Is the ban leaking? Chipmaking equipment after the trilateral licence regime

Trade-flow companion to the price wedge (R72). Every prior entry in this corpus asked the same question of a molecule: did the controlled origin's line collapse, and did a country that cannot produce the material surge into the gap? This case asks it of a machine — and gets a result that inverts the method. DUAL-SCORE / alternative-track only; never folded into any Tier-1 exposure score. Research, not investment advice. Nothing here asserts that any individual shipment was unlawful.

This is the structural pair to the trilateral chip-equipment perimeter case, which mapped the instruments. This one measures whether they hold.

Verdict

The volume-implausibility detector — layer 1 of the standard method — returns nothing here, and its silence is not evidence that the perimeter is holding. It is evidence that the detector cannot see this chokepoint at all. Three independent reasons, each visible in the trade data:

1. The controlled origins' direct lines to China went *up*, not down. Dutch exports of chipmaking machines to China are +264% from the pre-control year; Japan's +32%. A node-specific licence leaves the tier just below the threshold fully saleable, so "the direct line collapsed" — the antimony signature — never forms. 2. The obvious transit hubs are genuine manufacturers. Singapore accounts for roughly 20% of world semiconductor-equipment production (Singapore EDB). A Singapore surge is plausible on its face, so the ~0%-of-world-capacity test that convicted Thailand on antimony has no purchase. 3. The one place the fingerprint does appear, the buyer's own statistics do not record. Hong Kong reports re-exporting $5.87bn of chipmaking machines to mainland China in 2025. China reports importing $20m from Hong Kong. That is a two-orders-of-magnitude mirror gap on a $3bn+/yr channel.

The working detector for equipment is therefore not the origin field. It is the end-user field — and the only public instrument that reads it is the enforcement docket. In September 2025 BIS named two Chinese companies that had bought US-origin chipmaking equipment for Entity-Listed SMIC fabs, and revoked the entity-keyed authorisation it called a "loophole" for three foreign-owned fabs inside China. Origin true, destination true, buyer false.

Layer 1a — the controlled lines rose

Exporter-reported shipments of chipmaking machines (HS 8486) to China, USD m:

Source: UN Comtrade free public preview API, annual, HS 8486, each row as reported by the named exporter. Committed artifact `data/intelligence/semi-equipment-circumvention.json`. 2020 omitted (COVID distortion); China had not yet reported 2025 annual at pull time.

Two readings matter. First, only the US line fell — and it fell hardest in 2025, after the December-2024 package and alongside the VEU revocation. Second, the Dutch line tripled through the licence regime. Read naively that says the Dutch control failed. It says something more specific: the control is keyed to a node, the buyer bought everything underneath it, and each announced expansion (Sept 2024, Jan 2025) gave the buyer a dated deadline to buy against. A rising direct line is the expected signature of a threshold licence, not evidence of leakage — and a falling one is not proof of a bite.

Layer 1b — where the control actually bit: parts, not tools

China's own reported imports, USD m:

Source: UN Comtrade, reporter China (156), flow imports, HS 8486 / 848690.

Whole-tool imports grew 36% while the spares line shrank 17%, and the US spares line halved (−61% from its 2021 peak). That composition inversion is the mirror-image of the plywood tell: there, China's exports flipped from finished good to input; here the buyer keeps buying machines and loses the ability to keep them running on US parts and US-person servicing. For a fab, a tool without a spares and service pipe is a depreciating asset with a fixed lifespan. The binding constraint the trade data actually shows is uptime, not acquisition — and no origin-relabelling scheme fixes that, because what is restricted is a support relationship, not a molecule.

Layer 1c — the Hong Kong mirror gap

The one implausible-origin fingerprint in the dataset, USD m, HS 8486:

Source: UN Comtrade, reporter Hong Kong (344) flows X and M, and reporter China (156) flow M, HS 8486.

Hong Kong is ~99% a re-export economy — re-exports were HK$5,174.2bn of HK$5,240.3bn total exports in 2025, i.e. 98.7% (HK Trade & Industry Dept; C&SD, 27-Jan-2026). It has exactly one significant equipment manufacturer — ASMPT's Tsing Yi back-end plant — and ASMPT's entire global FY2025 revenue was HK$13.7bn ≈ US$1.8bn (ASMPT financial information), about one-third of the US$5.87bn that left Hong Kong for the mainland in 2025 under this heading. Hong Kong manufacture cannot account for the line; it is overwhelmingly tools made elsewhere, passing through.

And China's books do not show it arriving. The likeliest explanation is mundane and is itself the finding: China records imports on a country-of-origin basis while Hong Kong records exports on a country-of-consignment basis, so a Singapore-built tool trans-shipped through Hong Kong is booked to Singapore by the buyer and to Hong Kong by the transit territory. If that is right, then on the importer's own statistics the trans-shipment leg is invisible by construction — not concealed, just unrepresentable. That is a far more general warning than any single laundered cargo: the mirror line, which this corpus has leaned on repeatedly, cannot detect a transit hop when the importer reports on an origin basis.

One anomaly I cannot resolve and will not explain away: Hong Kong's 2025 exports of 8486 to the mainland (5,871) exceed its total 8486 imports from the entire world (4,109), and the same pattern holds in 2022 (2,109 vs 1,178). Candidate causes — re-export margin in the export valuation, bonded-warehouse timing, HS reclassification between entry and exit, or round-tripping of Chinese-origin goods (HK's imports from the mainland ran 2,233 in 2025) — are not separable with public data. Flagged as an open question, not a finding.

Layer 2 — the end-user tell (this is the whole signal)

With layer 1 structurally blind, everything rests on the entity layer. For equipment the falsified field is not the origin — it is the consignee.

The front-buyer designation. Effective 12 September 2025 (published 16-Sep-2025), BIS added 32 entities to the Entity List, 23 of them in China. Among them, in the agency's own words:

> "These entities acquired U.S. origin semiconductor manufacturing equipment for > two Entity List parties, SMIC Northern Integrated Circuit Manufacturing > (Beijing) Co., Ltd. and Semiconductor Manufacturing International > (Beijing) Corporation, without the requisite license or authorization from > BIS."

The two named acquirers are GMC Semiconductor Technology (Wuxi) Co., Ltd. and Jicun Semiconductor Technology (Shanghai) Co., Ltd., both now under a presumption of denial for all items subject to the EAR. Source: Federal Register, "Additions and Revisions to the Entity List", 16-Sep-2025, via [govinfo.gov](https://www.govinfo.gov/content/pkg/FR-2025-09-16/html/2025-17893.htm) (federalregister.gov is unreachable from this host; govinfo is the official mirror).

Note precisely what this is and is not. It is a procurement-front relationship asserted by the regulator — a domestic buyer of record standing in for a listed fab. It is not a common-ownership chain of the Youngsun → Thai Unipet → Youngsun & Essen kind, and I did not trace one in any registry. On this corpus's own scoring, the common-ownership tell here is N.

The authorisation loophole. On 2 September 2025 (effective 31-Dec-2025) BIS removed Intel Semiconductor (Dalian), Samsung China Semiconductor and SK hynix Semiconductor (China) from the Validated End-User list, describing the VEU programme as a "loophole" that let those fabs receive most US-origin chipmaking equipment, software and technology licence-free. VEU was keyed to the entity, not the country — so for three years the largest lawful channel for advanced US tooling into China was an end-user authorisation, and closing it required an end-user instrument. Country-pair trade data could never have shown either the opening or the closing.

Mode E, third instance — the domestic-front variant

This is not a new mode. It is the third row of mode E (end-user falsification / pass-through purchaser), defined by the advanced AI accelerators case and made GATE 1 by the dual-use microelectronics → Russia case: the control is on the buyer, the restricted good is the controlling state's own export, origin stays correct at every hop and is simply irrelevant, and what is falsified is the end-user declaration — a field that appears in no customs series anywhere. (Mode G already established that a mode can blind all three detector layers, so blindness here is not novel either.)

What is new is the geometry of the pass-through. In both earlier instances the purchaser sits in a third country, which is why mode E carries a detectability rule keyed to that country's baseline demand — auditable in Kyrgyzstan (€4,655 of EU lathes in 2021, nowhere for €4.0m to hide), invisible in Singapore. Here the front buyer sits inside the destination country: GMC Wuxi and Jicun Shanghai are Chinese companies buying for Chinese fabs. There is no pass-through jurisdiction, therefore no baseline denominator, therefore the mode-E detectability rule does not apply at all — it is not that the signal-to-noise is bad, it is that the ratio is undefined. Check whether a pass-through jurisdiction exists before reaching for the rule.

The VEU leg adds a second wrinkle that stretches the mode's own definition: for three years the largest channel for advanced US tooling into China involved no falsification whatsoever. VEU was a lawful, entity-keyed authorisation, and BIS closed it by revocation rather than by enforcement. Mode E's diagnostic field — the end-user — can therefore leak without anyone lying, which is closer to C⁰'s lawful-transformation logic than to a false declaration.

Three transferable rules:

1. Match the detector to the falsified field. Ask which field the evasion has to lie in before running the test. Where the control is keyed to an end user (VEU, Entity List, FDPR end-use), no country-pair statistic can ever fire, and an absence of implausible-origin surge means nothing at all. 2. The manufacturing-footprint blind spot. The co-producer blind spot found on Belarusian potash generalises: when the transit hub genuinely makes the controlled item (Singapore ≈20% of world equipment output; Malaysia's Penang cluster — US shipments to Malaysia +112% 2022→2024), the volume-plausibility test is defeated lawfully. For equipment, the technology's nationality and the shipment's origin have simply decoupled. 3. On a threshold licence, read composition, not level. The direction of the direct line carries no information. What carried information here was the split between machines (+36%) and spares (−17%, US −48%), which located the real bite on servicing rather than acquisition.

Why it matters for the buyer

1. An exposure model keyed on country-of-origin cannot see this chokepoint. Not "sees it imperfectly" — cannot see it. A risk team tracking "Chinese imports of Dutch lithography" would have read the 2023–24 tripling as a control failure and the 2025 dip as a control success; both readings are artefacts of a node threshold and a stockpiling deadline. 2. The exposure that matters is a service relationship, not a shipment. Fabs holding tools they cannot license parts or engineers for carry a dated, depreciating capability. That is a balance-sheet fact invisible to any trade feed. 3. Enforcement risk concentrates on the entity, so entity screening is the control. The buyers convicted by the public record here were domestic Chinese companies with clean-looking paperwork, not exotic offshore shells. Screening the origin country of a tool tells you nothing; screening the counterparty against the Entity List and VEU history tells you almost everything.

Method & honesty rails

• Trade data: UN Comtrade free public preview API (comtradeapi.un.org/public/v1/preview),

annual, HS 8486 (semiconductor/FPD manufacturing machines) and HS 848690 (parts). Reporters pulled: CN (imports), NL, JP, US, SG, MY, KR, HK (exports). Values are customs value in USD; net weight is meaningless for machinery and is not used. Artifact: data/intelligence/semi-equipment-circumvention.json.

• Known data defect, carried openly: the committed fetcher

scripts/py/iptm/fetch_semi_equipment_circumvention.py has an M49 partner map that silently drops partner codes 842 (USA), 757 (Switzerland) and 251 (France) — the US is China's #4 supplier of chipmaking machines ($4.5bn in 2024), so the omission is material and would have produced a false "the US is absent" finding. The tables above were built with a corrected map that preserves unmapped codes as M49_<code> instead of discarding them. The fetcher itself is still unfixed (out of this wake's write scope) — fix the map before re-running it.

• Malaysia's reported HS 8486 world total is unusable in Comtrade for

2019–2024 (it reads below individual partner lines); only Malaysia's partner-level rows and other reporters' Malaysia rows are used here.

• Attribution is explicitly unverified. That Singapore's and Malaysia's

China-bound equipment exports are substantially the output of US- and EU-headquartered toolmakers manufacturing there is consistent with those firms' documented footprints, but no per-shipment attribution is claimed.

• Alternative-track only: this never touches buyerRelativeScore or any base

exposure score. It sits beside them, like the China–West price wedge.

• Inference, not accusation: the only diversion asserted is the one the

Federal Register asserts, in its words, about the entities it names. No individual shipment is characterised as unlawful, and the Hong Kong mirror gap is presented as most likely a reporting-basis artefact, not smuggling.

• Gate: GATE 1 — every figure in every table traces to UN Comtrade, the HK

re-export share to HK TID/C&SD, the Singapore production share to Singapore EDB, ASMPT's revenue to its own financial statements, and the entity findings to the Federal Register. The common-ownership tell is honestly N; the layer-2 evidence is an enforcement designation, not a registry-traced ownership chain.