DeepSeek moves beyond AI models with custom chip project
The startup is developing an inference chip as Beijing pushes domestic AI companies to reduce reliance on foreign technology.
Chinese AI startup DeepSeek is developing its own artificial intelligence chip, according to three people familiar with the matter, in a move that could reduce its reliance on Nvidia and Huawei chips, which it has used to train and run its globally popular AI models.
The chip is being designed for inference, the stage of AI computing in which trained models generate responses for users, rather than for training new models, the sources said.
If successful, the effort would mark a major strategic shift for DeepSeek, which has been widely hailed as China's AI champion, while potentially intensifying competition for Huawei in China's rapidly evolving AI chip market.
DeepSeek rose to global prominence more than a year ago after releasing two highly efficient AI models that gained worldwide attention, surprising many in Silicon Valley and Washington with their performance and efficiency.
The company has long focused on advancing AI models rather than commercializing its technology or developing hardware.
Although Huawei's AI chips still trail Nvidia's most advanced processors, U.S. export restrictions have prevented Chinese companies from purchasing Nvidia's latest products, allowing Huawei to capture roughly half of China's estimated $50 billion AI chip market by supplying companies including DeepSeek.
However, Huawei's dominance is facing growing competition as Chinese technology giants Alibaba and Baidu develop their own AI chips and expand their market presence.
DeepSeek's chip initiative remains in its early stages. According to the sources, the Hangzhou-based company has been holding discussions with chip design firms, foundries, and memory suppliers while quietly expanding its chip engineering team without advertising positions publicly. One source said development began about a year ago.
The sources requested anonymity because the project has not been publicly disclosed. DeepSeek declined to comment.
Following a global trend
With an in-house AI chip, DeepSeek would join a growing number of AI developers seeking greater control over the hardware powering their models while reducing dependence on Nvidia.
Last month, OpenAI unveiled Jalapeno, its first custom inference chip developed with Broadcom. Reuters also reported in April that Anthropic has been evaluating the development of its own AI chips.
For DeepSeek, however, the effort carries additional strategic importance. U.S. export controls prohibit Chinese companies from purchasing Nvidia's most advanced AI chips, while Beijing has been encouraging domestic technology companies to build alternatives.
Founder Liang Wenfeng acknowledged in a rare 2024 interview with Chinese media that U.S. chip export restrictions posed a significant challenge for the company.
DeepSeek has relied on both Nvidia and Huawei hardware. The company previously said the foundation model behind its R1 reasoning model, which triggered a selloff in U.S. technology stocks after its January 2025 debut, was trained using Nvidia's H800 chips, a China-specific product that Washington later banned.
Since then, DeepSeek has increasingly shifted toward Huawei. In April, it released its V4 model optimized for Huawei's Ascend processors, while Huawei said its chips were used in part of the training of V4-Flash, a lighter version of the model. Reuters previously reported that orders for Huawei's Ascend 950 chips surged following those releases.
Betting on inference
DeepSeek's chip would target inference, the fastest-growing segment of AI computing. As AI adoption accelerates, a growing share of computing demand is shifting from training models to running them, increasing demand for specialized inference processors that are generally cheaper and more power-efficient than general-purpose GPUs.
The project, however, faces significant hurdles. Designing a competitive AI chip typically requires years of development and billions of dollars in investment. Manufacturing is also constrained by U.S. restrictions that prevent Chinese chip designers from using the world's most advanced overseas foundries, while separate export controls limit China's access to high-bandwidth memory, a key component in advanced AI chips.
DeepSeek's semiconductor ambitions come as the company prepares to accept outside investment for the first time. Reuters reported in June that it is planning to raise $7 billion at a valuation of between $52 billion and $59 billion, reversing its longstanding policy of rejecting external capital.
