New research suggests that several leading Chinese AI models were developed using restricted Nvidia Blackwell processors. The findings add to concerns about advanced computing hardware reaching Chinese technology companies despite U.S. export controls.
The investigation into Nvidia Blackwell chips in China examined AI models from DeepSeek, Moonshot AI, MiniMax and Ant Group. Researchers found technical evidence linking six models to Nvidia’s Blackwell architecture.
However, the findings do not establish where the processors were located or whether export restrictions were violated.
Nvidia Blackwell Chips in China Linked to Six AI Models
Will DePue, founder of Unemployment Labs, published the findings on November 12, 2026.
His checkpoint forensics research examined numerical patterns embedded in publicly released AI model weights.
The analysis identified Blackwell hardware signatures in DeepSeek V4, Moonshot’s Kimi K3 models, MiniMax M3 and Ant Group’s Ling-3.1-flash.
Earlier models from the same developers showed signatures associated with Nvidia’s older Hopper architecture.
Researchers reproduced the identified patterns using physical Blackwell processors for two models, strengthening their findings.
The method identifies hardware associated with model initialization. It supports, but does not conclusively prove, that Blackwell processors performed the subsequent training.
Research Raises Questions About Chip Smuggling Routes
U.S. export controls restrict sales of advanced Blackwell processors to China without authorization.
However, Chinese companies may access computing power through overseas facilities or unauthorized hardware transfers.
A separate Epoch AI investigation identified discrepancies in server trade records between Malaysia and China.
China recorded $3.8 billion in server imports from Malaysia between April 2024 and June 2025. Malaysia reported only $600 million in corresponding exports.
Epoch AI said the discrepancy was consistent with possible chip diversion but did not prove smuggling.
An October investigation into Malaysian data centers also examined concerns about Chinese access to overseas AI computing infrastructure.
These concerns highlight a distinction between physically importing restricted processors and accessing computing resources located abroad.
Bloomberg Links Restricted Nvidia Chips to Shanghai Facility
Separate reporting has intensified scrutiny of Nvidia’s distribution network.
According to a November 13 Bloomberg investigation, a Chinese company founded by Alice Huang obtained restricted Nvidia processors.
The equipment was reportedly installed at a Shanghai data center connected to Singapore-based Megaspeed.
Aperia Cloud Services, an Nvidia cloud partner in Singapore, originally purchased some servers.
The reporting does not establish Alice Huang’s personal involvement in the equipment purchases.
A separate Financial Times report examined Singapore’s decision to bar Alice Huang from returning after an investigation into inaccurate information provided to police.
The reported intelligence allegations involving Huang remain unproven.
Earlier Nvidia Research Alleged Hidden China Sales
Concerns about Nvidia’s Chinese business predate the latest technical findings.
In May, short-selling firm Culper Research published a report alleging undisclosed Chinese demand through Southeast Asian intermediaries.
Culper claimed that more than 20% of Nvidia’s fiscal 2026 computing revenue remained connected to China.
The firm disclosed a short position in Nvidia. Its financial estimates and allegations are disputed claims, not established findings.
Additional reporting on Megaspeed examined the company’s business relationships and U.S. regulatory scrutiny.
TechAmerica previously covered allegations involving $7.5 billion in Nvidia chip shipments and the companies reportedly involved.
U.S. Export Controls Face Growing Challenges
Despite restrictions, Chinese developers continue to seek advanced processors to train increasingly capable AI models.
A New York Times investigation examined concerns about U.S. approvals allowing certain Chinese-linked companies to receive advanced Nvidia chips.
Meanwhile, Reuters reported that approximately 600,000 Nvidia chips were awaiting shipment in Hong Kong as Chinese authorities withheld import approval.
The situation reflects competing pressures from U.S. export restrictions and China’s efforts to control foreign semiconductor imports.
TechAmerica also reported on Nvidia H200 shipment delays and rising unofficial-market prices.
What the Findings Mean for AI Chip Restrictions
The Unemployment Labs research offers technical evidence that Chinese AI developers accessed advanced computing hardware associated with Nvidia’s Blackwell generation.
However, the analysis cannot identify the processors’ physical locations or establish how developers obtained access.
Determining whether export laws were broken requires additional evidence about shipments, ownership, licensing and the operation of overseas data centers.
For regulators, the findings highlight the difficulty of controlling advanced AI computing through hardware export restrictions alone.
