A framework for auditing diversion-thesis claims at the primary-source-disclosure layer (Culper Research 2026-05-13 NVIDIA short thesis as worked example). The magnitude studio recomputes peak and effective ExaFLOPS live as trade-volume, chip-mix, MFU, and ASP move; a Tier-1-only counterfactual toggle strips the Tier-3 input and renders what the framework establishes without it. Every claim carries the authority tier of its source.
A structural framework for assessing claims that restricted AI compute reaches unauthorized end users via Southeast-Asian intermediaries. Scrutica builds the cascade chain from corporate filings, flags where Culper’s trade-record claims diverge from disclosed corporate topology, and cross-references Chinese hyperscaler peering at SE-Asian facilities via PeeringDB.
Diversion-rate inference is structurally bounded above by absent CY2023+ BIS license data. The CY2018-CY2022 license-action substrate documents what was licensed (denied + RWA + approved + ECCN top-10) for the pre-Oct-2023-rule window; the post-rule window — exactly the period the Culper-NVDA worked example operates over — has no BIS license-action substrate to corroborate against. The framework substitutes bilateral-trade-flow measurement of HS 854231/8542 advanced-chip-class proxies and corporate-disclosure topology, and labels every inference accordingly.
See what we can / cannot answer about the post-2022 window for the per-question crosswalk. Substrate refresh will land on the BIS publication, not on inference; the “Verify, never hedge” principle (CLAUDE.md DQP §8) holds — analyst commentary is not a substitute for the government measurement.
If Culper's $4.74B trade-flow figure holds and the assumed chip mix is broadly correct, the alleged channel represents ~160K GPUs / 166 ExaFLOPS peak BF16 dense / 67 ExaFLOPS effective at 40% MFU — approximately 1.58× China's MIIT 2025 intelligent-computing target (105 EFLOPS; six-agency Computing Power Infrastructure Action Plan, October 2023), sufficient to train ~98 GPT-4-class models annually on aggregate effective compute, and accounting for 4.1% of Nvidia's FY2025 Data Center revenue.
How to read this number. Per-card FLOPS data is Tier 1 (Nvidia spec sheets). Chip-mix and MFU are Tier 2 derived metrics. The trade-volume input is Tier 3 (Culper Research 2026-05-13 citing Tradesparq) and is unverified at the primary-source layer. Use the controls below to test how the magnitude moves with each assumption, or toggle the counterfactual Tier-1-only mode to see what the framework reports when the Tier-3 input is stripped.
Cards / FLOPS by SKU at current scenario
| SKU | Mix | ASP $K | PFLOPS / card | Cards |
|---|---|---|---|---|
| H200 SXM5 | 50% | 30.0 | 0.990 | 79,033 |
| H100 SXM5 | 30% | 30.0 | 0.990 | 47,420 |
| H800 SXM5 | 15% | 25.0 | 0.990 | 28,452 |
| B300 Blackwell Ultra | 5% | 45.0 | 2.500 | 5,269 |
| Total | 100% | — | 166.5 EFLOPS | 160,174 |
| At 40% MFU → 66.6 EFLOPS effective | ||||
Caveats
Each edge is one corporate-filing or trade-record relationship in the alleged Culper-NVDA chain. Edge color encodes primary-source corroboration (green), single-sourced via Tier-3 analyst citation (amber dashed), or absent from the licensed supply-chain disclosure topology (stone dotted). Toggle “fade Tier-3 edges” to render only the structural skeleton that survives at primary-source corroboration.
The cascade graph rendered as a cross-tab. Each of the 11 alleged edges appears in exactly one matrix cell (corroboration status × tier) and one relationship-kind row — surfaces which channels survive primary-source corroboration and which sit at Tier-3 single-sourced via Tradesparq, WSJ, or the Culper narrative.
Status × cascade tier
| Corroboration status | OEM | Intermediary | End-customer | Total |
|---|---|---|---|---|
| Primary-source corroborated | 2 | 2 | · | 4 |
| Tradesparq via Culper | · | 4 | · | 4 |
| WSJ via Culper | · | · | 2 | 2 |
| Culper narrative only | · | · | 1 | 1 |
| Total per tier | 2 | 6 | 3 | 11 |
Relationship-kind breakdown · ordered by primary-source share
1 primary-source corroborated
1 primary-source corroborated
1 primary-source corroborated
1 primary-source corroborated
3 tradesparq via culper
2 wsj via culper
1 culper narrative only
1 tradesparq via culper
Each edge appears in exactly one row and one column. The matrix exposes the structural asymmetry surfaced by the cascade graph as a single read: the corporate-topology edges (wholly-owned subsidiary, corporate rename, shared directors, financing chain) are primary-source-corroborated by Singapore Bizfile / California SoS / Hong Kong company filings / Malaysian SSM; the trade-flow and lease-chain edges are single-sourced via Tradesparq or Wall Street Journal reporting — both reach the framework at Tier 3.
If BIS designates Aivres alongside Inspur Group; if the corridor behind the March 2026 DoJ indictment is severed (per the DoJ release, a front company purchased ~$2.5B of AI servers from the manufacturer, unnamed in the release, and at least ~$510M worth was diverted to China; Sen. Warren’s letter ties the case to a Supermicro co-founder); if a Beijing- or Washington-side action cuts the post-restriction Asian-routed Nvidia flow — what propagates through the 3,416-node Scrutica supply-chain graph downstream? Four scenarios. Each carries severity, decay, direction, and per-tier weighted impact with the top affected nodes; a Tier-3-aware probe swaps in the 3,428-node graph that includes Culper’s 14 alleged additional edges so the Tier-3 delta is read directly. The supply-chain graph is built from a licensed supply-chain database held under subscription and not redistributed.
The DoJ-indicted corridor as proxy. Severity 100% at the Supermicro OEM node, downstream-only propagation with 0.85 per-hop decay. The narrowest enforcement footprint of the four scenarios.
Weighted impact by supply-chain tier
Top affected nodes by tier · per-node impact = severity × decay^depth
L0 Materials/Foundry
L3 Systems/OEM
L4 Cloud/End-User
+4 more in substrate
Substrate-side canonical-name resolution gap: the cascade-counterfactual writer emitted these rows with name === id (the org- namespace slug, not a display name). The label is the slugified id; the canonical name lives in canonical_org_slim.json and should be joined at substrate-build time. Tracked at A8 D8.3.
Tier-3-aware probes · cannot be modeled on the canonical graph alone
The Aivres → Speedmatrix → Megaspeed → Novagate → Aolani chain rides on 14 additional edges that exist only at Culper Research’s analyst synthesis. Their reach into Alibaba, ByteDance, IFLYTEK, PT Indosat, and Opera surfaces here at the Tier-3 layer — and is absent from the canonical-graph scenarios above by construction.
TIER-3-AWARE: Aivres channel shutoff (Aivres + alleged downstream)
BIS enforcement extends across the full Culper-alleged Aivres + Speedmatrix + Megaspeed + Novagate + Aolani chain. Cannot be modeled on the canonical graph alone — depth-4 reach into Alibaba, IFLYTEK, PT Indosat, Opera requires the Tier-3 edges.
NARROW: Aivres alone shutoff (the cleanest Culper-thesis test)
The cleanest Culper-thesis test: BIS extends to Aivres only. Depth-2 reach via Speedmatrix and Aolani — ByteDance and Alibaba surface at depth 2 only because Tier-3 edges connect them.
Substrate · cascade_counterfactual_2026_05_14.json · cascade_with_tier3_2026_05_14.json. Weighted impact = per-node severity × decay^depth, normalized by graph-level node weighting. Propagation uses the existing Scrutica cascade-simulation engine against the licensed supply-chain graph. The single load-bearing Tier-3 input is the set of 14 additional Culper-alleged edges; with the toggle off, no Tier-3 edge enters the model.
Where public primary-source disclosure (SEC, A-share + HK + Bizfile / SSM) and the licensed corporate-disclosure substrate read against the Tier-3 narrative for the same entity; the Tier-1 measurement sits next to the Tier-3 interpretation, and both can be true at once.
cross listing state ownership disclosure divergence
Insider category 'Government' = 50.88% of outstanding shares
Licensed-database record (held under subscription, not redistributed)· 2026-05-14
Insider category 'Private Company' = 33.26%; no Government category disclosed
Licensed-database record (held under subscription, not redistributed)· 2026-05-14
Culper Research 2026-05-13 reads the cross-listing asymmetry as evidence that the BIS Entity List 50% subsidiary-extension threshold is not effectively binding on the Inspur Group's downstream OEMs (Aivres) because the 'Private Company' classification on the A-share listing decouples the U.S. compliance view from the underlying state ownership.
geographic revenue methodology change within disclosure window
Nvidia stated change reflects 'better representation of the geographic profile of our revenue' — companies' HQ location now overrides Singapore-based billing
Nvidia Q3 FY2026 10-Q (October 2025) — methodology change disclosure· 2025-10
Post-shift LTM Jan '26 georev: US 69.3% (+47.6% Y/Y), Taiwan 19.6% (+24.4%), Mainland China 8.9% (-30.5%), Singapore not in top 8
Licensed-database record (held under subscription, not redistributed)· 2026-05-14
Per Culper, the methodology change makes pre-2025 vs. post-2025 geographic-revenue compositions directly non-comparable; absent the disclosure shift, Singapore would likely still surface as a material exposure even with operational changes.
revenue growth counter cycle to export restrictions
11.7% Mainland China; +46.5% Y/Y; LTM Jun '25 $33.7B total revenue
Licensed-database record (held under subscription, not redistributed)· 2026-05-14
USD 2.5B alleged smuggling via 'Company-1' (Bloomberg-identified as OBON Corp)
DoJ SDNY indictment of Wally Liaw (March 2026)· 2026-03
Per Culper, the +46.5% Y/Y growth during a tightening window is itself a signal worth scrutinizing; the DoJ indictment quantifies one corridor at $2.5B but Culper's narrative interpretation is that this 'is the tip of the iceberg' with multiple parallel OEM/intermediary channels.
consolidated subsidiary revenue not visible at parent level
Mainland China 97.6%, HK 2.2%, Macau 0.3%, US 0%, Malaysia 0% of LTM Dec '25 CNY 153.4B
Licensed-database record (held under subscription, not redistributed)· 2026-05-14
Culper Research 2026-05-13 cites this as the canonical 'consolidated-disclosure / Tradesparq-disclosure tension' that motivates the Diversion Pipeline thesis.
supplier graph density anomaly for alleged largest buyer
651 counterparty edges; 254 disclosed customers; top-ranked customers include Alphabet (rank 1), Penguin Solutions (rank 10), IEIT Systems A (rank 100), Indosat partner; Gigabyte direct customer (3M price corr 0.87)
Licensed-database record (held under subscription, not redistributed)· 2026-05-14
Culper Research 2026-05-13 reads the absence as evidence of an intermediary-mediated diversion path: Nvidia ships to Elite OEMs (Aivres / Giga Computing / Supermicro), which then route to SEA intermediaries (Speedmatrix / Novagate / OBON), which then route to Chinese end customers. The Tier-1 graph view is structurally consistent with this hypothesis but does NOT independently corroborate the downstream chain.
segment disclosure lag
Last segment row Jun '23; FY24 and FY25 columns empty
Licensed-database record (held under subscription, not redistributed)· 2026-05-14
Disclosure-lag alone is not evidence of wrongdoing; surfaced so a researcher cannot inadvertently apply stale segment splits to current-period analysis.
segment disclosure lag
Last segment row Dec '23; FY24 and FY25 columns empty
Licensed-database record (held under subscription, not redistributed)· 2026-05-14
Disclosure-lag alone is not evidence of wrongdoing; surfaced so a researcher cannot inadvertently apply stale segment splits to current-period analysis.
capex acceleration outpacing revenue recognition
FY22 capex -1,937M MYR; FY25 capex -6,545M MYR; FY25 FCF -419M MYR (first negative-FCF year since FY22)
Licensed-database record (held under subscription, not redistributed)· 2026-05-14
YTL Power International Malaysia revenue +243.6% Y/Y at LTM Jun '25 — consistent with early data-center revenue recognition as the Johor facility ramps
Licensed-database record (held under subscription, not redistributed)· 2026-05-14
Culper Research 2026-05-13 cites YTL AI Cloud (the brand under which YTL AI Labs operates the Johor facility) as the channel through which Megaspeed reportedly operates servers; the capex acceleration is consistent with this thesis but does not by itself constitute evidence — large publicly disclosed AI-data-center buildouts produce identical capital-allocation signatures regardless of end-customer mix.
Public IX peering tables disclose which networks are present at which SE-Asian facilities. A direct peering between a Culper-named facility and a Chinese-hyperscaler AS-number is primary-source evidence of network-level proximity, independent of any Tier-3 analyst interpretation. The True IDC East Bangna ↔ Huawei Cloud (AS136907) finding leads; four representative adjacent disclosures and the calibrated nulls follow.
True IDC (Charoen Pokphand Group / True Corporation subsidiary)·IX: SYMPHONY THAI - IX BANGKOK - 1
Culper relevance. True IDC signed MOU with Siam AI per Culper p.38 (Dec 2024) — 'establish Thailand as the Regional AI data center hub'. Direct Culper-named entity.
AS136907 · ContentChinese hyperscaler / state-telecom AS-numbers disclosed at PeeringDB-listed facilities across the Malaysian, Thai, Indonesian, Vietnamese, and Singaporean data-center cluster. Not directly Culper-named but in the same diversion-corridor metropolitan areas.
| Facility | Operator | Disclosed Chinese peers |
|---|---|---|
| STT Bangkok 1 | ST Telemedia (Singapore-Thai joint venture) | Alibaba (Aliyun)AS45102 · ContentHuawei Cloud GlobalAS136907 · Content |
| NTT Cyberjaya Data Center (CBJ) | NTT Global | Alibaba (Aliyun)AS45102 · Content |
| DCI Indonesia (JK5) | DCI Indonesia (Salim Group) | Huawei Cloud GlobalAS136907 · ContentByteDance / TikTokAS396986 · Content |
| China Mobile International - Singapore DC | — | China Mobile InternationalAS58453 · NSPChina Mobile International - NIIAS58807 · NSP |
YTL Green Data Center / YTL AI Cloud / YTL Communications · Equinix JH1 - Johor · Singapore Equinix SG1-SG5 cluster.
PeeringDB visibility partitions Asian data centers into two structurally different topologies: (a) MID-TIER facilities (NTT, BDx, DCI, STT, TM ONE, True IDC East Bangna) where Chinese hyperscaler peering is openly disclosed via public IX — these are 'Chinese tech is here' facilities; (b) HYPERSCALE facilities (YTL Green, Bridge DC, Singapore Equinix flagships) where tenant relationships are private cross-connect, invisible to public peering tables. The single direct Culper-corroborating finding is True IDC East Bangna ↔ Huawei Cloud — a primary-source BGP-table disclosure that strengthens Culper's CP Group / True IDC / Chinese-tech-tenancy thesis with non-narrative-dependent evidence. 105 SE-Asian facilities scanned; 13 with Chinese-hyperscaler peering. Source: PeeringDB, analyzed 2026-05-14.
6 milestones from the 2023 BIS designation of Inspur Group through the 2026-06-01 Warren letter to Nvidia's General Counsel and Audit Committee Chair. Each clicks through to its primary source; 2 additional Tier-3 milestones (NYT shipment, Tradesparq window) live in the substrate and are not rendered here — surface scope is editorial, gated on the substrate's render_in_timeline field.
U.S. Department of Commerce, 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.” Designation is the upstream regulatory anchor for every subsequent disclosure-anomaly card.
California Secretary of State business filing records the corporate name change from Inspur Systems Inc. to Aivres Systems Inc. — approximately two months after the BIS designation of parent Inspur Group. The rebrand opens the question the Tracker’s cascade graph renders: does the 50%-subsidiary-extension rule uplift through Inspur Information (A-share-listed sibling, owner of Aivres) to the post-rebrand US entity?
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’s share of disclosed revenue collapses; United States share rises to ~70%. Cross-source disclosure-anomaly card “da-nvidia-georev-methodology-shift” anchors to this filing.
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.” The indictment provides primary-source corroboration of the structural pattern — OEM→intermediary→end-customer — the Tracker’s cascade graph encodes.
Culper Research publishes its NVDA short thesis, synthesizing the Inspur / Aivres / Speedmatrix structural topology, the Tradesparq trade-volume aggregations, the Nvidia disclosure-methodology shift, and the BIS designation chain. The Tracker’s render of this publication is the worked example — not endorsement of every claim, and not a stand-in for the underlying primary sources.
Senator Elizabeth Warren (Ranking Member, Senate Banking Committee) writes to Nvidia General Counsel Tim Teter and Audit Committee Chair Brooke Seawell — not to CEO Jensen Huang — requesting 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.
Two highest-leverage levers, each naming the statutory authority under which reform could proceed and the structural-engineering gap the lever would close. Descriptive — not a prediction of legislative likelihood, and not advocacy.
Cross-listing state-ownership asymmetry — Inspur Group is BIS-designated (2023-03), Inspur Information (its A-share-listed subsidiary, owner of Aivres LLC) is not, and the 50% subsidiary-extension rule does not uplift through the foreign-listed parent layer.
Reform direction
Amend the 50% rule so that designation of any entity in an ownership chain uplifts to every entity that ultimately controls, or is ultimately controlled by, the designated party — regardless of jurisdiction of intermediate holding entities. Closes the cross-listed-parent gap surfaced by the Inspur Group / Inspur Information / Aivres LLC topology.
SEA beneficial-ownership disclosure exists on paper (Singapore Companies Act §386AF; Malaysian Companies Act 2016 §60A; Indonesian Presidential Regulation 13/2018) but is not API-accessible to allied export-control authorities. End-customer identity verification requires bilateral mutual-legal-assistance requests that move at jurisdictional-court timescale.
Reform direction
Bilateral agreements with SG / MY / ID / TH / VN to provide allied export-control authorities API-level access to registered beneficial-ownership data — analogous to FATCA-style information sharing but for dual-use technology end-use verification. Closes the verification-latency gap the Culper thesis turns on.