How Banned Nvidia Chips Power China’s Top AI Models

Export-banned Nvidia graphics cards are driving some of China’s most powerful artificial intelligence models, according to independent technical investigations. The findings reveal that leading open-weight AI systems developed by top Chinese laboratories were trained using advanced silicon strictly restricted by U.S. export controls.

This widespread access to high-end hardware highlights major vulnerabilities in international trade policies. It also raises difficult corporate accountability questions for the world’s leading semiconductor designer as regulatory scrutiny intensifies. As global AI capabilities accelerate, enforcing geographic boundaries on cutting-edge compute infrastructure—whether through offshore cloud proxies or intricate smuggling networks—has proven remarkably difficult.

Uncovering the Silicon Fingerprints in AI Weights

The revelation stems from advanced technical work conducted by Will DePue, a former OpenAI researcher leading the nonprofit research group Unemployment Labs. By examining the weights—the underlying mathematical configurations—of large language models released openly by prominent Chinese entities like DeepSeek, Moonshot AI, MiniMax, and Ant Group, researchers discovered distinct signatures left by the hardware used during training.

According to this analysis, models such as Moonshot’s Kimi releases and DeepSeek’s V4 series exhibit structural footprints indicative of training on Nvidia’s powerful Blackwell chips. These powerful accelerators are strictly prohibited from direct export to mainland China or Chinese-owned firms under current U.S. federal policy. White House technology adviser Michael Kratsios previously noted intelligence indicating Moonshot’s reliance on restricted hardware, corroborating these technical findings.

Meanwhile, data compiled by Epoch AI estimates that roughly two million advanced processors entered China throughout recent years via diversion schemes designed to obscure final destinations. While tech enterprises continue building autonomous systems, the foundational compute hardware remains intensely contested geopolitical territory.

How Secondary Markets and Diversion Networks Operate

Training frontier artificial intelligence models requires immense computational power, traditionally housed in massive clusters of specialized graphical processing units. These chips perform trillions of parallel floating-point calculations necessary to optimize neural network weights during training. U.S. export controls attempt to deny rivals access to this specific capability by restricting direct shipments of cutting-edge silicon.

However, the global semiconductor supply chain relies on complex international logistics networks. Rather than shipping directly into restricted zones, hardware has frequently been funneled through third-party jurisdictions. Epoch AI data highlights that approximately 275,000 Nvidia accelerators were imported into Malaysia across 2024 and 2025 alone, where offshore data centers provide remote compute access to international clientele without triggering direct border alarms.

At the architectural level, hardware training runs generate specific bottlenecks in memory bandwidth, interconnect latency, and tensor core utilization. When analysts dissect open-weight models, residual optimization patterns reveal fingerprints corresponding directly to the hardware instruction sets of Nvidia’s Hopper and Blackwell microarchitectures. Because different GPU generations handle tensor parallelisms and collective communication primitives in unique ways, training logs and weight distributions expose the precise computational scaffolding used during training.

Market Repercussions and Enterprise Competitiveness

The presence of restricted hardware in competitive training pipelines challenges the foundational premise of American technology policy, which assumes trade restrictions can successfully bottle up AI progress. Security researchers argue that current rules miss the mark by regulating ownership within specific borders rather than governing compute utilization and remote access globally.

Comparing this situation with industry benchmarks shows that leading American labs like OpenAI, Anthropic, and Google DeepMind typically train models on massive domestic hyperscale facilities. The revelation that Chinese competitors achieved comparable algorithmic efficiency despite fragmented, diverted hardware clusters suggests that algorithmic innovations are increasingly neutralizing raw hardware advantages.

For everyday enterprise consumers and global businesses, these supply chain fractures introduce pricing volatility and regulatory compliance hurdles. Companies deploying automated software solutions must navigate an increasingly fractured global digital infrastructure. Ironically, some market analysts point out that strict U.S. controls may ultimately accelerate China’s domestic semiconductor independence by forcing local firms to invest heavily in indigenous silicon alternatives.

The Chronology of Export Controls and Trade Loopholes

The regulatory and corporate fallout continues to unfold rapidly. Understanding how this situation developed requires tracing key chronological milestones from early trade restrictions to the current crisis:

• October 2022: The U.S. Department of Commerce implements sweeping export controls targeting advanced AI chips, specifically restricting exports of high-end Nvidia A100 and H100 processors to China.
• October 2023: The administration updates export rules to close loopholes, capping performance density and introducing gray-zone country licensing requirements.
• 2024–2025: Secondary smuggling networks emerge, routing hundreds of thousands of restricted chips through transshipment hubs in Southeast Asia, while Chinese labs adapt training pipelines to leverage distributed cluster architectures.
• Late 2025 / Early 2026: Independent investigators analyze model weights of frontier Chinese open-weight models, definitively identifying mathematical fingerprints of Blackwell-class silicon.
• Projected Future Milestones (2026–2027): The U.S. Department of Commerce anticipates rolling out stringent “know-your-customer” mandates for cloud computing providers, targeting remote access leasing models and forcing hardware verification at the data-center virtualization layer.

Frequently Asked Questions

How did Chinese labs acquire export-banned Nvidia chips?

Research and intelligence reports indicate that chips bypassed direct trade bans by being routed through third-party nations like Malaysia, utilized in offshore cloud data center facilities, or smuggled via specialized intermediary logistics networks that obscure final corporate destinations.

What evidence proves these models used banned chips?

Researchers analyzed the structural model weights—the unique mathematical architecture of large language models—released by Chinese labs, finding distinct hardware fingerprints matching the tensor core configurations and instruction set efficiencies unique to Nvidia’s advanced processors.

Are U.S. companies currently breaking laws by selling chips abroad?

Current U.S. regulations impose limited tracking obligations on manufacturers once controlled goods are sold to licensed intermediaries, though ongoing federal investigations are examining whether specific companies or executives failed to exercise adequate due diligence.

How do remote cloud leasing models bypass geographic trade restrictions?

Offshore data centers located in non-restricted jurisdictions purchase physical hardware legally and then lease computational time over the internet to international clients. This allows engineers anywhere in the world to train massive neural networks without physical possession of the restricted silicon inside their home borders.

Sources and further reading

Read the original investigative report on TechCrunch.