Nvidia Chips in the PLA: What the Procurement Records Show
Two investigations - Wirescreen and Bloomberg - document over 500 PLA procurement attempts for Nvidia AI chips since 2019, and seven military-linked universities seeking the H200.
"Irrefutable": How US Chips Are Reaching China's Military
For years, the working assumption behind American semiconductor export controls was that tightly drawn rules — with escalating restrictions in October 2022 and October 2023 — could keep the most capable AI hardware out of the hands of the People's Liberation Army. Two investigations published this week, one by Bloomberg and one drawing on a report by business-intelligence firm Wirescreen, offer the most systematic evidence yet that this assumption has been tested — and, in documented cases, broken.
"This makes it irrefutable that US technology is arming the Chinese military," John Costello, a Wirescreen analyst and the report's principal author, told The New York Times, which first obtained the research. Wirescreen, which specialises in Chinese corporate intelligence, reviewed approximately 3,800 Chinese public procurement records relating to high-end chips and computing hardware. Among those records, Costello's team found more than 500 instances in which units of the PLA specified Nvidia products — by model name or performance specification — as a procurement requirement, covering the period from 2019 through 2025.
The Bloomberg investigation, published 1 June 2026 by reporters Kate O'Keeffe and Maggie Eastland, took a complementary approach: examining procurement announcements and university lab websites to identify Chinese institutions with documented military ties seeking access to the H200 — the most powerful AI processor the US has ever permitted for legal export to China. Bloomberg identified at least seven such universities.
Together, the two inquiries describe not isolated leakage but a sustained, multi-channel procurement effort that adapted to each successive layer of US restrictions.
What the Procurement Records Show
Wirescreen's methodology relied on Chinese public tendering documents, which institutions — including military ones — are required to publish under Chinese procurement law. The firm worked with Sunny Cheung, a fellow at the Washington-based Jamestown Foundation who studies China's defence technology sector. Cheung helped establish the criteria for classifying an institution as a direct military collaborator rather than a peripheral one. The threshold was deliberately high: universities were included only if they had signed memoranda of understanding with the PLA, undertaken joint research, engaged in documented technology transfer to the defence sector, or are formally part of the "Seven Sons of National Defense" — an elite group of Chinese universities legally designated to advance PLA capabilities.
The Wirescreen findings cover four Nvidia chip families in the procurement records:
- A100 and its China-market variant, the A800 (introduced after the October 2022 controls)
- H100 and its China-market variant, the H800 (introduced in the same cycle)
These models appear in requests from across the PLA's institutional structure. Notably, the Jiangnan Institute of Computing Technology — one of the principal research centres of China's Cyberspace Force, responsible for cyberwarfare, reconnaissance, and domestic surveillance — appears in the records, despite having been added to the US Commerce Department's trade blacklist in 2019 specifically for its role in developing military supercomputers. Wirescreen found that the Jiangnan Institute's procurement attempts became less frequent after the 2022 controls took effect, but did not stop.
Across service branches, the PLA's Cyberspace Force emerges as the single largest identified buyer of American AI computing hardware in the procurement records reviewed.
Cheung separately analysed several hundred tenders published on the PLA Procurement Network over a six-month window in 2025. He found that the Chinese military had increased its use of contractors that publicly claim to use only domestically manufactured hardware — notably Huawei's Ascend AI chips — while simultaneously continuing to seek Nvidia hardware through other channels.
The Bloomberg Layer: H200 and the Seven Sons
Bloomberg's analysis focuses on a more recent and politically sensitive data point: institutions seeking access to the H200, which BIS first permitted for export to China on a case-by-case basis in January 2026 under a rule that shifted the default for certain chips from a presumption of denial to case-by-case review. The H200 represents the ceiling of what the US has ever allowed into China legally; the Blackwell generation (B100, B200) remains fully restricted.
The earliest documentation Bloomberg identified showing a military-linked Chinese university seeking H200 access was dated June 2025 — months before BIS formalised the case-by-case review policy. This implies institutions were positioning themselves ahead of anticipated rule changes.
Two of the seven universities Bloomberg identified — Beihang University and Northwestern Polytechnical University — are members of China's Seven Sons of National Defense and are on the US Commerce Department's Entity List. Being Entity-Listed means US exporters are legally prohibited from shipping to them without a specific BIS licence, which carries a presumption of denial for military-end-use applications.
The fact that these institutions appear in procurement documents seeking H200 chips does not necessarily mean the chips were successfully obtained through legal channels. But Bloomberg's methodology — cross-referencing procurement announcements with university lab disclosures detailing installed computing resources — suggests some institutions have found paths to hardware even after blacklisting.
The Architecture of Failure: How Export Controls Are Being Circumvented
The Downgraded-Variant Cycle
The most documented workaround is structural. Each time BIS tightened controls, Nvidia produced a performance-capped variant for the Chinese market: the H800 after October 2022 (reduced interconnect speed), the H20 after October 2023 (inference-optimised, roughly 15% of H200 arithmetic performance). During the window when H800 exports were legal — October 2022 to October 2023 — Nvidia's disclosed China revenues exceeded $9 billion, implying a large installed base. The H20 required a new export licence from April 2025 but was cleared for licensed export again in July 2025.
Third-Country Diversion
Singapore became the most-discussed waypoint after Nvidia's reported revenues from that jurisdiction surged from under $1 billion in Q4 2022 to nearly $8 billion in Q3 2024 — a trajectory that tracks almost precisely with the imposition of China controls. Third-country diversion involves purchasing hardware through distributors in jurisdictions with weaker end-use verification, then re-routing to China. BIS has expanded the Foreign Direct Product Rule (FDPR) to extend US jurisdiction to chips produced using American semiconductor manufacturing equipment, meaning foreign-made chips can still be subject to US export law if US-origin tooling was used in their fabrication.
Front Companies and Smuggling Networks
The most overt channel is outright smuggling. The US Justice Department has indicted multiple individuals across several separate networks. The most prominent recent case involves Yi-Shyan "Wally" Liaw, a co-founder of Super Micro Computer, and two associates who were arrested for allegedly shipping approximately $2.5 billion worth of AI servers — containing controlled Nvidia GPUs — to China via shell companies and Southeast Asian warehouses. Prosecutors alleged that once servers arrived at regional staging points, their serial numbers were transferred to dummy units and new documentation was issued before final shipment to China. Fortune magazine, drawing on encrypted message records, reported in May 2026 that at least eight separate chip-smuggling networks — each conducting transactions exceeding $100 million — had been operating simultaneously.
Remote Access and Cloud Computing
Perhaps the least visible channel is cloud-based access. Research institutes linked to the Chinese military have increasingly rented capacity from commercial data centres, gaining remote access to Nvidia hardware without triggering a physical export. This route raises a distinct regulatory question about whether providing cloud compute to Entity-Listed end users constitutes a violation — one BIS has not resolved in published guidance.
The Export Control Timeline
| Date | Restriction | Chips Affected | Known Workaround |
|---|---|---|---|
| Oct 7, 2022 | BIS initial advanced computing rule; performance thresholds on interconnect speed and compute | Nvidia A100, H100; AMD MI250 | Nvidia releases A800, H800 (reduced interconnect, legally saleable) |
| Oct 17, 2023 | BIS update; tighter TPP + performance density thresholds; expanded Entity List additions | A800, H800, L40, L40S, RTX 4090 | Nvidia releases H20 (inference-optimised, ~15% of H200 arithmetic performance) |
| Dec 2, 2024 | BIS Entity List expansion (140 entities); FDPR scope extended to HBM, advanced packaging, DRAM | Broad semiconductor manufacturing equipment | Third-country diversion; smuggling networks |
| Apr 2025 | H20 subjected to export licence requirement | H20 | DOJ indictments of smuggling networks begin; cloud compute workaround grows |
| Jul 2025 | H20 and AMD MI308X cleared for licensed export to China | H20, MI308X | Entity-Listed institutions not eligible for licences |
| Jan 15, 2026 | BIS revises H200 licence policy from presumption-of-denial to case-by-case review | H200, AMD MI325X | Military-affiliated universities seeking H200 ahead of policy formalisation |
Nvidia's Position
In response to reporting on the Wirescreen and Bloomberg findings, Nvidia spokesperson John Rizzo said the number of chips identified in PLA procurement records was small relative to modern AI cluster sizes. "Advanced AI systems typically run on networks of 100,000 or more chips," Rizzo noted, adding that "as Chinese technological capabilities have improved, the Chinese military appears to be focusing on Huawei chips" and characterising the claim that the PLA relies on US technology as "silly and untrue."
The statement echoes a position Jensen Huang, Nvidia's chief executive, has articulated in previous public settings — that the Chinese military does not rely on Nvidia's products. The Wirescreen report directly challenges this framing: the procurement records document active, sustained efforts to acquire specifically Nvidia hardware by model name, and the Cyberspace Force is identified as the largest single PLA buyer in the dataset. The records do not resolve what proportion of PLA AI compute runs on Nvidia versus Huawei hardware at present, a distinction Rizzo's statement does not quantify with sourced data.
BIS did not issue a statement directly addressing the Wirescreen report. The bureau's most recent policy action — the January 2026 case-by-case review rule for H200 chips — contains conditions requiring end-user verification and certification, but no enforcement action tied to the Bloomberg or Wirescreen findings has been announced as of publication.
The India Dimension
India occupies a structurally different but directly adjacent position in this story. Under the tiered AI diffusion framework the Biden administration introduced in January 2025, India was classified as a Tier 2 country — subject to quantitative caps of approximately 49,901 H100-equivalent GPUs per country through 2027, placing it alongside Singapore and Israel rather than the Tier 1 group that receives US-ally treatment.
That classification has practical stakes. The IndiaAI Mission's initial target of 10,000 GPUs has expanded to approximately 38,000 deployed for researchers and startups by late 2025. The 2026-27 Union Budget announced ISM 2.0 with a mandate covering semiconductor equipment, materials, and domestic IP. Tata Semiconductor's Dholera fab — a $11 billion front-end fabrication facility backed by a Tata-ASML agreement signed in May 2026 — marks India's first serious wafer-fab commitment, though it targets automotive and mobile chips, not AI GPUs.
India updated its SCOMET export control list in September 2025 to cover advanced lithography equipment below 45-nanometre resolution — an alignment with, but not full membership of, the Wassenaar Arrangement's semiconductor controls. The February 2026 US-India interim trade agreement was read in New Delhi as a signal that India's Tier 2 status would not harden into an effective embargo. A 2024 bilateral agreement for a joint US-India semiconductor facility in Uttar Pradesh, producing defence-grade chips for drones and missile seekers at roughly 50,000 units annually, adds a layer of formal cooperation that distinguishes India from China in the US licensing calculus.
The China-PLA diversion evidence matters to India in a second register. The same third-country routing — Singapore, Southeast Asian warehouses, shell-company resellers — documented in the Wirescreen and Fortune investigations runs through India's neighbourhood. Washington's expectation that Tier 2 countries tighten their own re-export verification will be harder for New Delhi to defer as the mechanics of diversion become more publicly documented.
What to Watch
- Whether BIS moves against the Entity-Listed universities identified in the Bloomberg investigation — Beihang and Northwestern Polytechnical are already blacklisted, but enforcement action directly tied to documented H200 procurement attempts would mark a new phase of active enforcement rather than list maintenance.
- The scope of cloud compute guidance from BIS — remote access to controlled chips via data centres is the least-regulated acquisition channel the Wirescreen report documents; a formal rule or advisory opinion would close a significant gap.
- Wirescreen's full dataset — the firm has signalled it will release additional details from the 3,800-record review; the list of named PLA entities seeking chips beyond the Jiangnan Institute and the Cyberspace Force units will sharpen the enforcement picture.
- Huawei Ascend scaling — Cheung's PLA tender analysis found growing use of Huawei branding as a public-facing cover; whether Ascend 910C and 910D can displace Nvidia for PLA training workloads remains the central unknowable, and a technical assessment from SemiAnalysis or Epoch AI on cluster performance parity would be the most consequential near-term data point.
- India's SCOMET alignment — whether Washington accelerates a Tier 1 equivalency for Indian AI compute access in exchange for tighter Indian re-export enforcement will determine how much room Indian AI developers retain as the US tightens the overall regime.
- Super Micro trial proceedings — the $2.5 billion smuggling case involving co-founder Yi-Shyan Liaw will produce documentary evidence on serial-number substitution and shell-company routing that will directly shape BIS guidance for hardware manufacturers and distributors.
