Chinese artificial intelligence start-up DeepSeek is advancing plans to develop its own AI chips, aiming to reduce its reliance on Nvidia and Huawei chips and gain a foothold in the country's rapidly expanding AI inference chip market. According to people familiar with the matter, the project is still in its early stages, but has become an important part of the strategic transformation of the company regarded as "China's AI champion".
Multiple people close to the company said that DeepSeek is designing an AI chip specifically for the inference stage, focusing on scenarios such as model deployment and online response to user requests, rather than model training. With the popularity of generative AI applications, the industry's computing power demand is gradually shifting from the training stage to the inference stage. Such specialized chips are usually more energy-efficient and cost-effective than general-purpose GPUs.
If this attempt is successful, DeepSeek will not only join the ranks of self-developed chips by global AI companies such as OpenAI, but will also respond to China's policy orientation of promoting "domestic substitution" under the pressure of export controls. The U.S. government previously restricted Nvidia from exporting high-end GPUs to China, allowing Huawei to rapidly expand its share of the Chinese AI chip market and become an important supplier to many leading companies, including DeepSeek.
DeepSeek became famous around the world with its efficient large models. The inference model R1 it launched triggered sharp fluctuations in U.S. technology stocks in early 2025 due to its low cost and high performance. The company has publicly stated that the basic model supporting the model uses the NVIDIA H800 chip customized for the Chinese market during the training phase. This chip has been further tightened by the United States in 2023. As restrictions escalate, DeepSeek gradually shifts more computing power to Huawei's Ascend series, and will launch the V4 model optimized for Huawei chips and its lightweight version V4-Flash in 2026.
However, Huawei’s advantage in the domestic AI chip market is also facing challenges from other Chinese technology companies. Alibaba and Baidu have respectively promoted self-developed AI chip solutions to compete for more shares in data centers and cloud service scenarios. People familiar with the matter said that DeepSeek's attempt to develop its own inference chips was launched under this competitive landscape. In the past year, the company has been in contact with a number of chip design companies, wafer foundries and storage manufacturers to evaluate cooperation paths.
In addition to external cooperation, DeepSeek is also quietly increasing its recruitment of chip design talents. Two people familiar with the matter said that the company has recently hired a number of additional chip architecture and design engineers, but the relevant positions are not announced through public recruitment platforms, but are conducted internally and in a targeted manner to maintain a low-key operation of the project. Since the information has not yet been made public, these people requested anonymity. DeepSeek itself has always maintained a low profile and did not respond to external inquiries.
Self-developed chips are not an easy project for any AI company, involving high capital investment, long research and development cycles, and complex industrial chain collaboration. In the manufacturing process, Chinese chip design companies still face U.S. restrictions on advanced process wafer foundry services, and they also encounter export controls on key components such as high-bandwidth storage. These may bring additional difficulties to the development and mass production of high-performance inference chips.
At the capital level, DeepSeek’s chip plan is advancing almost simultaneously with changes in the company’s financing path. For many years, the Hangzhou-based company has insisted on rejecting external equity investment and is regarded as a rare "self-financing" AI company in the industry. However, people familiar with the matter said that DeepSeek has recently launched an initial round of financing of about US$7 billion, with a valuation range between US$52 billion and US$59 billion. This change is interpreted as the company's preparations to support larger-scale computing power and chip research and development.
Globally, leading AI companies such as OpenAI and Anthropic have begun to explore or evaluate self-developed inference chips in order to take greater initiative in performance optimization, cost control and supply security. For Chinese companies facing export controls, independent and controllable chip solutions also have stronger strategic significance: they are not only a response to external restrictions, but also an important link in the improvement of the domestic technology industry chain. Many industry observers believe that DeepSeek’s move into the chip field may further reshape the technical route and competitive landscape of China’s AI market.