The logic rests on a chokepoint. Training a frontier AI model requires enormous quantities of specialized chips, and the supply chain for those chips is astonishingly concentrated: a handful of firms design the leading processors, one company in the Netherlands makes the lithography machines required to fabricate them, and a small number of foundries actually produce them. This concentration gives the states that host those firms a lever over global AI development that exists nowhere else in the technology stack.

Since 2022 the United States has restricted the sale of the most capable AI chips and the tools to make them, aiming to slow rivals' access to the hardware that frontier AI requires. It is the most muscular AI measure in force, and a revealing case study in the limits of unilateral action.

How the controls work

Beginning in October 2022 and tightened repeatedly since, US regulations have restricted the export of the most advanced AI chips and the semiconductor manufacturing equipment needed to produce them, particularly to China. The measures use performance thresholds to define controlled chips, place specific companies on restricted lists, and (through rules that reach products made abroad with US technology) extend American jurisdiction across the global supply chain. Washington has also pressed allies, notably the Netherlands and Japan, to align their own controls, since the chokepoint only holds if the key suppliers act together.

Why this is a governance tool, not just trade policy

Export controls matter for AI safety because they demonstrate, in practice, that frontier development can be influenced through the hardware layer. The same chokepoint that lets a state slow a rival's access could, in principle, underpin a cooperative regime: if compute is the bottleneck, then compute is where verification and limits can bite. Every serious proposal for compute governance (tracking chips, monitoring large clusters, conditioning access on safety compliance) depends on the concentrated, controllable supply chain that export controls have shown is real.

  • They prove the chokepoint exists. Controls work because advanced-chip supply is concentrated enough to regulate, the precondition for any compute-based governance.
  • They build the tracking machinery. Enforcing controls requires knowing where chips go, creating monitoring capacity a treaty could later use cooperatively.
  • They establish that access can be conditioned. Once access to compute is a policy lever, it can be tied to safety obligations, not only to geopolitics.

The risks and limits of the unilateral approach

But export controls as currently used are a competitive weapon, not a safety measure, and that distinction carries real risks. Because they are aimed at denying a rival capability rather than reducing shared risk, they intensify the very race dynamics that make AI dangerous. A China that feels it is being strangled has stronger incentives to race, to build an indigenous supply chain, and to reject any cooperative regime it sees as a Western instrument of containment. Controls can accelerate the fragmentation of the technology world into rival blocs, the opposite of the shared governance safety requires.

They are also leaky and self-eroding. Chips are smuggled, thresholds are gamed by designing just under the line, and sustained restriction spurs exactly the indigenous development it aims to prevent, potentially dissolving the chokepoint over time. And a purely unilateral tool has no reciprocity: it constrains the target without constraining the wielder, so it cannot deliver the mutual, verified restraint that reduces catastrophic risk on all sides.

Export controls prove that the hardware chokepoint is real and usable. The question is what we use it for, to win a race, or to govern one. The same lever that fuels the competition could anchor the cooperation that ends it.

From weapon to instrument of governance

The deepest lesson of export controls is that the compute chokepoint they exploit is the single most promising point of leverage for ASI governance, and that it is currently being used for competition rather than safety. A wiser strategy would repurpose the same mechanism cooperatively: shared, verifiable limits on the most dangerous scales of development, access to compute conditioned on safety compliance rather than on geopolitical alignment, and allied coordination aimed at reducing collective risk rather than denying a rival. Export controls show the lever works. The task for governance is to grip it for a different purpose, turning a chokepoint now used to run the race into the mechanism that could verifiably slow it.