The new Fort Worth factory is not diversifying supply chains—it is locking down a monopoly for a single company on American soil.
It is true, in the narrow sense of physical geography, that moving some assembly of Nvidia’s superchips out of Taiwan reduces the risk that a cross‑strait conflict could interrupt the global supply of the most advanced AI hardware. The trouble is that what is being built in that 324,000‑square‑foot facility is not a general‑purpose semiconductor plant that could serve multiple firms. It is a dedicated assembly line for a single company’s proprietary modules—the Grace Blackwell GB300 and the forthcoming Vera Rubin products—designed from the silicon up to lock customers into Nvidia’s ecosystem and to lock competitors out of the manufacturing capacity that the U.S. government is now subsidizing to the tune of tens of billions of dollars.
The plant, opened this week by Taiwanese contract manufacturer Wistron at a cost of $700 million, will produce “tens of thousands of boards a month” of Nvidia’s custom AI modules. Wistron is best known as one of the Foxconn‑style firms that built its fortune assembling consumer gear for Apple. Now it is pivoting, like much of the global electronics supply chain, to become a captive logistics arm for the most profitable chip company in history. The factory is not a neutral foundry; it is a single‑customer assembly line whose output is tied to a single firm’s proprietary architecture. That matters because a factory built to produce Nvidia’s Grace Blackwell modules cannot, without substantial retooling, produce anyone else’s AI accelerator. The capacity is dedicated, and the public dollars that helped make it viable are, in effect, deepening a monopoly.
Nvidia’s announcement frames the investment as part of a $500 billion commitment to “advanced AI platforms in the U.S.”—a number too round to audit, too large to question seriously, and, for now, almost entirely notional. The hard, physical manifestation is Wistron’s $700 million plant and the $100 billion that Taiwan Semiconductor Manufacturing Co. has pledged for its Arizona fabs, the latter also overwhelmingly dedicated to Nvidia and Apple. The CHIPS and Science Act of 2022 was supposed to catalyze a broad semiconductor renaissance; instead, its largest beneficiaries are the two most valuable chip firms on Earth, each with gross margins above 60 percent, each with market capitalizations larger than the GDP of most nations, each now employing the language of national security to secure public subsidies for production lines that serve themselves.
This is the chokepoint in its latest form. Cory Doctorow and Rebecca Giblin’s framework for the creative industries—intermediaries inserting themselves between creators and audiences, squeezing both—applies with equal force to the AI hardware stack. Nvidia controls the design layer (the creator of the platform). Through captive manufacturers like Wistron, it controls the assembly layer (the intermediary), which absorbs the cost of dedicated capacity, tooling, and labor with no recourse to any other buyer. And it controls access to the end users (the audience), who are locked into CUDA, locked into Nvidia’s proprietary architectures, and priced accordingly. The squeeze is the same: the creator and the audience are served, and the intermediary gets what is left. Nvidia already controls the design layer through its CUDA software moat, which ensures that almost all machine‑learning workloads are written for its GPUs. It influences the fabrication layer through its relationship with TSMC, where it is the largest customer and where allocation of leading‑edge capacity is a zero‑sum game among a handful of hyperscalers. Now it is moving into the assembly and test layer, building dedicated, subsidized plants in Texas and ensuring that the physical assembly of its modules happens inside the United States, on terms that make it harder for any potential rival to access equivalent capacity at equivalent cost.
The Trump administration’s push to reshore high‑tech manufacturing is, in principle, defensible. A supply chain concentrated in Taiwan is vulnerable, and some geographic diversification is prudent. But the form that diversification is taking is not a resilient, multi‑sourced industrial base; it is the replication of the dominant firm’s captive logistics network on American soil, with the U.S. taxpayer underwriting the transition. The administration is not building an AI manufacturing commons; it is building a company town.
Jensen Huang is not subtle about the dynamic. Standing in front of the GB300 modules at the opening ceremony, he told the assembled dignitaries, “We’re producing them like phones… cranking them out in volume because the world needs all of these machines to drive the intelligence infrastructure.” The comparison to phones is more revealing than he perhaps intended. The phone supply chain is the canonical lesson in how captive manufacturing erodes bargaining power: Apple’s contractors bid against one another for razor‑thin margins while Apple captures the bulk of the value. Wistron’s pivot to Nvidia modules is the same story in a different casing. The Taiwanese firm is installing expensive equipment and hiring American workers to serve a single customer who, by design, will be the only buyer of what that factory produces. The power asymmetry is structural, and the plant’s output will be priced accordingly.
What is absent from the announcement is any commitment to interoperability, to open standards, or to capacity that could be used by anyone other than Nvidia. Lina Khan’s foundational insight—that platform power is exercised not only through pricing but through the control of the infrastructure on which rivals depend—applies here in full force. A competitor who wants to build an AI accelerator must not only design a better chip; it must find a way to manufacture it without access to the TSMC capacity already locked up by Nvidia, without a software ecosystem to match CUDA, and now without access to the dedicated assembly lines that the U.S. government is helping Nvidia build under the banner of national resilience.
The rebate is real, but the resilience is narrow. A factory that can only produce one firm’s proprietary modules does not make the supply chain less brittle—it merely relocates the same single point of failure onto American territory. The cross‑strait risk that justified this entire reshoring push remains unmitigated: the plant cannot serve anyone but Nvidia, the assembly chain still depends on TSMC’s leading‑edge fabrication and single‑source high‑bandwidth memory from Korea, and the “reindustrialization” narrative masks the fact that the single point of failure has simply moved from Taiwan to Texas. If the geopolitical tail risk that justifies these subsidies ever materializes, Fort Worth will not be building AI modules for a diversified set of competitors; it will be building them for Nvidia. And if Nvidia’s proprietary supply chain is interrupted anywhere—by a TSMC outage in Arizona, by a shortage of advanced packaging, by a trade dispute over that Korean memory—the Fort Worth plant stops, and the entire AI infrastructure that has been so carefully “reshored” stops with it.
The plant is open. The ribbon is cut. The tax abatements are locked in. The public has bought itself a captive supply chain for a single company, disguised as a revival of American manufacturing. The competition is still waiting for a turn at the line.