DeepSeek is developing its own AI chip, sources say

BEIJING: Chinese startup DeepSeek is working on an artificial intelligence chip of its own, according to three people familiar with the matter, in a move that could lessen its dependence on chips from Nvidia and Huawei that it has used to train and run its widely known models.

The people said the proposed chip is being designed for inference, the phase in which a trained AI model produces answers for users, rather than for training fresh models. They said the project is still at an early stage, with the Hangzhou-based company in contact with outside partners and in discussions with firms involved in chip design, foundry work and memory.

If the effort succeeds, it would represent a significant strategic shift for DeepSeek, which has been known more for advances in AI models than for commercialising its technology. The company began the effort about a year ago. It has also quietly stepped up recruitment of chip-design engineers in recent months without posting openings on public hiring platforms, according to two of the people. All three spoke on condition of anonymity because the plans are not public.

DeepSeek did not respond to a request for comment.

Shift in China’s AI chip landscape

DeepSeek gained international attention more than a year ago when it released two highly efficient AI models that spread quickly around the world and surprised many in Silicon Valley and Washington. In China, the company has been widely seen as a leading standard-bearer for the country’s AI ambitions.

Its move into semiconductor development would come as China’s domestic AI chip market changes. While Huawei’s products remain well behind Nvidia’s most advanced chips, US export restrictions on those products have helped Huawei capture about half of China’s $50 billion AI chip market, supplying DeepSeek and several other major players.

But Huawei’s position has already been coming under pressure as Alibaba and Baidu develop AI chips of their own and expand their share of the market. A DeepSeek entry into chip design could therefore add another challenger in the local market.

US-based Nvidia’s shares slipped about 1.6pc in premarket trading after the report. Analyst Richard Windsor of Radio Free Mobile said the development would not materially affect the American chipmaker.

"Nvidia is at zero in China and staying ​there. DeepSeek has almost no chance of selling silicon outside of China unless it gets access to leading edge manufacturing"

Windsor added that the development does not affect Nvidia.

Export controls and hardware strategy

DeepSeek’s chip effort mirrors a broader global push by AI developers to gain more control over the hardware used for their systems and reduce reliance on Nvidia. Last month, OpenAI introduced Jalapeno, described as its first custom inference chip developed with Broadcom, while Reuters reported in April that Anthropic had been considering building its own AI chips.

For DeepSeek, the push also carries a strategic dimension because US export controls prevent Chinese firms from buying Nvidia’s top-end chips, while Beijing has been urging major technology companies to develop domestic alternatives. In a rare interview with a Chinese media outlet in 2024, DeepSeek founder Liang Wenfeng said chip export controls posed a challenge for the company.

DeepSeek has used both Nvidia and Huawei chips. The company has said the foundation model behind R1 — the reasoning model whose low-cost performance contributed to a sell-off in US technology stocks in January 2025 — was trained on Nvidia’s H800 chip, a China-specific product that Washington banned in late 2023.

Since then, DeepSeek has relied more heavily on Huawei. In April, it released its V4 model adapted for Huawei’s Ascend chips, and Huawei said its processors were used in part of the training for V4-Flash, a lighter version of that model. Reuters has also reported that orders for Huawei’s Ascend 950 chips from Chinese technology conglomerates jumped after the launch.

Inference demand and obstacles

An inference chip from DeepSeek would target the fastest-growing part of AI computing demand. As AI tools spread more widely, a greater share of computing is moving away from model training and toward running models, which depends on specialised chips that can be cheaper and use less power than general-purpose GPUs.

Even so, the project faces major hurdles. Building a competitive AI chip generally requires years of work and large amounts of capital. Manufacturing is another challenge because US restrictions stop Chinese chip designers from using the most advanced overseas foundries, while separate curbs have limited China’s access to high-bandwidth memory, an important component for AI inference chips.

The chip push comes as DeepSeek also prepares to take outside funding for the first time. Reuters reported in June that the company was set to raise $7 billion in an initial funding round that would value it at between $52 billion and $59 billion, reversing its longstanding approach of rejecting external investment.