issued Tuesday, July 7, 2026 · 08:42 PM
[BEIJING] Chinese startup DeepSeek is developing its own AI chip, which could reduce the company’s reliance on Nvidia and Huawei chips to train and run its globally popular models, according to three people familiar with the matter.
The chip is designed for inference, the stage of AI computing where a trained model generates a response to the user, rather than for training new models, the sources said.
If successful, DeepSeek’s expansion into semiconductor development would mark a major strategic shift for the company widely hailed as a champion of AI in China, and could add to the challenges facing Chinese technology giant Huawei.
Shares of U.S.-based Nvidia fell about 2% in premarket trading.
DeepSeek gained global fame over a year ago after releasing two highly efficient AI models. This spread around the world and surprised many people in Silicon Valley and Washington.
The company has long been known for focusing on breakthroughs in AI models over commercializing the technology.
Although Huawei’s products still lag far behind Nvidia’s cutting-edge chips, the U.S. ban on exports to China has given Huawei roughly half of the US$50 billion domestic AI chip market, supplying DeepSeek and several other major industry players.
But Huawei’s market power is already waning as technology rivals Alibaba and Baidu develop their own AI chips and gain market share.
DeepSeek’s efforts to join the race are still in the early stages, with the company reaching out to outside partners and holding discussions with chip design, foundry and memory companies, the three people said. The effort began about a year ago, one of the people said.
Related items
The Hangzhou-based company has been hiring more chip design engineers in recent months, but the hiring is done privately without posting jobs on public hiring platforms, two people said.
All three declined to be named because the information is private. Despite becoming the standard-bearer for China’s AI ambitions, DeepSeek has remained under the radar. The company did not respond to requests for comment.
Follow global trends
By adopting its own chips, DeepSeek, like other global AI developers, aims to gain more control over the hardware behind its models and reduce its dependence on Nvidia.
OpenAI announced its first custom inference chip, Jalapeno, developed with Broadcom last month, but Anthropic is considering building its own AI chip, Reuters reported in April.
For DeepSeek, this effort adds a strategic dimension. U.S. export controls have barred Chinese companies from buying Nvidia’s cutting-edge chips, and Beijing has put pressure on technology advocates to develop domestic alternatives.
DeepSeek founder Liang Wenfeng said in a rare interview with Chinese media in 2024 that chip export controls were a challenge for the company.
DeepSeek uses chips from both Nvidia and Huawei. The company said the underlying model behind R1, the inference model whose low cost performance triggered the U.S. tech stock crash in January 2025, was trained on Nvidia’s H800, a chip designed for the Chinese market that the U.S. government banned in late 2023.
Since then, the company has increased its reliance on Huawei. In April, the company released a V4 model compatible with Huawei’s Ascend chip, and Huawei said its processors were used for part of the training for V4-Flash, a lighter version of the model. Orders for Huawei’s Ascend 950 chip from Chinese tech conglomerates soared after its launch, Reuters reported.
Leveraging inference demands
DeepSeek inference chips will target the fastest growing segment of AI computing demand. As AI applications become more prevalent, much of the industry’s computing work is shifting from model training to model execution, relying on specialized chips that are cheaper and consume less power than general-purpose GPUs.
However, there is no guarantee of success. Designing a competitive AI chip typically takes several years and significant capital. Manufacturing faces new hurdles as the United States bans Chinese designers from accessing cutting-edge overseas foundries, while another U.S. regulation cuts off China’s access to high-bandwidth memory, a vital component of AI inference chips.
DeepSeek’s chip push coincides with the company’s first infusion of outside capital. The company plans to raise $7 billion in its first round of funding, Reuters reported in June, valuing it between $52 billion and $59 billion, reversing its long-standing strategy of refusing outside investment. Reuters
