Checked 2026-10-04 · unchanged since 2026-08-21RSS
Export Controls
Diversion Pipeline Tracker
Export controls on AI chips and computing infrastructure, with designation records and semiconductor trade data.
3 of 11 company relationships in Culper’s Nvidia diversion report have primary-source citations
Licensed corporate records support one additional relationship.
Alleged chain · evidence grade
3
Alleged relationships with primary-source citations, across 7 records in corporate registries and a DoJ indictment
8
Alleged relationships supported here by a single Tier-3 citation or subscription-only corroboration. Drawn in amber. Primary-source evidence only fades these and the 8 firms that have no corroborated connector, leaving 3 connectors in view
8
Distinct relationship kinds across the 11 connectors — ownership, rename, financing, trade flow, lease, and director clustering among them
20
Evidence records behind the figure, each with its own source, authority tier and vintage. A connector’s tooltip gives the strongest tier among its records and the source at that tier
The alleged OEM → intermediary → end-customer chain: one connector per claimed relationship, over 14 organizations.
Relationships are grouped by the tier they lead into. An open tick marks a firm with no upstream relationship in this case.
Firms are placed by chain position, and 2 connectors skip a position.
Worked example Culper Research’s 13 May 2026 short thesis alleges that Nvidia’s China-linked business continued through chip diversion and access to servers in Southeast Asia, including Aivres shipments to Speedmatrix. Culper disclosed a short position in Nvidia. The graph separates ownership and financing records from reported shipments and alleged customer access, with sources for each relationship. A documented corporate relationship does not establish where chips went or who used them.
IEIT Systems A is among the organisations this worked example covers. The panels below show the chain alleged in Culper’s 2026-05-13 report and are the same whichever organisation is selected.
4 of 11 relationships have primary-source citations or subscription-source corroboration. For the other 7, the only recorded evidence is from Culper’s report or the sources it quotes. Search attempts are not recorded.
Assuming all $4.74B of server/product invoice value purchases GPUs, the selected spending shares and assumed card prices give the counts below. Peak throughput uses BF16 dense arithmetic; effective throughput applies 40% model FLOPs utilization.
Sensitivity controls
$1.0Bstarting $4.742B$10.0B
0.70× (30% lower)1.00× baseline1.30× (30% higher)
Comparison anchors
—Comparison withheld. MIIT’s October 2023 plan targets above 300 EFLOPS of total computing power and a 35% intelligent-computing share by 2025. Its unspecified precision and aggregation conventions prevent comparison with BF16 dense throughput.Tier unavailable
~208Annual compute in units of the EU AI Act Article 51 training threshold of 10²⁵ FLOPs (1.75 days per unit at the selected utilization). This assumes a full year of operation at that rate.T4
~4.1%Trade-value scale relative to NVIDIA’s approximately $115B FY2025 Data Center revenue. Server/product invoices cover a different period and include costs beyond NVIDIA components.T4
Starting assumptionsThe $4,742M starting value exceeds the named flows’ $4,486M subtotal by $256M; the difference remains unreconciled. Culper’s trade reporting is Tier 3; the spending mix, prices and utilization are Scrutica assumptions (Tier 4). NVIDIA specifications support H100, H200 and HGX B300 throughput; H800 uses secondary reporting.
The $4,742M starting assumption exceeds the three named flows’ $4,486M subtotal by $256M; reconciliation requires the underlying trade rows.
Culper reports server/product invoices. Converting their entire value at bare-GPU prices assumes all that value purchases GPUs.
GPU spending shares, prices and model FLOPs utilization are scenario assumptions. Shipment-level composition and transaction prices have not been verified.
H100, H200 and H800 use the same BF16 dense throughput in this calculation. The lower assumed H800 price buys more GPUs for a given spend.
NVIDIA specifications support the H100, H200 and HGX B300 values; H800 throughput is retained from secondary reporting.
Annual compute assumes simultaneous operation of the calculated GPUs for 365.25 days at the selected utilization. Installation, uptime and connectivity remain unknown.
Cascade graph
The same 11 alleged relationships, with their source citations. The “Primary-source only” switch highlights relationships supported by citations to corporate-registry or government records.
Channel taxonomy
Each of the 11 relationships sits in one corroboration-status × chain-tier cell and in one relationship-kind row.
Relationship-kind breakdown · ordered by primary-source share
Financing via pledged charge1
1 primary-source corroborated
Corporate rename1
1 primary-source corroborated
Shared-director clustering1
1 primary-source corroborated
Trade flow (unilateral)3
3 Tradesparq via Culper
Lease through intermediary2
2 WSJ via Culper
Wholly-owned subsidiary1
1 subscription-source corroborated
Cloud-capacity lease (alleged)1
1 Culper narrative only
Trade flow · optical components1
1 Tradesparq via Culper
The financing via pledged charge, corporate rename and shared-director clustering relationships have 7 primary-source citations. The wholly-owned subsidiary relationship is corroborated against subscription-only corporate records; their contents are withheld. 7 relationships each rest on a single Tier-3 citation.
Speedmatrix Malaysia Statement of Charges filed 18 June 2024 (Malaysian SSM)
Megaspeed International Pte Ltd 2024 annual report (Singapore Bizfile)
Singapore ACRA Bizfile — Megaspeed International Pte. Ltd. (UEN 201535940Z)
Hong Kong corporate filings (OBON BVI director list)
Singapore Bizfile — Siam AI Corporation Pte. Ltd. officer listings
March 2026 DoJ Supermicro indictment + May 2026 Bloomberg identification
Enforcement-shock scenarios
Enforcement-shock cascade · downstream propagation through 3,416 supply-chain nodes
Narrow: Supermicro's OEM channel cut off, standing in for the Wally Liaw indictment
The corridor named in the March 2026 DoJ indictment, severed at the Supermicro OEM node.
Affected nodesT4
50
1.46% of 3,416
Weighted compute impactT4
3.21%
max depth 1
SeverityT4
100%
at 1 input node
PropagationT4
0.85 decay
downstream (toward customers)
Weighted impact by supply-chain tier
L0 Materials/Foundry · 1 node
1.510%
L3 Systems/OEM · 1 node
0.062%
L4 Cloud/End-User · 48 nodes
1.637%
Top affected nodes by tier · per-node impact = severity × decay^depth
L0 Materials/Foundry
Super Micro Computer, Inc.100.0%
L3 Systems/OEM
Applied Digital16.8%
L4 Cloud/End-User
Compuware Technology, Inc.20.4%
WhiteFiber18.0%
One Stop Systems15.8%
Penguin Solutions14.6%
Logicom14.4%
Shinden Hightex13.7%
+4 more on record
Scenarios including relationships alleged by Culper
These scenarios include the additional relationships alleged by Culper among Aivres, Speedmatrix, Megaspeed, Novagate, Aolani and their customers. Results depend on those allegations; the additional relationships are absent from the base supply-chain graph.
Aivres channel cut off, with the relationships Culper alleges downstream of it
A hypothetical restriction on the Aivres channel propagates through the relationships alleged by Culper. Alibaba, IFLYTEK, PT Indosat and Opera are reached within four hops in this graph.
Affected nodes
16 · 0.47% of 3,428
Weighted impact
0.438%
Max depth
4
Aivres alone cut off, including Culper’s alleged relationships
Only Aivres is directly restricted in this scenario. The projected effects extend two hops via Speedmatrix and Aolani to ByteDance and Alibaba, using relationships alleged by Culper.
Affected nodes
6 · 0.18% of 3,428
Weighted impact
0.117%
Max depth
2
Weighted impact = per-node severity × decay^depth, normalized by graph-level node weighting. The 14 Culper-alleged edges are the one Tier-3 input, and they enter the model only while Include Tier-3 Culper edges is switched on.
4 enforcement scenarios run on Scrutica’s supply-chain graph, which is built from a database held under subscription and not redistributed: Supermicro’s OEM channel severed, standing in for the March 2026 DoJ indictment; Nvidia’s flow to Southeast Asian server makers cut by 30%; Aivres treated as inheriting Inspur Group’s Entity List designation; and a full US-China decoupling as a comparator. Each reports its severity, decay rate and direction, the weighted impact by supply-chain tier, and the firms hit hardest. A companion graph adds the 14 further relationships alleged only in Culper’s report.
Cross-source disclosure tensions
Company disclosures and government records compared with Culper’s interpretations. 2 of 8 recorded comparisons are published here; each source retains its own authority tier. Subscription-only evidence is identified, with its values withheld.
Geographic-revenue methodology change within the disclosure window
Nvidia changed its geographic-revenue methodology from billing location to customer headquarters location in Q3 FY2026, which removed Singapore from the disclosed geographic split. Singapore-billed revenue had risen before the change. The amounts were checked against licensed records and are not reproduced here.
T1Primary-source evidence
Nvidia stated change reflects 'better representation of the geographic profile of our revenue' — companies' HQ location now overrides Singapore-based billing
Corroborated against subscription corporate-data sources that cannot be republished (1 record, values withheld). See how those records are held and audited.
T3Analyst interpretation
Culper argues that Singapore would remain a material revenue exposure under the earlier billing-location method. Figures before and after the change use different geographic definitions.
China revenue growth during tighter export controls
Supermicro's Mainland China revenue rose over a period spanning the April 2025 export-control tightening. The amounts and growth rate were checked against licensed records and are not reproduced here. The March 2026 DoJ indictment alleges that a front company — since identified by Bloomberg as OBON Corp — bought approximately $2.5B of servers from an unnamed US manufacturer, of which at least approximately $510M was diverted to China.
T1Primary-source evidence
USD 2.5B of servers bought via 'Company-1' (Bloomberg-identified as OBON Corp), part of which the indictment alleges was diverted to China
DoJ SDNY indictment of Wally Liaw (March 2026)· 2026-03
Corroborated against subscription corporate-data sources that cannot be republished (1 record, values withheld). See how those records are held and audited.
T3Analyst interpretation
Culper alleges additional channels through manufacturers and intermediaries beyond the transactions described in the indictment.
PeeringDB facility-presence corroboration
PeeringDB lists facilities where networks report having equipment. Co-location in this directory does not establish interconnection, BGP sessions, traffic or tenancy.
Lead facility-presence record
True IDC - East Bangna, Thailand · Huawei Cloud Global (AS136907) declared present
True IDC (Charoen Pokphand Group / True Corporation subsidiary)·Internet exchange: SYMPHONY THAI - IX BANGKOK - 1
Culper relevanceCulper’s report (p. 38) describes a December 2024 memorandum of understanding between True IDC and Siam AI to “establish Thailand as the Regional AI data center hub”.
4 of the 12 other Southeast Asian facilities with a Chinese network declared present
One representative facility each in Thailand, Malaysia, Indonesia and Singapore at which a selected Chinese hyperscaler or PRC state-telecom network declares a presence. A declared presence does not establish GPU shipments or an export-control violation.
Tier 2 · research
Facility
Operator
Declared Chinese networks
STT Bangkok 1
ST Telemedia (Singapore-Thai joint venture)
Alibaba (Aliyun) AS45102Huawei Cloud Global AS136907
NTT Cyberjaya Data Center (CBJ)
NTT Global
Alibaba (Aliyun) AS45102
DCI Indonesia (JK5)
DCI Indonesia (Salim Group)
Huawei Cloud Global AS136907ByteDance / TikTok AS396986
China Mobile International - Singapore DC
—
China Mobile International AS58453China Mobile International - NII AS58807
3 checks without a selected-network presence record
Culper-named sites checked: YTL Green Data Center / YTL AI Cloud / YTL Communications · Equinix JH1 - Johor · Singapore Equinix SG1-SG5 cluster.
A check without a presence record either matched a listed facility whose participants name no selected Chinese network, or found no listed facility to match. Neither outcome rules a presence in or out. Of 105 Southeast Asian facilities scanned, 13 have a selected Chinese hyperscaler or PRC state-telecom network declared present. Source: PeeringDB, analyzed 2026-05-14.
Evidence timeline
7 events from Inspur Group’s March 2023 Entity List designation to Senator Warren’s June 2026 letter to Nvidia, with a source for each. 2 additional events are omitted from this timeline.
2 Mar 2023Regulatory anchorT1
Inspur Group added to BIS Entity List
The Bureau of Industry and Security designates Inspur Group Co., Ltd. on the Entity List “for acquiring and attempting to acquire U.S.-origin items in support of China’s military modernization efforts.”
Inspur Systems Inc. renames to Aivres Systems Inc.
A California Secretary of State business filing records the name change from Inspur Systems Inc. to Aivres Systems Inc., about two months after the BIS designation of Inspur Group.
Six Inspur subsidiaries added to the Entity List — including the A-share issuer that owns Aivres
BIS adds six Inspur Group subsidiaries: Inspur (Beijing) Electronic Information Industry Co., Ltd.; Inspur Electronic Information Industry Co., Ltd.; Inspur Electronic Information (Hong Kong) Co., Ltd.; Inspur (HK) Electronics Co., Ltd.; and Inspur Software Co., Ltd. under China, plus Inspur Taiwan, all for all items subject to the EAR, under a policy of denial, with a Footnote 4 designation. The second of those is the Shenzhen-listed issuer that owns the US OEM Aivres Systems Inc., recorded in the Federal Register under the aliases “Inspur Information” and “IEIT Systems Co. Ltd.” BIS designated that intermediate itself, two years after the parent.
Nvidia’s Q3 FY2026 10-Q discloses a shift in geographic-revenue methodology, from billing-location-of-customer to customer-headquarters-location, because “the end customer and shipping location may be different from customer’s billing location.” Singapore drops out of the disclosed geographic split and the United States share rises to ~70%.
DoJ unseals indictment of three individuals · ≥$510M in AI servers diverted to China
The U.S. Department of Justice unseals an indictment of Yih-Shyan (“Wally”) Liaw, Ruei-Tsang (“Steven”) Chang, and Ting-Wei (“Willy”) Sun for allegedly conspiring to divert high-performance AI servers, assembled in the United States, to China in violation of U.S. export-control law. Per the DoJ release: a front company purchased approximately $2.5B of servers from the U.S. manufacturer (unnamed in the release) across 2024–2025, and at least approximately $510M worth was diverted to China between late April and mid-May 2025 alone, using false documents, staged dummy servers, and transshipment. Senator Warren’s June 1, 2026 letter to Nvidia characterizes the case as involving a Supermicro co-founder and “$510 million in diverted servers.”
Culper Research publishes its NVDA short thesis, synthesizing the Inspur / Aivres / Speedmatrix corporate structure, the Tradesparq trade-volume aggregations, the Nvidia disclosure-methodology shift, and the BIS designation chain.
Sen. Warren letter to Nvidia General Counsel and Audit Committee Chair on export-control compliance
Senator Elizabeth Warren (Ranking Member, Senate Banking Committee) writes to Nvidia General Counsel Tim Teter and Audit Committee Chair Brooke Seawell for information about Nvidia’s compliance with U.S. export control laws and the accuracy of its public statements about chip diversion. The letter cites three DOJ enforcement actions: the Nov 20, 2025 arrests for exporting AI technology to China via Malaysia and Thailand (“millions of dollars in GPUs”); the Dec 8, 2025 shutdown of a China-linked AI tech smuggling network (“$160 million in H100 and H200 chips”); and the Mar 19, 2026 indictment (“$510 million in diverted servers”). It quotes CEO Jensen Huang’s public statements that “[t]here’s no evidence of any AI chip diversion” and that Nvidia chip market share in China has “dropped to zero,” and puts four numbered questions to the Audit Committee, response deadline June 18, 2026.
