DeepSeek, the Chinese artificial intelligence lab that shook global markets with its low-cost AI model last year, has broken with industry norms by withholding its upcoming flagship AI model, V4, from major U.S. chipmakers, two sources familiar with the matter told Reuters. Instead, early access has been given to domestic suppliers, including Huawei Technologies, signaling a sharp pivot in the lab’s global strategy.
Typically, AI developers share pre-release versions of major models with leading chipmakers such as Nvidia and Advanced Micro Devices (AMD) to ensure compatibility and performance on widely used hardware. DeepSeek has previously collaborated closely with Nvidia’s technical teams. For its forthcoming update, originally expected around the Lunar New Year, U.S. chipmakers were left out while Chinese firms received a head start of several weeks to optimize the software for local processors. Nvidia and AMD declined to comment, and DeepSeek and Huawei did not respond to inquiries.
Industry analysts suggest the move could have broader geopolitical motivations. Ben Bajarin, CEO of research firm Creative Strategies, said the decision is likely part of a strategy to maintain a technological edge for China while keeping U.S. hardware and software at a disadvantage in the domestic market. He noted that the impact on Nvidia and AMD is limited because most enterprises do not run DeepSeek models at scale, and AI coding tools now reduce the time required to optimize software on hardware from months to weeks.
The development also raises potential regulatory concerns. A senior Trump administration official told Reuters that DeepSeek’s latest AI model was trained using Nvidia’s most advanced Blackwell chip within mainland China, a move that may violate U.S. export controls. The official added that DeepSeek may attempt to obscure evidence of U.S. chip usage and instead publicly attribute the training to Huawei hardware.
DeepSeek has become a dominant force in the Chinese open-source AI sector. Its models have been downloaded over 75 million times on Hugging Face since the company emerged in January 2025, surpassing downloads of AI models from any other country. This surge has intensified debate in Washington over the export of advanced U.S. AI chips to China. While Nvidia’s H20 and AMD’s MI308 inference chips have been allowed for export under limited licenses, the status of DeepSeek’s access to these processors remains unclear.
The H20 and MI308 chips are critical for running trained AI models, and demand for them is high; AMD reported $390 million in sales for the MI308 in its most recent quarter. DeepSeek is among several Chinese AI firms expected to announce new model releases in the coming weeks, continuing the rapid expansion of China’s AI capabilities and raising fresh questions about technology transfer and global AI competition.
