DeepSeek Trained AI Model on Nvidia Blackwell Despite U.S. Export Ban, Sources Reveal

Despite a sweeping U.S. export ban targeting advanced AI semiconductors, Chinese artificial intelligence startup DeepSeek has reportedly trained its latest generative AI model on Nvidia’s most powerful chip, the Blackwell architecture, according to a source familiar with the matter cited by Reuters. The model, expected to be publicly released as early as next week, represents a significant technological leap for China’s AI sector and underscores growing challenges in enforcing global technology controls.

The Blackwell GPU, unveiled by Nvidia in late 2023, is engineered for exascale AI training and boasts unprecedented computational density, memory bandwidth, and energy efficiency. U.S. authorities imposed strict export controls on Blackwell and its predecessors in October 2023, aiming to curb China’s ability to develop next-generation AI systems capable of military, surveillance, and strategic applications. The chips are subject to licensing requirements under the Export Administration Regulations (EAR), and unauthorized transfers are considered violations of national security policy.

Yet, according to Reuters, DeepSeek — a relatively young but rapidly ascending player in China’s AI ecosystem — managed to acquire and deploy Blackwell chips for training its latest model. The startup, known for open-weight models like DeepSeek-V2 and DeepSeek-MoE, has previously emphasized efficiency and cost-effectiveness in model development. This latest development suggests either a sophisticated supply chain maneuver, third-party intermediary involvement, or potential exploitation of loopholes in export enforcement mechanisms.

While the exact method of acquisition remains unconfirmed, analysts speculate that DeepSeek may have obtained the chips through third-party distributors, gray-market channels, or by leveraging existing inventory in jurisdictions with lax enforcement. Some experts suggest the company may have repurposed chips originally intended for non-AI applications, such as high-performance computing or data center workloads, which fall under less restrictive export categories. Others point to the possibility of reverse-engineering or chip modification techniques to evade detection.

The U.S. Department of Commerce has not yet issued an official comment on the report, but internal sources indicate that the Bureau of Industry and Security (BIS) is reviewing the case. A senior official familiar with the matter told Reuters that the agency is “actively investigating potential violations” and assessing whether the deployment constitutes a breach of export control regulations.

DeepSeek has not publicly responded to the allegations. However, the company’s recent public disclosures highlight its focus on open-source AI development and cost-efficient scaling — positioning itself as a challenger to U.S.-based giants like OpenAI and Anthropic. Its ability to train on Blackwell hardware would significantly narrow the performance gap between Chinese and American AI models, particularly in multilingual reasoning and long-context understanding.

The incident reignites debates over the effectiveness of unilateral export controls in an increasingly interconnected global tech landscape. Critics argue that such measures accelerate China’s push for domestic semiconductor self-reliance, while proponents maintain they are necessary to preserve U.S. technological superiority. Meanwhile, Nvidia, which has largely complied with U.S. regulations, has declined to comment on specific customer usage.

As the global AI race intensifies, this case may set a precedent for how nations enforce — or fail to enforce — technological embargoes. If confirmed, DeepSeek’s use of Blackwell chips could prompt tighter controls, expanded sanctions, or even new multilateral frameworks to prevent circumvention. For now, the world watches as one of China’s most promising AI startups quietly bridges the gap between policy and innovation — and perhaps redefines the boundaries of what’s possible under restriction.