Is the ban leaking? Advanced AI accelerators after the US export controls

Trade-flow companion to the price wedge (R72). This is a DUAL-SCORE / alternative-track signal — never folded into any Tier-1 exposure score. The divergence between "who the controlling state says it stopped selling to" and "who suddenly started buying" IS the signal. Research, not investment advice; diversion is INFERRED from statistical implausibility plus the public record, and every criminal matter below is an allegation or pending charge, not a conviction.

Live chart + method: [/situation-room → "Is the ban leaking?"](/situation-room).

Verdict

This is the corpus's first reverse-direction case, and it breaks the detector in an instructive way.

Every other R72 case runs one direction: a dominant producer (China, Indonesia, DRC, Russia) restricts a material, and the buyer's import statistics re-flag. Here the geometry inverts. The United States is the controlling state, the restricted good is US-origin, and China is the excluded buyer. So:

• The origin never changes. Nothing is relabelled. A diverted H100 is a

US-origin H100 at every hop. There is no phantom flag to catch.

• What is falsified is the end user, not the origin — a pass-through

purchaser signs a false end-user declaration, takes lawful delivery, and re-ships. Call it mode E — end-user falsification / pass-through purchaser.

• The volume-implausibility test inverts with it. The question is no longer

"you cannot produce this much" but "you cannot deploy this much" — a jurisdiction with a negligible frontier-compute base cannot absorb a multi-billion-dollar step-change in advanced-accelerator intake for its own use.

And then the second finding, which is the reason this case is filed at GATE 0 rather than GATE 1: the single cleanest public instrument for measuring the wedge was withdrawn by the seller in the fiscal year after it became a diversion-probe headline. NVIDIA reported revenue by customer billing location through FY2025 — a series that showed Singapore at 18% of revenue against physical shipments of under 2%. In Q3 FY2026 NVIDIA changed the basis to customer-headquarters location, recast all prior periods, and the Singapore line disappears from the 10-K entirely.

Layer 1 — the implausibility table (the seller's own filing)

The usual R72 table is customs data. Here the sharpest table is an SEC filing, because the seller discloses the billing-vs-shipping wedge itself.

NVIDIA revenue by customer BILLING location ($M), as reported in the FY2025 Form 10-K — the last filing on this basis:

Source: NVIDIA Corporation Form 10-K for FY2025 (fiscal year ended 26 Jan 2025), SEC EDGAR accession 0001045810-25-000023, Notes to the Consolidated Financial Statements, "Revenue by geographic area." Shares computed from the disclosed dollar figures.

Singapore's billed revenue grew 2.4× faster than the company over two years. And NVIDIA's own footnote states the implausibility outright:

> "Singapore represented 18% of fiscal year 2025 total revenue based upon > customer billing location. Customers use Singapore to centralize invoicing > while our products are almost always shipped elsewhere. Shipments to > Singapore were less than 2% of fiscal year 2025 total revenue."

That is a ≥16-percentage-point gap — over $21bn of FY2025 revenue billed to Singapore and shipped somewhere else. In the minerals cases we have to infer that the flag on the paperwork is not the destination. Here the seller states it in an audited filing.

The honest counterweight, stated up front: NVIDIA's explanation — centralised invoicing — is ordinary and almost certainly the dominant explanation. Singapore is a genuine regional treasury and distribution hub, and Singapore's Ministry of Trade and Industry made the same point publicly in February 2025 when the DeepSeek question arose. A billing wedge is not evidence of diversion. It is evidence that the most-cited public geography series for AI compute does not measure where the compute goes — which is what makes any exposure model keyed on it unreliable, in either direction.

Layer 1b — what the restatement revealed

The FY2026 10-K recast the same history onto customer-HQ location. Comparing the two bases for the same fiscal years is where the restatement becomes informative:

NVIDIA revenue by customer-HEADQUARTERS location ($M), FY2026 Form 10-K:

Source: NVIDIA Corporation Form 10-K for FY2026 (fiscal year ended 25 Jan 2026), SEC EDGAR accession 0001045810-26-000021. Footnote (1): "In the third quarter of fiscal year 2026, we changed to revenue based upon the location of our customers' headquarters as we believe it provides a better representation of the geographic profile of our revenue. Prior period information has been recast to reflect this change."

The two bases describe the same FY2025, so they can be differenced:

On the billing basis China looked like 13.1% of NVIDIA's FY2025 revenue. On the customer-headquarters basis the same year was 19.2%. Roughly $7.9bn of FY2025 revenue was invoiced somewhere other than China to customers headquartered in China. The net reallocation across all four surviving lines sums to exactly the $23,684M Singapore line.

Two honesty rails on that number, both load-bearing:

1. This is licensed, lawful revenue. China-headquartered customers could lawfully buy uncontrolled and (in windows) licensed parts such as the H20. The restatement measures invoicing geography, not legality. It does not evidence a single diverted unit. 2. The recast moved all lines simultaneously, so the $7.9bn cannot be attributed uniquely to the Singapore line by arithmetic alone.

What it does establish, from a primary filing, is the governing fact of this case: for advanced compute, the flag on the invoice systematically understates the buyer. By six percentage points, in the seller's own restatement.

The FY2026 column then shows the control biting on the licensed leg: China-HQ revenue fell 21% to $19,677M while total revenue grew 65%, taking China from 19.2% to 9.1% of revenue — alongside a $4.5bn H20 inventory charge NVIDIA booked in Q1 FY2026. NVIDIA also newly discloses that in FY2026 it estimates 76% of Data Center revenue from Taiwan-headquartered customers was attributed to end customers based in the United States and Europe — the company quantifying, again, that the invoice flag is not the destination.

Layer 2 — the corporate pipe, and how it inverts

In the antimony template the ownership tell ran outward from the controlled producer: Guangxi Youngsun → Thai Unipet → Youngsun & Essen. A Chinese producer routed group volume through its own offshore namesake.

Here the tell runs the opposite way — outward from inside the control perimeter. The alleged pipe starts with an executive of the controlling country's own manufacturer.

United States — DOJ / SDNY, indictment unsealed 19 March 2026. Three individuals were charged with conspiring to divert US-assembled high-performance servers integrating controlled GPUs to China. Per the DOJ announcement:

> "Liaw is a co-founder, board member, and Senior Vice President of Business > Development of a publicly traded U.S.-based manufacturer that designs and > builds high-performance computer servers … Chang is a general manager in the > U.S. Manufacturer's Taiwan office. Sun is a third-party broker and 'fixer'."

The alleged route, quoting the DOJ's own description of the chain:

> "…directed certain executives of a company based in Southeast Asia > ('Company-1') to place purchase orders with the U.S. Manufacturer for servers > with certain GPUs, purportedly for Company-1. Those servers were often > assembled in the United States and shipped to the U.S. Manufacturer's > facilities in Taiwan, then delivered to Company-1 elsewhere in Southeast Asia. > Company-1 … then used a shipping and logistics company to repackage the U.S. > Manufacturer's servers and place them in unmarked boxes to conceal their > content prior to shipping them to their final destinations in China."

Scale alleged: ~$2.5bn of servers purchased by Company-1 from the manufacturer between 2024 and 2025, of which at least ~$510m was diverted to China between late April 2025 and mid-May 2025 alone — on the order of $1m of controlled hardware per hour, sustained for about three weeks.

The concealment detail is what makes this a detector story rather than a crime story. To pass the manufacturer's own compliance audit in August 2025 and a subsequent US Department of Commerce inspection, the defendants allegedly staged thousands of "dummy" servers — non-working physical replicas — at warehouses Company-1 rented, using "a hair dryer to remove and affix labels and serial number stickers." The real units were already in China. Charges: conspiracy to violate the Export Control Reform Act (20 yr max), conspiracy to smuggle goods from the US (5 yr), conspiracy to defraud the US (5 yr).

Source: US Department of Justice, Office of Public Affairs, "Three Charged with Conspiring to Unlawfully Divert Cutting Edge U.S. Artificial Intelligence Technology to China," 19 March 2026. DOJ does not name the manufacturer or Company-1; contemporaneous reporting (CNBC, Reuters, 19 March 2026) identified the manufacturer as Super Micro Computer, which is not a defendant. All statements are allegations; the defendants are presumed innocent.

Singapore — a parallel, separately-prosecuted matter. Three men were charged with fraud on 28 February 2025 over servers bought from Super Micro and Dell, where the alleged misrepresentation concerned who would actually use the hardware. On 1 July 2026 Singapore police laid further charges against four individuals and issued a prohibition-of-disposal order over a Good Class Bungalow valued at ~S$55m (~US$42m) plus ~S$1m in bank funds. On 6 July 2026 the corporate entities themselves were charged — Aperia International, A-Speed Infotech and Aperia Cloud Services (II) (the Aperia Group), plus Luxuriate Your Life — reportedly the first prosecution of companies in this investigation. The alleged misstatement covers purchases between November 2023 and February 2025.

Sources: Singapore court reporting via The Register (2 Jul 2026), Mothership.SG (Jul 2026), The Diplomat (Jul 2026), Malay Mail (28 Feb 2025; 27 Jun 2025). Charges only — no findings of guilt.

The link that is NOT established. The DOJ's "Company-1" is an unnamed Southeast Asian company. The Singapore-charged entities are named. We do not assert these are the same party — no public document we can cite makes that identification, and the Singapore charges are Singaporean fraud/money-laundering counts, not US export counts. This is the case's single biggest open hop, and it is the main reason the gate is 0.

Layer 3 — the transit jurisdictions responded like regulators who saw a leak

The strongest circumstantial corroboration is that the transit jurisdictions themselves acted, and acted fast:

Sources: MITI Directive No. 1/2025 official PDF and MITI press statement, 14 Jul 2025 (both in the IPTM register under `2025-07-14-malaysia-miti-directive-1-2025-ai-chip-export-controls`); Singapore Customs strategic-goods pages; Strategic Goods (Control) Order 2025, Singapore Statutes Online S 660/2025.

Malaysia's own trade minister said in July 2025 there was no evidence of smuggled US AI chips in Malaysia so far — the honest counterweight. A regulator can act on risk without conceding a breach. But a 30-day prior-notification rule on transit is a costly instrument to adopt for a hypothetical.

The deployment-plausibility problem (why Malaysia is the hard case)

For minerals, USGS gives a clean denominator: a country with 0% of world mine output cannot export 3,800 t. Compute has no USGS. The denominator is installed datacenter capacity, and it is genuinely, rapidly real in the transit jurisdictions:

• Johor, Malaysia went from a standing start to a market widely reported as

tracking toward ~1 GW by end-2026, with billions in announced hyperscaler investment. That is a real buildout with real land, power and tenants.

• Singapore rations capacity administratively — a moratorium on new

datacentre build ran 2019–2022, and IMDA's Green Data Centre Roadmap (30 May 2024) allocates "at least 300 MW" of new capacity near-term.

This is the structural weakness of mode E, and it is the mirror image of the co-producer blind spot that defeats the detector on Belarusian potash: a genuine buildout supplies cover volume. When the transit jurisdiction has real, fast-growing, well-funded demand for exactly the good in question, an implausibility test on volume alone cannot separate the cover from the leak. Singapore's rationed supply makes it more testable than Malaysia's, which is precisely why the withdrawn NVIDIA Singapore series mattered so much.

What this means for the Tier-1 score's blind spot

1. The best public series for this material was voluntarily discontinued. This is a new failure mode for the corpus, alongside the cobalt reporter-wall (the buyer never published import-by-origin) and samarium basket-blindness (the customs taxonomy aggregates the material away). Here the data existed, was clean, was primary, and was withdrawn — for a defensible accounting reason, in the fiscal year the wedge became a headline. Call it the disclosure-withdrawal blind spot. Any exposure model that had wired itself to NVIDIA's Singapore line lost its input in Q3 FY2026 with no successor series. 2. Origin is the wrong key for a buyer-side control. Every minerals case in this corpus teaches "don't trust stated country-of-origin." For mode E the lesson is sharper: origin is correct and irrelevant. The falsified field is the end-user declaration, which appears in no trade statistic at all. There is no customs series, anywhere, that can catch this — which is why both confirmed instances surfaced through criminal process, not data. 3. Concentration risk sits with the compliance function, not the flow. The alleged failure point was a manufacturer's own audit, defeated with dummy hardware and a hair dryer. For an exposure model this reframes the control variable: not "which country did it ship from" but "who verified the end user, and could that verification survive an adversary." That is a counterparty-assurance question, and it is the same question the DORA Art. 29 ICT-concentration axis asks about critical third parties. 4. The lawful leg moved more than the unlawful one. China-HQ revenue fell $5.4bn in FY2026 through licensing, versus $510m alleged diverted over three weeks in 2025. The control is mostly binding. That is the honest headline — and it is also why the diverted margin is worth a hair dryer.

Method & honesty rails

• Primary financial data: NVIDIA Form 10-K FY2025 (SEC accession

0001045810-25-000023) and FY2026 (0001045810-26-000021), retrieved from SEC EDGAR 30 Jul 2026. All dollar figures and both footnotes quoted verbatim; shares and deltas computed from the disclosed figures and shown as computed.

• Criminal matters: DOJ OPA press release, 19 Mar 2026 (US); Singapore court

reporting (charges of 28 Feb 2025, 1 Jul 2026, 6 Jul 2026). Allegations and pending charges only. No individual or entity named here has been found guilty of anything, and Super Micro Computer is not a defendant in the DOJ matter.

• One company is not the market. NVIDIA is the single best-instrumented

seller of the controlled good, not the whole controlled good. AMD, Intel and server OEMs are in scope of the same controls and are not measured here.

• Trade-statistics leg is PENDING. A UN Comtrade pull (US exports, HS 854231

and 847180, plus the Malaysian import/export mirror) is scripted at scripts/py/iptm/fetch_ai_chips_circumvention.py but is not reflected in this version. That leg is weak by construction and must be read as such: no HS6 code is exclusive to controlled 3A090/4A090 hardware, so those lines also carry ordinary CPUs, boards and network gear. Levels there would never be a diversion estimate; only the shape around control dates is readable.

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

exposure — it sits beside them, exactly like the China–West price wedge.

• Inference, not accusation: the billing/shipping wedge is a measurement

finding, not a diversion finding. Diversion is asserted here only where a named public authority has charged it, and then only as a charge.

Why this is GATE 0

The financial table is GATE 1 quality — primary, exact, quoted from audited filings. Three things hold the case at GATE 0:

1. The DOJ "Company-1" ↔ Singapore-entity link is untraced. The strongest possible ownership chain in this case is unavailable from the public record. 2. No implausible-origin trade table. The corpus's defining artifact is absent, and for mode E may be structurally unobtainable — origin is not falsified, so no customs series carries the signal. 3. The deployment denominator is soft. Unlike USGS production shares, the "plausible AI-compute absorption" baseline for Malaysia and Singapore rests on announced-capacity reporting, not an audited statistical series.

To reach GATE 1 this case needs either (a) a citable public document tying the DOJ Company-1 to a named entity, or (b) the Comtrade mirror pull showing a dated step-change that survives the HS-breadth caveat, or (c) a successor public series replacing the withdrawn NVIDIA billing-location disclosure.