US chip export restrictions on China

A forensic analysis of publicly released artificial intelligence models has produced the first technical evidence that China's leading AI laboratories are utilizing Nvidia's most advanced hardware. The findings indicate that top Chinese developers initialized their newest frontier models on export-controlled Blackwell graphics processing units.[1][2]

The research was published by Will DePue, a former OpenAI engineer who worked on the Sora and GPT-4o models before founding the nonprofit research group Unemployment Labs. DePue's investigation relied entirely on the published weights of the Chinese models, requiring no leaks or insider accounts.[1][3]

According to the report, six models released since April 2026 by DeepSeek, Moonshot AI, MiniMax, and Ant Group contain numerical patterns that match Blackwell hardware. The discovery provides concrete data to support previous U.S. intelligence allegations that Chinese firms had accessed the restricted technology.[1][4]

Tracing the Hardware Fingerprint

The forensic method exploits a quirk in how neural networks are initialized before training begins. Software must assign billions of starting values to a model's weights, typically using a pseudorandom number generator called Philox that distributes the computational work across a chip's processing units.[1][4]

Because Nvidia's architecture divides this task among its streaming multiprocessors—using 2,048 threads per core—chips with different core counts scatter the starting values differently. Traces of this initial scattering survive the training process and remain visible in the final published weights of the open-source models.[1][3]

DePue's team searched for the hardware configurations that could reproduce these specific initial values. They found that Nvidia's Blackwell B200 and B300 chips, which feature 148 streaming multiprocessors, perfectly matched the initialization patterns of DeepSeek-V4-Flash and MiniMax-M3.[1][3]

"In doing so, I found conclusive evidence that six Chinese AI models—including DeepSeek-V4-Pro, Kimi K3, and MiniMax-M3—were initialized on Nvidia Blackwell systems," DePue stated in his public release. He noted that the forensic method matched deterministic initialization patterns to specific chip and seed combinations.[4]

The more powerful Grace Blackwell GB200 and GB300 systems, which utilize 152 streaming multiprocessors, matched four other models. DeepSeek-V4-Pro, Moonshot AI's Kimi K3 and Kimi K3.1-Pro, and Ant Group's Ling-3.1-flash all carried the distinct 152-core signature.[1][4]

To verify the statistical findings, the researchers tested the recovered seeds on physical hardware. Using a Blackwell B300 chip, they successfully reproduced the untouched starting weights of DeepSeek-V4-Flash bit for bit, while a rented GB200 system perfectly recreated the preserved rows of Kimi K3.[1][4]

Scale of the Training Operations

The statistical matches are exceptionally strong, ranging from 49 to 1,385 standard deviations against chance. When tested on Nvidia's older Hopper H100 chips, the initialization patterns failed to reproduce, confirming that the models were specifically initialized on the newer Blackwell architecture.[1][4]

The report estimates that the five main training runs required approximately 12,000 Blackwell GPUs in total. Kimi K3 alone likely utilized about 4,500 Blackwell processors over a 70-day period, while DeepSeek-V4-Pro required an estimated 3,000 chips for its development cycle.[1][3]

At an estimated unit price of $70,000 per GPU, the hardware required for these training clusters represents a massive capital investment. The largest individual cluster would cost between $200 million and $300 million, pushing the combined hardware value past half a billion dollars.[1][3]

The analysis also reveals a clear chronological shift in the hardware used by these Chinese laboratories. Older models from the same companies, such as DeepSeek-V3 and Kimi K2.5, carry the fingerprint of Nvidia's previous-generation Hopper chips, aligning with their earlier release dates.[1][3]

Export Controls and Offshore Compute

The U.S. government strictly prohibits the export of Blackwell GPUs to China in an effort to curb the country's artificial intelligence capabilities. However, the forensic evidence of Blackwell initialization does not definitively prove that export control laws were violated by the developers.[1][2]

The chips could be physically located in offshore data centers outside of China, which has become a popular strategy for Chinese firms seeking advanced compute. Recent estimates suggest hundreds of thousands of Nvidia AI accelerators have been imported into neighboring countries like Malaysia.[2][4]

While the fingerprint confirms the hardware used at initialization, it cannot definitively prove that the entire training run occurred on Blackwell chips. However, researchers note that transferring a massive model to different hardware immediately after initialization would be highly unusual and inefficient.[1][3]

The researchers also considered whether specially configured AMD hardware, which is also restricted from export to China, could have produced the patterns. By analyzing vendor-differentiating values, they conclusively ruled out AMD's MI300X for DeepSeek-V4-Flash and deemed it highly unlikely for the others.[1][4]

AI

Geopolitical and Corporate Fallout

The technical revelations arrive during a period of intense scrutiny for Nvidia and its corporate leadership. The company's chief executive, Jensen Huang, recently resigned following an investigation into his relationship with an executive at Megaspeed, a firm implicated in smuggling chips to China.[2][3]

The crisis has prompted lawmakers to demand stricter oversight of the agencies responsible for enforcing U.S. export controls. The Trump administration official overseeing these programs is now facing intense pressure to explain how Chinese developers continue to access the most advanced American technology.[2][5]

Senior researchers at both OpenAI and Anthropic have independently verified DePue's forensic methodology, adding significant weight to the findings. The technique establishes that open-weight AI releases carry permanent, verifiable hardware signatures that developers cannot easily hide without degrading model quality.[3][4]

The four Chinese laboratories identified in the report have not publicly responded to the forensic analysis. Nvidia has also declined to comment on the specific findings, though the company previously maintained that its offshore data center partners operate within legal boundaries.[1][2]

Implications for Global AI Policy

The discovery fundamentally shifts the debate over U.S. export controls from intelligence anecdotes to reproducible technical data. Policymakers now have concrete evidence that physical shipment restrictions are failing to prevent Chinese frontier labs from utilizing the latest generation of American AI hardware.[1][4]

AI

The findings highlight the growing challenge of regulating compute access in a globally connected cloud infrastructure. As long as Chinese developers can rent Blackwell clusters in unrestricted jurisdictions, physical export bans will struggle to contain the proliferation of advanced artificial intelligence models.[2][4]

Not all Chinese developers have made the transition to the restricted hardware. The newest readable models from Alibaba, Tencent, and Xiaomi still produce initialization readings consistent with the older Hopper generation, suggesting that Blackwell access remains concentrated among a few leading labs.[1][3]

The forensic technique developed by Unemployment Labs provides a new tool for monitoring global AI development. By analyzing the permanent signatures left in model weights, researchers can now independently track the deployment of advanced semiconductors across the international artificial intelligence landscape.[1][4]

The reliance on offshore data centers underscores a significant loophole in the current regulatory framework. While the physical chips may never cross the Chinese border, the computational power they generate is seamlessly transmitted back to developers in Beijing and Hangzhou.[2][5]

As the U.S. government evaluates its next steps, the focus is rapidly shifting toward identifying the specific procurement channels and cloud providers facilitating this access. The ability to trace models back to their silicon origins will likely become a standard component of future export enforcement.[1][5]