DeepSeek starts designing in-house AI inference chip

The Chinese artificial intelligence startup DeepSeek is developing its own semiconductor for AI inference, according to multiple reports, a strategic move aimed at reducing its reliance on external chip suppliers like Nvidia and Huawei. The project, which sources say began about a year ago, is still in its early stages and focuses on designing a chip specifically for running trained AI models to answer user queries.

This initiative marks a significant shift for the Hangzhou-based company, which rose to global prominence last year with the release of highly efficient AI models. DeepSeek has historically depended on a mix of hardware, initially using Nvidia's H800 GPUs before U.S. export restrictions blocked access to the most advanced chips. The company subsequently shifted much of its workload to Huawei's Ascend processors. Developing an in-house inference chip is seen as a step toward greater independence and long-term control over its AI infrastructure.

The effort aligns with a broader industry trend where leading AI developers, including OpenAI, Meta, and Microsoft, are pursuing custom hardware to lower costs and optimize performance. For DeepSeek, the push also carries a...