A U.S.-funded nonprofit has documented how billions of dollars in export-controlled Nvidia accelerators flow into China despite parallel bans from Washington and Beijing, revealing a supply chain that operates in plain sight through universities, transshipment hubs, and layered shell companies.
The three channels
C4ADS, a monitoring group financed largely by the U.S. government, identifies three primary conduits. Research institutions embed restricted GPUs inside sprawling multi-vendor contracts handled by small regional integrators. Southeast Asian testing and logistics centers, Vietnam, India, Malaysia, serve as waystations for chips that never actually stay there. And a nested structure of opaque ownership vehicles moves the largest volumes. The group cautions that its tally covers only explicitly declared shipments, so actual flows are likely higher. An earlier Epoch AI estimate placed smuggled GPUs at roughly one-third of China’s total AI compute capacity, possibly more.
Research institutions as cover
The quasi-legal university channel is modest in dollar terms but symbolically potent. C4ADS tracked 56 chips valued at $1.7 million moving this way between July 2025 and January 2026. A broader 2024 review of government records covering multiple years found $6.48 million in similar transactions. Some buyer institutions maintain ties to the Communist Party, defense research, and intelligence agencies, according to the report.
Southeast Asian transshipment hubs
The drop-ship route accounted for $13.4 million in A100, H100, GH100, and AD102-series processors between 2022 and 2025. Chips manufactured at TSMC in Taiwan travel to Vietnam, where testing facilities provide a ready pretext, then onward to Hong Kong. Two firms, Profit New Limited and ELB International Limited, dominate the import side there. Profit New moved $8.7 million in a single March 2025 session, shortly before tighter export rules took effect in April. Several high-value shipments in the dataset lacked cost, insurance, freight, or weight data, and carried suspiciously round import valuations.
The shell company layer
The largest category by far is opaque ownership. C4ADS attributes $4.6 billion to this method, describing a matryoshka-doll structure of nested entities that obscures ultimate beneficiaries. The report does not name the specific companies behind the largest flows, nor does it specify the chip generations involved. U.S. rules currently bar H100, A100, and Blackwell-family sales to China; H20 and H200 units, along with AMD’s MI325X, are eligible for case-by-case licenses, with the H200 subject to a 25 percent tariff. Beijing officially discourages American procurement to boost domestic alternatives led by Huawei, though authorities granted exceptions to ByteDance, Alibaba, and Tencent in 2026.