Artificial Intelligence: Supply-Chain Chokepoints and the Reach of Industrial Policy

The Hidden Bottlenecks of the AI Stack

While most people worry about a few big companies controlling AI models, the real danger lies much further back in the supply chain. The most critical parts—like the machines that make chips and the minerals used to build them—are controlled by very few companies or even single countries. These create massive chokepoints. This study suggests the competition debate is looking at the wrong end of the telescope.

Looking at the wrong end of the stack

Current discussions regarding artificial intelligence competition typically focus on the "downstream" layer. This refers to the developers of frontier models (large-scale AI systems). Regulators worry if a handful of firms are forming an oligopoly (a market dominated by a few large players). However, the authors argue this focus ignores how AI is actually built.

Existing antitrust (laws preventing unfair competition) frameworks monitor mergers and market shares among domestic firms. But the most intense concentration does not happen among companies selling AI services. Instead, it rises as you move "upstream"—away from software and toward physical hardware and raw materials. While the model layer might appear competitive, the layers providing essential ingredients are often monopolies or state-controlled.

Mapping the concentration gradient

To visualize this, the authors use a production-function framework. They treat AI capability as an output produced from a hierarchical stack of inputs. A trained model requires compute (processing power); compute requires chips; chips require logic, memory, and advanced packaging (the process of connecting components). These components require electricity and refined minerals.

The study uses the Herfindahl-Hirschman Index (HHI) to provide a common yardstick. The HHI is a standard economic metric that sums the squares of market shares. It ranges from 0 to 10,000. A score of 10,000 represents a total monopoly. By applying this scale to everything from cloud providers to gallium refineries, the authors reveal a stark "concentration gradient" .

The mechanism of this gradient is a transition of power. At the downstream end, the market is fragmented among various software providers. As you move into the midstream—the hardware and fabrication stages—the index climbs sharply. Finally, at the upstream base, the index hits near-total saturation. Here, single nations or companies control the flow of essential physical matter.

From software competition to statecraft

The evidence for this gradient is mathematically severe. The authors report that frontier-model usage is relatively diffuse. It has an HHI of approximately 1,150 . Even the cloud infrastructure layer remains below the 1,800 mark. This is the threshold United States agencies use to define a "highly concentrated" market.

However, the landscape shifts once you enter the hardware supply chain. The paper finds that high-bandwidth memory (HBM) has an HHI of 4,174 [Table 1]. Advanced packaging reaches 8,100, dominated by a single firm [Table 1]. At the extreme edge, leading-edge EUV lithography (machines that print microscopic circuit patterns) hits the theoretical ceiling of 10,000 .

This concentration is not limited to chips. The authors report that the most extreme bottlenecks are in the mineral layers. For example, gallium production is dominated by a single country. Its HHI reaches 9,804 [Table 1]. Because these upstream layers are often in foreign jurisdictions or managed by states, traditional antitrust tools are ineffective. You cannot use merger review to regulate a foreign state's refining capacity. Instead, the struggle shifts to export controls and industrial policy (government actions to support domestic industry).

Built capacity vs. geological luck

A critical distinction the authors make is whether these chokepoints are "geological" or "built." Geological concentration is dictated by where minerals are in the earth. Built concentration is created by industrial capacity. This distinction determines if a policymaker can solve the problem.

The study compares mineral reserves with actual production and refining levels . The authors find a consistent pattern. Concentration is lowest in the ground (reserves) and highest at the refinery. For instance, cobalt mining is somewhat concentrated in the Democratic Republic of the Congo. Yet, cobalt refining is even more heavily concentrated in China [Table 3]. Similarly, graphite is widely available in nature. However, its processing is controlled by a single state with an HHI of 8,649 [Table 1].

This reveals that the AI supply chain's vulnerability is largely a matter of industrial capacity. It is not necessarily about resource scarcity. For minerals like gallium, concentration is entirely a result of who owns the processing plants. The paper notes that for many inputs, the primary threat is not a price spike. Rather, the threat is the total loss of availability [Section 3.2].

Assessing the strategic risk

The authors conclude that the AI economy faces a mismatch between perceived risks and actual vulnerabilities. While the world watches the "model wars," systemic risk accumulates in the silent, upstream layers.

The paper does not explore how new technologies might flatten this gradient. It also does not quantify the exact cost of a total supply cutoff. Additionally, it relies on proxies for certain layers. For example, it uses cloud revenue as a stand-in for installed compute capacity.

However, the core verdict is clear. The stability of the AI revolution depends on physical control of the supply chain. For practitioners and policymakers, the takeaway is vital. Building resilient AI requires moving beyond merger oversight. It requires active industrial strategy. Focus must shift toward refining capacity, mineral recovery, and diversifying physical processing hubs. Code and the full data pipeline are reportedly available via the companion dashboard at https://aistackmap.org.

Figures from the paper