How Export Controls Work, and Where They Leak

A chip designed in California, fabricated in Taiwan on machines built in the Netherlands, and sold to a data centre in Malaysia may never touch American soil. Understanding how export controls work means understanding why Washington can still stop that sale.

Export controls are licensing rules. A government designates certain goods, software and technical knowledge as controlled, then requires a licence before any of it reaches particular destinations, buyers or end uses. The licence is the entire instrument. Refusing one is how a state blocks a transaction it has no other way to reach.

That sounds narrow. In practice it has become the most far-reaching economic tool any government currently operates, and the least understood.

What counts as an export

The word does most of the damage. In ordinary speech an export is a physical thing crossing a border, and almost everyone reading a news story about export controls imagines a shipping container.

The regulations mean something much wider. Technical data counts. Software counts. So does a conversation.

The clearest illustration is the deemed export. If a company inside the United States gives a foreign national access to controlled technology — a schematic, a process specification, source code — the government treats that disclosure as an export to that person’s home country, and the same licensing analysis applies as if the technology had been shipped there.

This is why university research offices, not just weapons manufacturers, spend real money on compliance. A lab hiring a doctoral student can trigger the rules without anything leaving the building.

Three agencies, three rulebooks

American export control is not one system. It is three, with different statutes, different lists and different cultures, and the first question in any real case is which one applies.

The three overlap constantly, and a single transaction can sit under all of them at once. The overlap with the programmes run by Treasury’s Office of Foreign Assets Control matters most, because the two instruments are usually deployed together and are routinely confused in coverage — though they work differently, as the mechanics of how international sanctions work make clear.

Penalties are not nominal. Criminal violations carry fines reaching $1 million and up to 20 years in prison, counted per violation rather than per case.

How an item becomes controlled

Nothing is controlled in the abstract. An item is assigned a classification number on a control list, and that number is then read against four things at once: what the item is, where it is going, who will receive it, and what it will be used for.

Change any one of those and the answer changes. The same oscilloscope can ship freely to Germany, require a licence for one buyer in a third country, and be flatly prohibited for another.

This four-way structure is what makes the system flexible enough to be used as foreign policy. Governments rarely need to ban a technology outright. They adjust the threshold — a performance ceiling on a chip, a resolution limit on a sensor — and the same rule now captures a different slice of the market.

It is also why compliance is expensive and why small exporters get caught. The classification is the exporter’s responsibility, not the government’s.

The Entity List names names

Alongside the item lists sits a list of parties. The Entity List, maintained by BIS, identifies companies, institutes and individuals the US government has determined are acting against its national security or foreign policy interests.

Landing on it does not ban trade outright. It imposes a licence requirement for dealings that would otherwise be unrestricted, and most listings carry a presumption of denial — meaning the licence exists on paper and is not going to be granted.

The obvious workaround is a subsidiary. Set up a differently named company, and the parent’s designation does not formally attach to it.

Regulators closed much of that route in 2025, when BIS extended restrictions to entities that are at least 50% owned by listed parties, individually or in aggregate. The rule matters less for what it prohibits than for what it signals: the agency now expects exporters to investigate ownership structures rather than check a name against a list.

How the rules reach factories outside America

Here is the mechanism that most coverage skips, and it is the one that makes the whole system unusual. Ordinarily a country’s export rules govern goods leaving that country. The Foreign Direct Product Rule does something else.

If controlled American technology or software was used to design an item, or if the plant producing it runs on major equipment that is itself a direct product of American technology, then the finished item falls under US jurisdiction — regardless of where it was made or by whom.

It is a provenance test. It asks what made the product, not what is inside it. The practical effect is that a factory in another country, owned by a company with no American presence, can be barred from selling to a particular buyer because of the lineage of its tooling. No American component needs to be present in the finished chip.

The rule is not new — it dates to 1959 — but for most of its life it was a technicality invoked rarely. It was expanded dramatically in 2020 to cut Huawei off from foreign-made chips, and again through 2022 and 2024 to restrict China’s access to advanced semiconductors and the equipment that makes them.

That expansion is the reason export controls now function as an instrument of great-power competition rather than a nonproliferation housekeeping tool. It converted a national licensing system into something with global reach.

Allies are asked, not ordered

Reach is not the same as control, and the limit shows up quickly when other governments are involved.

The main multilateral forum for dual-use goods is the Wassenaar Arrangement, founded in 1996 and now counting 42 participating states. It is worth being precise about what it is: not a treaty, not legally binding, and not an enforcement body. Participants exchange information about their national licensing decisions and maintain broadly similar control lists.

Every actual licence is still issued by a national authority answering to a national government.

Semiconductor tooling is where this bites. The machines for extreme ultraviolet lithography — required for the most advanced chips — are built by a single Dutch company, which means the restriction only holds while the Netherlands agrees to enforce it. Washington can persuade and it can apply the provenance rule, but it cannot legislate for The Hague.

Alignment among allies is therefore a political variable, renegotiated with each change of government, and it is the softest joint in the structure.

Where the system leaks

Controls on physical goods are only as good as the enforcement at the far end of a supply chain that nobody fully sees.

The routine failure is transshipment. Controlled items are sold legitimately into a country that is not itself restricted, then re-exported to a buyer who is. Singapore, Malaysia and Thailand recur in these cases, not because their governments sanction the trade but because they are large, well-connected trading hubs where volume provides cover.

The mechanics are unglamorous, and they repeat across cases:

• Shell companies with no operating history placing orders sized to look ordinary
• Falsified end-user statements, sometimes supported by staged compliance audits
• Intermediaries who buy legally, take delivery, and re-sell across a border

Two unrelated smuggling cases involving American AI chips surfaced in March 2026, and both followed that pattern. Enforcement works after the fact, through prosecutions and new designations, which means each success is also evidence that the shipment already happened.

Nobody credibly claims the leakage is total. The honest framing is that controls raise cost and delay rather than build a wall, and the policy argument is about whether the delay is long enough to matter.

Substitution is the slower leak

The more consequential failure mode is not smuggling. It is that a restriction gives the restricted party a reason to build its own.

SMIC, China’s largest foundry, could not buy EUV machines, so it pushed older deep-ultraviolet equipment further than the industry expected, reaching 7-nanometre production through repeated patterning steps. The result costs more per wafer and yields worse. It also exists, which the controls were meant to prevent.

This is the central tension in the policy and the reason serious analysts disagree about it. Controls buy time; they also convert a customer into a competitor with state backing and a guaranteed domestic market.

Which effect dominates depends on how long the lead being defended actually is, and that is not something anyone can measure in advance. The demand driving all of it — the enormous compute build-out behind modern AI — is also reshaping the physical infrastructure underneath, from data centres to the undersea cables carrying the traffic.

What to watch when the next package lands

Announcements are written to sound decisive and rarely explain their own mechanics. Three questions cut through most of it. First, does the measure name items or parties? Item thresholds catch a whole market segment and get engineered around; party designations are precise and get restructured around.

Second, does it rely on the provenance rule to reach foreign factories, and if so, have the governments hosting those factories agreed to enforce it? Third, what is the stated licence policy — a case-by-case review and a presumption of denial are very different instruments wearing the same word.

If a package tightens thresholds without addressing transshipment, it is aimed at the legal market and will not touch the illegal one.