Chinese AI startup DeepSeek is developing a processor for its models, sources told Reuters. According to them, work began about a year ago.
According to the agency, the chip is designed for inference rather than training new models. The development is currently in its early stages: DeepSeek is discussing the project with external partners for chip design, contract manufacturing, and memory.
In recent months, the company has been quietly hiring chip design engineers. DeepSeek did not respond to Reuters' request for comment.
This move could reduce reliance on Nvidia and Huawei. The company has previously utilized their solutions: the team mentioned that the base model for R1 was trained on the Nvidia H800, made for the Chinese market.
Later, the startup increasingly relied on chips from Huawei. In April, DeepSeek introduced the V4 model, adapted for Ascend. Huawei stated that its processors were used in part of the training for V4-Flash.
The attempt to create its own hardware comes amid U.S. export restrictions and a growing demand for inference solutions. DeepSeek's founder, Liang Wenfeng, noted in 2024 that export controls on chips posed a challenge for the company.
Currently, the race for proprietary AI infrastructure is led by OpenAI, Broadcom, Anthropic, Alibaba, and Baidu. Reuters emphasized that the project's success is not guaranteed due to high competition, significant investments, and limited access to advanced foreign factories.
In May, it was reported that DeepSeek is forming a new team to develop Code Harness—a tool for autonomous programming. This will directly compete with Claude Code from Anthropic and Codex from OpenAI.