A strange thing happens when an AI data center moves from a financial model to a construction site: the GPU stops looking like the whole story. The AI infrastructure industry increasingly discusses capacity in terms of accelerators, racks, power availability and computing clusters. Yet those machines cannot operate without a dense physical system that sits underneath them. Transformers move electricity into the right voltage range. Switchgear distributes and protects it. Batteries provide backup. Optical modules and fiber infrastructure move data between computing systems. Cooling equipment keeps the resulting heat under control.

These components rarely dominate AI headlines, but shortages and long procurement timelines for critical infrastructure can materially affect when new computing capacity comes online. That creates a more uncomfortable supply-chain question for the U.S. AI buildout. America can expand domestic semiconductor manufacturing, assemble servers onshore and spend heavily on new power capacity while still relying on globally distributed manufacturing networks for the less glamorous equipment that connects those pieces. The vulnerability, in other words, may not sit inside the GPU rack. It may sit several layers upstream.

China’s Role Is Bigger Than the Hardware Americans Usually Associate With China

China’s relevance to the AI infrastructure supply chain does not depend on Chinese companies supplying frontier GPUs to U.S. data centers. The exposure can appear through equipment categories that receive considerably less attention. Transformer manufacturing provides one example. Chinese manufacturers are part of the global transformer supply chain, while U.S. authorities continue to identify limited domestic production capacity, foreign supply dependence and extended procurement timelines as challenges for critical transformer equipment. DOE’s current supply-chain program specifically targets transformers, components and materials while acknowledging dependence on imported inputs.

Optical infrastructure presents another layer. High-performance AI clusters increasingly depend on optical connectivity to move enormous volumes of data between servers and across networking systems. Industry reporting from China indicates that Chinese companies hold significant positions in optical-module manufacturing, with several Chinese vendors ranking among the world’s largest suppliers. That does not mean every U.S. AI facility depends directly on Chinese-made optical equipment.

It does mean that the global manufacturing base behind high-speed AI networking remains geographically concentrated in ways that the U.S. semiconductor narrative can obscure. Supply-chain exposure is not the same thing as dependence on a single Chinese supplier. Modern infrastructure rarely follows such a clean line. A component can involve Chinese manufacturing, non-Chinese intellectual property, materials from another country, final assembly somewhere else and a U.S. integrator before it reaches a data center. That complexity makes the vulnerability harder to see, not necessarily smaller.

The Real Question Is Whether America Is Building or Assembling

Washington’s industrial push increasingly focuses on bringing strategic production closer to home. NVIDIA, for example, now highlights U.S. expansion across parts of its AI manufacturing ecosystem, including chips, packaging, optics and systems. But an independent AI industrial base requires more than domestic production of the components that receive the most attention. It requires depth. That means the ability to manufacture transformers at scale, produce specialized electrical equipment, secure critical materials, build optical systems, fabricate cooling equipment and replace foreign inputs without sending project schedules into another planning cycle.

The distinction between building and assembling therefore becomes important. America could assemble an enormous amount of AI capacity domestically while remaining exposed to overseas industrial ecosystems for some of the equipment and materials that support that capacity. The resulting infrastructure would look American at the site level but remain international at the component level. That is not automatically a failure. Global specialization exists because it can deliver lower costs, greater scale and faster production. The problem emerges when geopolitical friction or sudden demand exposes how few alternative suppliers exist.

The Most Important AI Supply Chain May Be the Least Photogenic

There is an irony in the industry’s obsession with increasingly sophisticated silicon. The more powerful the accelerator becomes, the more demanding the surrounding infrastructure becomes. A next-generation AI cluster does not merely need more computing. It needs more electrical capacity, more sophisticated power distribution, higher-bandwidth networking, more cooling and increasingly complex supporting systems. The physical requirements compound as compute density rises. That creates a supply-chain hierarchy in which a comparatively mundane component can become strategically important simply because there are not enough of them.

DOE’s recent assessment makes the point without relying on AI hype: critical grid equipment can face lead times of two years or more, while the department is pursuing programs to increase domestic manufacturing and reduce reliance on foreign suppliers. The implication for AI developers is straightforward. The fastest path to more compute may not always involve buying more accelerators. It may involve securing the equipment that allows those accelerators to receive power, communicate and operate continuously. That is a much less glamorous industrial problem, but it may become the more consequential one.

America’s AI Independence Will Be Tested Outside the Server Rack

The U.S. does not need to eliminate every foreign supplier to create a resilient AI infrastructure base. That would be commercially unrealistic and would confuse self-sufficiency with resilience. The more useful test is whether a disruption involving one major manufacturing region can stop a large portion of the AI construction pipeline. If the answer is yes, then domestic AI capacity remains vulnerable even when the most visible technology sits inside U.S. facilities. That is why China’s supply-chain shadow matters. It is not primarily a story about Chinese GPUs appearing inside American data centers. It is about how deeply manufacturing specialization has become embedded in the physical machinery surrounding advanced computing, including components and systems supplied through international production networks.

The next phase of America’s AI buildout will therefore reveal something the GPU race cannot show on its own: whether the country is developing an industrial ecosystem capable of supporting frontier compute, while reducing its exposure to bottlenecks that depend on overseas manufacturing capacity. The uncomfortable part is that those bottlenecks may look almost boring. They may be copper, steel, fiber, transformers, switchgear and power electronics. And if those components can materially affect how quickly AI capacity comes online, they deserve to be treated as part of the AI race itself.