concentration is most acute and the geopolitical risk highest.

Packaging:

Increasingly critical and frequently the actual bottleneck. Modern AI chips combine multiple components, including stacked memory, into single packages, and advanced packaging capacity is concentrated in Taiwan with growing capability in Malaysia, South Korea and elsewhere.

Packaging shortages have limited accelerator supply repeatedly, independent of fabrication capacity.

Memory:

High-bandwidth memory is essential to AI accelerator performance and is produced by a small number of firms, principally in South Korea with significant American and Chinese participation. Memory supply has been the binding constraint on accelerator production in multiple periods.

Assembly and systems:

Servers and racks incorporating these chips are designed and largely built by Taiwanese firms, often manufactured in Taiwan, China, Mexico or Southeast Asia.

Raw inputs:

Underneath sits mining: copper, principally from Chile, Peru and the DRC; cobalt overwhelmingly from the DRC; rare earths processed predominantly in China; and various specialty materials with concentrated sources.

The chokepoints that matter:

Four points concentrate risk: Dutch lithography, Taiwanese fabrication and packaging, Korean memory, and Chinese processing of critical minerals.

Each is difficult to replicate quickly. Each has become an instrument of policy, with export controls, subsidies and investment screening all deployed to shape access.

Why diversification is slow:

Building advanced fabrication requires capital in the tens of billions, years of construction and, most importantly, accumulated process expertise that cannot be purchased. New facilities routinely take longer than announced to reach production yields.

Subsidy programmes in the United States, Europe, Japan, India and elsewhere are attempting it, and results are measured in years rather than quarters.

The outlook:

The supply chain will diversify gradually and remain concentrated for the foreseeable future. Anyone forecasting AI capability, national or corporate, is implicitly forecasting the stability of these few points, which is why a strait in East Asia and an export licence in the Netherlands matter more to AI's trajectory than most model releases.