Why DeepSeek is hiding new AI model from US chipmakers
What's the story
DeepSeek, a Chinese artificial intelligence (AI) lab, has broken from industry norms by not showing its upcoming flagship model to US chipmakers such as NVIDIA and AMD. The move is in stark contrast to the standard practice of sharing pre-release versions with leading companies for performance optimization. Instead, DeepSeek has given early access to domestic suppliers like Huawei Technologies.
Unconventional approach
DeepSeek's departure from industry norms
AI developers usually share pre-release versions of major models with top chipmakers such as NVIDIA and AMD. This is done to ensure their software runs efficiently on widely used hardware. However, DeepSeek has not followed this norm for its upcoming model, which was expected to launch around the Lunar New Year holiday. Instead, it gave Chinese chipmakers such as Huawei a head start of several weeks to optimize the software for their processors.
Market implications
New AI coding tools are cutting down optimization time
Ben Bajarin, CEO of research firm Creative Strategies, downplayed the impact on NVIDIA and AMD. He said most enterprises are not running DeepSeek, which serves as a benchmarking model more than anything else. Bajarin also noted that new AI coding tools are cutting down the time to optimize software for hardware from months to weeks. DeepSeek's move could be part of a larger strategy by the Chinese government "to try to keep US hardware and models disadvantaged" in China.
Export concerns
DeepSeek may try to hide use of US chips
A senior Trump administration official told Reuters that DeepSeek's latest AI model was trained on NVIDIA's most advanced chip, Blackwell, using a cluster in mainland China. This move appears to violate US export controls. The company may try to remove technical indicators showing its use of American AI chips and plans to publicly claim it used Huawei's chips to train its model, according to the US official.
Model influence
Surge in Chinese open-source models
DeepSeek's models have been downloaded over 75 million times on the open-source platform Hugging Face since its debut in January 2025. This has sparked a wave of Chinese open-source models challenging US AI labs. In the last year, downloads for Chinese models have outpaced those from any other country on the platform. The rapid growth of these models has reignited discussions in Washington about exporting advanced US AI chips to China.