DeepSeek may be developing its own AI chip

According to Reuters, DeepSeek's chip development project was actually launched a year ago. Unlike the general-purpose GPU line, this chip is specifically optimized for the inference phase, the process where the AI ​​model has finished processing data and responding to user requests, instead of being used to train new models.

If successful, the move will mark an important strategic shift for DeepSeek, a startup known as "China's AI champion" thanks to its philosophy of focusing on algorithmic breakthroughs instead of software commercialization.

This move will also create more direct pressure on Huawei, the technology giant that currently holds nearly half of China's $50 billion AI chip market by taking advantage of the gap after the US export ban. Before DeepSeek, other large technology companies such as Alibaba and Baidu also had autonomous supply of AI chips and began to gradually take over Huawei's market share.

Currently the project is still in the early stages. The company is actively contacting and discussing with design partners, semiconductor foundries and memory chip suppliers. According to sources, this startup has also increased recruitment of chip design engineers in the past few months, but all processes are carried out confidentially, not posted publicly on traditional recruitment platforms.

DeepSeek's decision to make its own chips fits a global trend by major AI developers to gain control over hardware and reduce dependence on Nvidia. Last month, OpenAI unveiled Jalapeno, its first custom inference chip in partnership with Broadcom, while Anthropic is also considering a similar path.

For DeepSeek, the project is also vital, as Liang Wenfeng, founder of DeepSeek, once admitted that US chip export control measures are a big challenge for the company.

DeepSeek's R1 model, which caused a series of US technology stocks to plummet in January 2025, was originally trained on Nvidia H800 - a chip line refined for the Chinese market before being banned by the US at the end of 2023. After that, DeepSeek switched to relying on Huawei's Ascend chip to train updates such as V4 or V4-Flash.

Although the inference chip market is the most fertile and growing land due to increasingly popular AI applications, analysts say that DeepSeek's journey will not be easy. Designing a competitive chip requires many years and huge capital.

This startup's biggest barrier lies in the production stage, when the US prohibits advanced chip foundries around the world from accepting orders from China, while also restricting access to high-bandwidth memory chips, core components of AI inference chips.

However, DeepSeek's abandonment of the principle of "saying no to foreign capital" to open a record capital raising round of 7 billion USD, helping to value the company at 52-59 billion USD last month, is considered a solid financial preparation step for this ambitious semiconductor strategy.