Chinese AI startup DeepSeek has quietly developed its own custom inference chip over the past year, media reports. DeepSeek spent approximately twelve months building the chip in secrecy without public announcements, representing a major strategic shift for a company previously focused solely on AI model development.
DeepSeek targets inference rather than training with the new chip. Inference is the stage where trained AI models generate responses to user queries. This segment represents the fastest-growing area of AI computing as millions of users interact with chatbots, coding assistants, and enterprise applications daily. Inference chips require specialization for efficiency and power consumption rather than the raw performance needed for training.
The project reflects DeepSeek’s push to reduce dependence on external hardware suppliers. US export controls bar Chinese companies from purchasing NVIDIA’s most advanced processors.
DeepSeek founder Liang Wenfeng acknowledged chip export controls posed challenges to the company’s operations in a 2024 interview. The company has increasingly relied on Huawei’s Ascend processors for newer models after Nvidia H800 chips faced export restrictions.
The Chinese AI company, which shocked the world upon its AI reveal, recruits chip design engineers quietly rather than through public job postings. The company holds discussions with chip designers, semiconductor foundries, and memory suppliers to develop the project. No formal announcements have confirmed the effort, and sources requested anonymity due to the project’s sensitivity.
Designing a competitive AI chip demands years of engineering work and billions in capital investment. US restrictions ban Chinese chip designers from accessing the most advanced overseas foundries. Additional US curbs restrict China’s access to high-bandwidth memory critical for AI inference chips.
A DeepSeek inference chip is designed to tap into the rapidly expanding segment of AI computing demand. As AI applications become more widespread, there’s a noticeable shift in the industry from training models to actually running them. This transition relies on specialized chips that are often more cost-effective and energy-efficient compared to traditional general-purpose GPUs.
