人工智能：供应链瓶颈与产业政策的影响范围
Artificial Intelligence: Supply-Chain Chokepoints and the Reach of Industrial Policy
浏览论文内容
中文总结 AI 辅助
本文采用赫芬达尔-赫希曼指数测算AI供应链各环节集中度，发现上游是战略瓶颈，其集中度为人为形成，需通过出口管制和国内产业政策应对。
中文摘要 AI 辅助
人工智能依赖多层投入：模型依赖算力，算力依赖芯片，芯片依赖电力与精炼矿物。本文采用赫芬达尔-赫希曼指数（HHI），基于可引用、可复现的数据，对各层的集中度进行测算，得出三项核心发现：第一，集中度呈现清晰梯度：下游集中度较低，处于美国机构认定的高度集中度阈值1800以下，也是公众与监管关注最集中的领域；上游集中度急剧上升，公众讨论较少，其中先进封装的HHI达8100，前沿光刻达到10000的上限，某一国家主导了镓等多种关键矿物的生产或精炼，其集中度接近上限。第二，这些上游环节是整个AI经济的瓶颈和战略弱点，反垄断法无法触及，因为它们掌握在外国或国有主体手中，需通过出口管制和国内产业政策应对。第三，在精炼环节旁设置储备显示，集中度是人为形成而非地质决定，属于产业政策可调控的变量；上游投入品价格上涨对下游产品成本影响极小，因为投入品仅占成本的一小部分，因此瓶颈的威胁在于投入品本身的短缺，而非价格上涨。
英文摘要
Artificial intelligence depends on a stack of inputs, models on compute, compute on chips, and chips on electricity and refined minerals. This paper measures the concentration of each layer on one scale, the Herfindahl-Hirschman Index (HHI), from cited and reproducible data. Three findings follow. First, concentration forms a clear gradient. It is modest downstream, where public and regulatory attention is heaviest and model usage and cloud fall below the 1{,}800 mark United States agencies treat as highly concentrated. It rises steeply upstream, where public discourse is sparse. There advanced packaging scores 8{,}100, leading-edge lithography reaches the ceiling of 10{,}000, and one country dominates the production or refining of several critical minerals, gallium near that maximum. Second, these upstream layers are chokepoints and a strategic vulnerability for the whole AI economy. Antitrust cannot reach them, because they lie in foreign or state hands. Contesting them falls to export controls and domestic industrial policy. Third, placing reserves beside refining shows the concentration is built rather than geological, a variable industrial policy can move. A rise in an upstream input's price barely changes the cost of the product built from it, because the input is only a small share of that cost. A chokepoint's threat is therefore the loss of the input itself rather than a higher price.
