Chinese AI startup DeepSeek is developing its own custom inference chip, marking a strategic pivot toward hardware independence as US export controls tighten access to advanced semiconductors.
The project, confirmed by three people familiar with the matter and first reported by Reuters, has been underway for roughly a year. DeepSeek has been quietly building a chip design team through private recruiting channels and holding discussions with design firms, foundries, and memory suppliers.
Inference, not training. The chip is designed specifically for inference, the stage where a trained AI model generates responses for users, rather than for training new models. As AI applications move from research labs into production at scale, inference has become the dominant source of compute demand, creating a growing market for specialized efficiency-oriented processors.
DeepSeek previously relied on Nvidia’s H800 GPUs, chips designed for the Chinese market before Washington imposed broader restrictions. The company has since shifted much of its AI workload to Huawei’s Ascend processors. Founder Liang Wenfeng acknowledged in a 2024 interview that US export controls had become a major obstacle for the company.
The move places DeepSeek alongside a growing roster of AI companies designing custom silicon. OpenAI unveiled Jalapeno, its first inference chip built with Broadcom, last month. Anthropic has also been exploring chip development. In China, Alibaba and Baidu have invested heavily in their own AI processors.
Nvidia shares dipped roughly 1.6% in premarket trading following the news, reflecting investor attention to the broader industry shift away from reliance on a single chip supplier.
The effort remains early-stage. Developing a competitive AI processor requires years of engineering work, billions in investment, and access to advanced manufacturing, all of which remain constrained for Chinese companies under current export regimes.