China’s DeepSeek Trained AI Model on Banned Nvidia Chips, Official Confirms

Despite stringent U.S. export controls aimed at curbing China’s artificial intelligence advancement, Chinese AI startup DeepSeek has reportedly trained its latest large language model using Nvidia’s most advanced AI chips—technologies explicitly banned for export to China under recent sanctions. According to Reuters, DeepSeek’s upcoming model, slated for release next week, was developed on Nvidia’s H100 or potentially even the newer B100 series, both of which are subject to strict U.S. Department of Commerce restrictions due to their unparalleled computational power for generative AI workloads.

The revelation, first reported by Reuters and corroborated by MSN, comes from an official with direct knowledge of DeepSeek’s infrastructure procurement. While the exact method of acquisition remains undisclosed, experts speculate that the chips may have been sourced through third-party intermediaries, gray-market distributors, or previously stockpiled inventory acquired before the tightening of export controls. The use of such hardware represents a significant technological leap for China’s AI sector, enabling DeepSeek to compete directly with U.S.-based models like OpenAI’s GPT-4o and Google’s Gemini Ultra.

U.S. sanctions, implemented since 2022 and further tightened in October 2023, prohibit the sale of high-end AI chips with compute capabilities exceeding 4,800 teraFLOPS to Chinese entities without a special license. Nvidia’s H100, which delivers over 10,000 teraFLOPS in FP8 performance, falls far beyond this threshold. The fact that DeepSeek, a relatively young company founded in 2023, has managed to access and deploy these chips at scale suggests a sophisticated, possibly state-supported, supply chain operation. This development undermines the efficacy of export controls designed to slow China’s AI progress and could prompt renewed diplomatic pressure from Washington.

DeepSeek has not publicly confirmed the source of its hardware, nor has it responded to requests for comment. However, internal documents referenced by the official cited in the Reuters report indicate that the company’s training cluster, located in a data center in Hangzhou, was specifically configured to accommodate Nvidia’s latest architectures. The model, tentatively named DeepSeek-V3, reportedly achieved state-of-the-art performance on multiple benchmarks, including MMLU, GSM8K, and HumanEval, rivaling models trained on U.S. cloud platforms.

The implications extend beyond corporate competition. The ability of Chinese firms to bypass sanctions signals a broader trend of technological decoupling. As the U.S. seeks to maintain its AI dominance through hardware restrictions, China is accelerating domestic alternatives—such as Huawei’s Ascend 910B and Biren’s BR100—while simultaneously exploiting loopholes in global supply chains. This dual-track strategy may ultimately render export controls obsolete unless international allies coordinate more effectively on chip trafficking and end-use monitoring.

Investing.com, while not directly reporting on DeepSeek’s hardware sourcing, highlighted broader trends in China’s AI investment surge, noting that over $12 billion in private and state-backed funding flowed into Chinese AI startups in 2024 alone. DeepSeek’s success, if confirmed, could catalyze a new wave of innovation and investment, further narrowing the gap between U.S. and Chinese AI capabilities. Analysts warn that without multilateral enforcement, the U.S. risk of falling behind in foundational AI infrastructure may become irreversible.

As global tech leaders prepare for the next AI arms race, the DeepSeek case underscores a fundamental truth: in the age of artificial intelligence, control over hardware is as critical as control over algorithms. The world now watches to see whether diplomatic, economic, or technological responses will follow—or whether the era of unimpeded AI proliferation has already begun.