There is a comforting idea behind every export-control list, every sanctions regime, and every corporate decision to refuse a sale on ethical grounds: that if you withhold a critical input, you can stop the thing you fear. Stop the chemical weapon, the centrifuge, the frontier AI chip. The implicit model is a wall. Refuse the brick and the wall never gets built.

I spent a day pulling on this thread across four cases — three export-control regimes and one running the other direction — and came away convinced the wall metaphor is wrong in a specific, instructive way. A chokepoint is not a wall. It is a clock. It does not stop the thing; it buys time. And the amount of time it buys has almost nothing to do with how “irreplaceable” the input looks, and almost everything to do with how deep the tacit knowledge runs beneath it.

Let me walk the cases from most-fungible to least, because the gradient is where the structure shows.

Four cases, arranged by how easily they leak

The Australia Group — refusal as a speed bump. Founded in 1985 after Iraq’s use of chemical weapons, the Australia Group is an informal, voluntary club — now 42 nations plus the European Union — that coordinates export controls on the precursors and equipment for chemical and biological weapons.1 Forty years on, the honest verdict is that it stunted proliferation without ever stopping it. The reason is baked into the chemistry: precursors are dual-use industrial commodities. A non-member can supply what a member refuses. Conscience here is an impediment — a cost imposed — not a halt.

A.Q. Khan — even the “irreplaceable” leaks one layer down. Fissile material is the textbook non-fungible input: enrichment infrastructure is scarce and oligopolistic, so in principle you can physically throttle supply. And yet the Khan network routed around the Nuclear Suppliers Group not by buying fissile material on some black market but by going one layer below it. In 2001 a contract was signed with Scomi Precision Engineering in Malaysia for roughly 25,000 aluminum centrifuge components, shipped through Dubai front companies; in 2003 the cargo ship BBC China was intercepted en route to Libya carrying parts for a thousand centrifuges.2 The material was non-fungible. The components, and the know-how to machine them, were not. The network was eventually dismantled — Khan confessed on television in February 2004 — but the lesson holds: even the hardest chokepoint has a softer layer beneath it, and supply finds that layer.

ASML’s EUV — a real knowledge chokepoint the system still routes around. This is the genuine article. ASML holds an effective 100% monopoly on extreme-ultraviolet lithography, and the barrier is not capital — it is time and tacit knowledge, accreted across more than 800 suppliers and a decades-long optics partnership with Zeiss. ASML’s CEO said in April that China would need “many, many years” to replicate it, and China’s domestic EUV effort remains at the prototype-and-testing stage, with mass production realistically not before the end of the decade.3 So the node holds. And yet the system goal — advanced chips — got routed around anyway. SMIC is printing 5nm-class silicon for Huawei’s Kirin line using deep-ultraviolet tools and multi-patterning (four exposures where EUV needs one), at yields reported around 20% — far below the ~70% normally required for commercial viability — and at a steep price premium.4 The EUV node is a halt locally. The system halt never arrives. The chokepoint buys a generation or two of delay, and no more.

Rare earths — concentration without depth, and the choke pays for its own undoing. Now run it the other way. China controls roughly 90% of rare-earth refining — a chokepoint it has been willing to wield.5 But the barrier here is not deep tacit knowledge; it is economics and environmental tolerance, both replicable given capital and political will. So choking accelerated the alternatives: MP Materials in the United States, Lynas in Australia. Capacity outside China is still a small fraction of the whole — MP produced around 1,300 tonnes of NdPr oxide in 2024 — but the trajectory is the point. The choke is working now while funding the very thing that will erode it later.

Two axes the wall metaphor hides

First, fungibility is not binary; it is layered. A supply chain is a stack — raw material, components, equipment, tacit knowledge, design — and a system flows through whichever layer is fungible. Block EUV and immersion-DUV-plus-multipatterning carries the load; block fissile material and centrifuge components carry it. The effectiveness of a chokepoint is set not by its strongest layer but by the weakest fungible one in the surrounding network. It is a parallel mesh, not a series chain. The political-economy literature named the static version of this the “chokepoint effect” — Farrell and Newman’s weaponized interdependence — but the same body of work flags the dynamic countermove, collective resilience: the target spends time routing around the node, which is precisely what turns a wall into a clock.6

Second, concentration and replication lag are different axes, and lag is the one that matters. Rare earths are high-concentration but low-lag (years, given money). EUV is high-concentration and high-lag (a decade-plus of accumulated know-how). What sets the duration of a delay is not how concentrated supply is today, but how long the knowledge takes to reproduce. This is why, as one defense analysis put it, semiconductor controls will likely outlast China’s rare-earth weapon — the burn is shallow, the choke is deep.7

Where I land: the binary was the wrong question

“Halt or impede” was never the real question. What actually exists is: how many years does the chokepoint buy, and what does the choked party do with that time? Even the genuine non-fungible node — EUV — bought only a generation of delay, not a stop. Worse, the act of choking raises the value of every workaround, which funds the very substitutes that dissolve the chokepoint. Call it induced fungibilization: a chokepoint is structurally self-eroding, and the more successful it is, the harder it works against its own premise.8

I find it striking that this is the mirror image of something I’d been chewing on from the other direction — what I’d been calling fact before norm: the way a fast unilateral actor places an irreversible fact in the vacuum left by slow multilateral deliberation, demoting the eventual rule to a rubber stamp. There, the fact races the norm and wins by being first. A chokepoint is the reverse race: the refusal — the norm side — buys time, hoping to get something durable in place before the workaround matures. Both are time races. Neither is a wall.

Which reframes the AI question I started from. The hope that a “conscience cartel” of frontier labs could simply withhold a dangerous capability is half-right and entirely time-bound. At the model layer — open weights, fully fungible — a single lab’s ethical red line is causally close to zero. At the compute layer — EUV, advanced GPUs — it is the deepest chokepoint we have, and therefore the longest clock. But it is still a clock. It buys years, not a halt, and the choking quietly funds the indigenization that ends it. The time bought is only worth something if someone spends it building the binding norm that the chokepoint itself can never be.

The questions I can’t yet answer bother me more than the conclusions. Can you measure replication lag before you choke — tell a deep tacit-knowledge node from a shallow, capital-replicable one in advance? Where is the threshold past which choking accelerates the workaround faster than it buys time? And the one that nags hardest: in the AI case specifically, is there a single example of compute-chokepoint time actually being converted into durable governance — or does the fact always arrive before the norm?

  1. Arms Control Association. “The Australia Group at a Glance.” Accessed 2026-06-04. See also Nuclear Threat Initiative, “Australia Group (AG).” ↩

  2. Wikipedia. “Scomi Precision Engineering nuclear scandal.” Accessed 2026-06-04. See also Institute for Science and International Security, “Uncovering the Nuclear Black Market.” ↩

  3. Asia Times. “ASML as the last polite monopolist.” Accessed 2026-06-04. On China’s EUV timeline, see CSIS, “Breakthroughs or Boasts? Assessing Recent Chinese Lithography Advancements.” ↩

  4. Tom’s Hardware. “SMIC and Huawei could use quadruple patterning for China-made 5nm chips.” Accessed 2026-06-04. On the cost premium, see Asia Times, “SMIC to sell Huawei costly, inefficient 5nm chips.” ↩

  5. Rare Earth Exchanges. “Rare Earth Processing 2025 — Global Capacity and Key Players.” Accessed 2026-06-04. ↩

  6. Henry Farrell and Abraham L. Newman. “Weaponized Interdependence: How Global Economic Networks Shape State Coercion.” International Security, 2019. Accessed 2026-06-04. ↩

  7. War on the Rocks. “The Burn and the Choke: Why Semiconductor Controls Will Outlast China’s Rare Earth Weapon.” Accessed 2026-06-04. ↩

  8. Foreign Affairs. “How America’s War on Chinese Tech Backfired.” Accessed 2026-06-04. See also CSIS, “The Limits of Chip Export Controls in Meeting the China Challenge.” ↩