SI Agent News
RumorOpen modelsChipsPolicyResearch

Traces of banned Nvidia Blackwell chips found in Chinese AI models

Researchers identified hardware signatures in leading Chinese AI models, raising questions about how export restrictions are being bypassed.

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

The short version

Why it matters

This discovery highlights the difficulty of enforcing hardware export bans when companies can utilize international data centers. It signals a potential shift toward stricter regulations targeting the end-users of technology rather than just the shipping destination.

Hardware signatures in AI

Researchers have discovered evidence that leading Chinese AI models were trained using Nvidia Blackwell chips. These specific chips are currently restricted from being exported to China or to firms owned by the Chinese government. The findings come from an analysis of model weights, which are the internal parameters that define how a large language model processes information.

Because models retain distinctive structures based on the hardware configurations used during their training, these traces act as a digital fingerprint. Will DePue, a former OpenAI researcher, identified these signatures in several high-profile models. His study included models from major Chinese entities such as DeepSeek, Moonshot AI, MiniMax, and Ant Group.

The analysis specifically highlighted the V4 models from DeepSeek and the K3 releases from Moonshot AI. While the findings do not prove that the training relied entirely on Nvidia hardware, they provide compelling evidence of Blackwell use. This discovery suggests that the technological gap between U.S. and Chinese AI capabilities may be narrowing despite current trade barriers.

Bypassing export restrictions

The presence of these chips in Chinese models raises significant questions about how export restrictions are being circumvented. One primary method for bypassing these rules is the use of overseas data centers. Chinese companies have increasingly accessed computing capacity in other countries to train their models without direct shipment to Chinese soil.

Malaysia has emerged as a notable destination for this type of infrastructure. According to estimates from the nonprofit Epoch AI, Malaysia imported 275,000 Nvidia AI accelerators across 2024 and 2025. These facilities allow Chinese firms to utilize high-end American technology while technically remaining outside the immediate reach of domestic export bans.

Epoch AI also estimates that roughly two million advanced chips reached China during this same two-year period. These shipments were reportedly made through arrangements that concealed their ultimate destination. This suggests that the current system of monitoring the final location of hardware is being successfully undermined by complex logistics.

U.S. policy and oversight

The United States maintains these export restrictions primarily to preserve its technological advantage in artificial intelligence. However, the current rules face criticism for being too easy to evade. Some advocates for tougher controls are now calling for regulations that cover the actual users of the chips rather than just the shipping routes or ownership structures.

Current U.S. companies face limited requirements to monitor how their customers ultimately use controlled products like AI chips. This lack of oversight allows for the

Specific model findings

The evidence of Blackwell hardware use was identified by researcher Will DePue through a detailed examination of several prominent models. His analysis specifically pinpointed traces in the V4 models released by DeepSeek as well as the K3 releases from Moonshot AI's Kimi series.

DePue, who previously worked at OpenAI and now leads the nonprofit Unemployment Labs, also looked into models from MiniMax and Ant Group. His work suggests that these specific high-profile Chinese models are the ones most likely utilizing the restricted hardware despite the current trade barriers.

Political and trade tensions

The debate over how to handle these exports has reached the highest levels of government. While China officially objected to new semiconductor restrictions on Friday, House Speaker Mike Johnson noted that there is growing support within Congress for even tighter controls on chip technology.

The Trump administration currently allows for the export of the less powerful H200 chip, provided it follows an indirect shipping route that enables the U.S. to collect a 25% tariff. This creates a complex landscape where some hardware is permitted under specific financial and logistical conditions while others remain strictly banned.

Accountability and the Entity List

Internal government friction has also emerged regarding which companies should be blocked from accessing American technology. Treasury Secretary Scott Bessent told CNBC that the Commerce Department was responsible for failing to include Megaspeed on the Entity List, which would have restricted its access to U.S. tech.

Bessent stated that the Treasury, Defense, and State departments, along with the U.S. trade representative, had all supported adding the company to that list. This highlights a lack of coordination in how the government manages the specific entities that are allowed to receive sensitive hardware.

What to watch

Try it

Look up the current U.S. Entity List to see which companies are currently restricted from receiving American semiconductor technology.

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 Technobezz · 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