The Hangzhou-based startup has been hiring chip-design engineers privately and holding talks with foundry and memory companies for roughly a year, targeting the fast-growing inference market
DeepSeek, the Chinese AI startup that rattled Silicon Valley when its efficient models went viral early last year, is now working on something beyond software. According to three people familiar with the matter, the company has been developing its own AI chip for roughly a year, a move that would mark a fundamental strategic expansion for a firm that has until now focused almost exclusively on model breakthroughs.
What Kind of Chip and Why
The device is optimized for the inference stage of machine learning, which refers to the point of AI model deployment to the end-users instead of the complex part of creating an AI model from the start. The inference chips are cheaper than general-purpose GPUs, and their number is increasing due to the increasing use of AI technologies in daily life.
The Supply Chain Problem DeepSeek Is Trying to Solve
DeepSeek has depended on both Nvidia and Huawei chips to train and run its models. US export controls have effectively shut Chinese companies out of Nvidia’s most advanced offerings, pushing DeepSeek increasingly toward Huawei’s Ascend processors. But Huawei’s chips still lag Nvidia’s by a significant margin, and its hold on China’s domestic AI chip market is already being challenged as Alibaba and Baidu develop their own silicon.
Building an in-house chip would reduce DeepSeek’s dependence on both suppliers while aligning with Beijing’s broader push for domestic technology self-sufficiency.
Early Stage, Real Obstacles
The effort remains at an early stage. DeepSeek has been reaching out to external partners and holding discussions with chip-design, foundry, and memory companies, while quietly hiring chip-design engineers without advertising positions on public platforms. Manufacturing presents a significant hurdle, as US restrictions bar Chinese designers from accessing the most advanced overseas foundries and limit China’s access to high-bandwidth memory, a critical component for inference chips.
The push coincides with DeepSeek’s first move to raise outside capital, with a reported $7 billion funding round valuing the company at between $52 billion and $59 billion.
