How Washington Learned to Sit at the Chokepoint

The United States built its compute leverage on an accident of industrial structure that it then learned, deliberately, to exploit. The accident is that the advanced-chip supply chain is a series of narrow monopolies and duopolies, and several of the narrowest sit under American jurisdiction or that of close allies. The deliberate part is the recognition, formalized from October 2022 onward, that controlling those narrow points could shape the global distribution of AI capability without America having to manufacture the chips itself.

Walk the chain. Frontier AI accelerators are overwhelmingly designed by American firms — Nvidia above all, with AMD and a set of custom silicon efforts at the hyperscalers. The most advanced of those chips are fabricated by one company, TSMC, in Taiwan, using lithography machines for the finest features made by exactly one company, ASML, in the Netherlands. The electronic design automation software used to design chips is dominated by a few firms, most American. At several of these points the number of viable suppliers is one or two. That is what a chokepoint is: a place where a single decision can constrain everyone downstream.

Washington's insight was that it did not need to control every link, only enough of them, and that its legal reach extended further than the chips physically made on its soil. Through mechanisms like the foreign direct product rule, the United States asserted authority over products made abroad if they were built using American technology or software — which, in advanced semiconductors, nearly everything is. That is how a chip fabricated in Taiwan with Dutch machines can still fall under American export jurisdiction. The chokepoint is not geographic; it is technological and legal.

The October 2022 controls were the first full expression of this strategy, aimed at cutting China off from the most advanced AI chips and, crucially, from the equipment to make them domestically. The logic was not merely to deny China chips today but to prevent it from building the capacity to make its own, freezing a capability gap in place. Subsequent revisions in 2023 and beyond tightened definitions as vendors designed compliant-but-still-powerful chips for the Chinese market, a cat-and-mouse dynamic that itself reveals how the controls work: they set a performance threshold, the market engineers just beneath it, and the threshold moves.

What elevated this from a China policy to a global compute-map policy was the extension of controls beyond a single adversary into a tiered framework governing much of the world. Rather than a simple in-or-out line, the emerging architecture sorted countries into categories — close allies with broad access, a large middle group subject to licensing and quantity caps, and restricted destinations. For the middle group especially, this changed everything. A Gulf state or a Southeast Asian nation seeking to build large AI data centers now had to consider not just commercial terms but whether American licensing would permit the volume of chips it wanted, and what assurances it would have to give in return. Compute access became something you negotiated with Washington, not only with Nvidia.

This is the mechanism behind the course thesis. When access to the essential input is gated by one government's rules, every other government's compute strategy becomes partly a strategy toward that government. The controls do not have to forbid a data center to shape it; they merely have to make the chips inside it contingent on conditions — on where the facility is, who operates it, what security measures it adopts, and sometimes on the buyer's other geopolitical relationships. Nations respond rationally: they seek terms, offer guarantees, take on American or allied partners, and structure deals to fit inside the permitted lanes.

The critical analytical habit is to see the controls as a routing system rather than a blockade. Capacity that cannot be built in one configuration gets built in another — with a different partner, under different assurances, financed differently, sited differently. The controls rarely make compute vanish; they change its citizenship and its terms. Understanding a sovereign program therefore starts with understanding which lane of the control regime that nation sits in, because the lane determines the menu of deals available to it. In the next lesson we examine what happens to the nations on the receiving end of that line.