DeepSeek Withholds V4 Model from Nvidia and AMD, Fueling US-China AI Rivalry

DeepSeek Withholds V4 Model from Nvidia and AMD, Fueling US-China AI Rivalry

Chinese AI lab DeepSeek has broken from industry norms by not providing U.S. chip giants Nvidia and AMD with pre-release access to its upcoming flagship model, V4. Instead, the company has reportedly granted early access to domestic suppliers like Huawei, signaling a new phase in the escalating tech rivalry between the United States and China.

A Break from Industry Practice

Typically, AI developers provide early versions of their new models to leading chipmakers like Nvidia and AMD. This collaboration is crucial for optimizing the software to run efficiently on the most widely used hardware. DeepSeek itself has previously worked closely with Nvidia's technical team.

However, for its V4 model, which was anticipated to be released around the Lunar New Year, DeepSeek has excluded the American chipmakers from this process. While no official reason has been given, the decision is seen as part of a broader Chinese strategy to foster technological self-reliance.

Collaboration with Huawei and U.S. Concerns

In place of Nvidia and AMD, DeepSeek has given a head start of several weeks to Chinese chipmakers, including Huawei, to optimize the V4 model for their processors.

This move has been met with concern from U.S. officials. One senior administration official noted that DeepSeek may have trained its V4 model on Nvidia's most advanced AI chip, the Blackwell, within a cluster in mainland China, which would appear to violate U.S. export controls. The official also suggested that DeepSeek might attempt to conceal its use of American chips and publicly claim that it used Huawei's hardware for training.

The Intensifying US-China AI Competition

DeepSeek's decision is likely more than just a corporate strategy; it aligns with the Chinese government's broader goal of cultivating a domestic AI ecosystem and reducing reliance on American technology. The influence of Chinese AI models is growing, with DeepSeek's models having been downloaded over 75 million times on the open-source platform Hugging Face.

This development also comes as DeepSeek faces other controversies. Its U.S.-based rival, Anthropic, has accused DeepSeek and other Chinese firms of improperly extracting data from its flagship model, Claude. This technique, known as a "distillation attack," involves using fake accounts to interact with a rival AI millions of times to learn from its outputs and improve their own models. DeepSeek has not yet publicly responded to these allegations.

The AI competition between the U.S. and China is becoming increasingly complex, encompassing not only the pace of technological innovation but also supply chain control and ethical considerations.