Chinese AI startup DeepSeek, known for its advanced models such as R1 and V4, is making a decisive move toward vertical integration. According to my data, the company has been working for about a year on creating its own processor designed for inference — the stage of executing models, not training them. This is a strategically important move, given the current geopolitical realities and export restrictions from the United States.
The development is in its early stages. DeepSeek is actively engaging with external partners in chip design, contract manufacturing, and memory. In recent months, the company has been quietly hiring chip design engineers, confirming the seriousness of its intentions. There have been no official comments from DeepSeek yet, but insider information indicates that the project has already moved into a phase of active discussions with key suppliers.
This step is a direct response to dependence on Nvidia and Huawei. Previously, DeepSeek used Nvidia H800 to train the R1 model, but later shifted its focus to Huawei Ascend chips, adapting the V4 and V4-Flash models for them. However, its own hardware gives the company not only independence but also the ability to optimize the architecture for its unique algorithms, which is critical for reducing costs and improving performance.
This is happening against the backdrop of growing demand for inference solutions and tightening export controls. DeepSeek founder Liang Wenfeng as early as 2024 called chip restrictions one of the main problems. Now the company is trying to solve it on its own, entering the race for its own AI infrastructure alongside giants such as OpenAI, Broadcom, Anthropic, Alibaba, and Baidu.
However, the project's success is not guaranteed. Competition is enormous, requiring massive investments, and access to advanced foreign fabs is limited. Nevertheless, DeepSeek has already demonstrated its ability to innovate by creating Code Harness, a tool for autonomous programming that challenges Anthropic's Claude Code and OpenAI's Codex. If the company manages to overcome technological and logistical barriers, it could fundamentally change the balance of power in the AI chip market.
Analytical conclusion: Creating its own chip is a logical but risky step for DeepSeek. Success here depends not only on engineering talent but also on the ability to circumvent sanctions restrictions. If the project is realized, DeepSeek will become one of the few players capable of competing with Nvidia on equal footing in the inference segment. This event could trigger a new wave of decentralization in AI infrastructure.