SI Agent News
RumorPolicyChips

Evidence suggests banned Nvidia chips powered China's top AI models

Research into model weights suggests Chinese firms are bypassing U.S. export controls to train frontier AI on restricted Blackwell GPUs.

275,000Nvidia AI accelerators imported to Malaysia in 2024 and 2025

The short version

Why it matters

If Chinese labs can successfully train frontier models on restricted hardware, U.S. efforts to maintain a technological edge may be failing. This creates significant corporate liability for Nvidia and raises questions about the effectiveness of current trade policies.

Fingerprints in the weights

A new research effort has identified evidence that leading Chinese AI models were trained using Nvidia chips restricted by the U.S. government. By analyzing the weights—the unique internal structures of large language models—researchers found specific configurations that match banned hardware. These findings apply to several high-profile models released by companies including DeepSeek, Moonshot AI, MiniMax, and Ant Group.

Will DePue, a former OpenAI researcher, led the analysis of these open-weight models. His findings suggest that the most closely watched releases, such as Moonshot’s Kimi’s K3 and DeepSeek’s V4 models, bear clear fingerprints from Blackwell GPUs. These chips are currently banned from being exported to China or any Chinese-owned entities by the U.S. government.

The discovery highlights a significant gap between policy and practice in the AI arms race. While the U.S. aims to protect its technological leadership, the data suggests that Chinese frontier labs are successfully obtaining the high-end compute they need. This access allows them to develop sophisticated AI capabilities despite the official trade restrictions currently in place.

Smuggling and offshore hubs

The path to these chips often involves complex logistics designed to hide their final destination. Epoch AI, a nonprofit research organization, estimates that advanced chips entered China in 2024 and 2025 through various smuggling schemes. These methods are intended to circumvent the strict rules governing the sale of high-performance semiconductors to geopolitical rivals.

One popular method involves the use of offshore data centers, which has sparked an industrial revolution in Malaysia. Epoch AI data indicates that 275,000 Nvidia AI accelerators were imported to Malaysia during the same period. By placing chips in these third-party locations, Chinese firms can access massive amounts of compute power without the hardware technically entering mainland China.

Critics argue that these loopholes undermine the primary goal of U.S. export controls. Sam Bresnick of the Georgetown Center for Security and Emerging Technology noted that it does not make sense to allow companies to simply move to Malaysia to do what is banned in China. This creates a situation where the hardware is accessible even if the official shipping destination is restricted.

Corporate and political fallout

The situation has created significant personal and professional complications for Nvidia leadership. CEO Jensen Huang recently left the company after admitting to an affair with Alice Huang, an executive at Megaspeed. Megaspeed is a company implicated in smuggling Nvidia chips into China through its subsidiary, Speedmatrix, and has been investigated by both U.S. and Singaporean authorities.

Nvidia and Huang have previously vouched for Megaspeed’s data centers, claiming that company officials visited them to confirm they were legitimate operations. However, the potential for corporate liability regarding export control violations may have influenced Huang's departure from the company. The company faces scrutiny over how it manages its relationship with customers who may be violating federal laws.

Government officials are also facing pressure regarding the oversight of these exports. Secretary of Commerce Howard Lutnick, who oversees the agencies licensing chip exports, has faced reports that his investigators failed to uncover suspicious buyers. These reports suggest that some officials may have been too cozy with Nvidia representatives, potentially allowing smuggling operations to continue unchecked.

The future of the chip trade

The long-term impact of these trade restrictions remains uncertain as both the U.S. and China react to the current landscape. Some analysts argue that the bans might eventually drive China to recommit to its domestic chip industry. If China succeeds in creating its own high-end hardware, billions of dollars in demand for Nvidia chips could eventually dry up.

Other theories suggest that the current hold-up of shipments might be related to broader trade negotiations. Some observers believe that the lack of clearance for certain shipments could be a preparation for a future trade deal involving large volumes of U.S. goods. There are also concerns regarding the potential for surveillance technology to be embedded within the chips themselves.

For now, U.S. companies have limited obligations to track what their customers do with controlled goods once they are sold. However, the evidence of Blackwell chips in Chinese models suggests that the current system is being bypassed. The next phase of this conflict will likely involve stricter controls on who uses the chips, rather than just who owns them or where they are physically installed.

What to watch

Try it

Look up the 'Blackwell' GPU specifications to see which specific hardware is currently restricted from export to China.

Sources and links

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

Listen to the AI Wire headlines

Listen to the AI Wire

Listen to our daily AI shows every morning: on your commute, in the shower, wherever you are

Listen to the daily shows

First reported by TechCrunch · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

More AI news