Chinese AI startup DeepSeek is developing its first in-house AI chip designed for inference, reported Reuters, citing three people familiar with the matter.
According to Reuters, the project began about a year ago and remains in its early stages. DeepSeek has been holding discussions with chip design firms, foundries, and memory suppliers while privately expanding its chip engineering team.
AIM could not independently verify the report.
The AI company is looking to reduce its dependence on NVIDIA and Huawei processors, which DeepSeek has relied on to train and deploy its AI models.
DeepSeek has used both NVIDIA and Huawei hardware across different model generations. The company previously said its R1 reasoning model was trained using NVIDIA’s H800 GPUs before those chips were restricted for China.
More recently, it has shifted further towards Huawei’s ecosystem, releasing its V4 model adapted for Huawei’s Ascend chips, with Huawei stating that its processors were used for part of the training of the lighter V4-Flash model.
Currently, NVIDIA’s H20 is the only advanced AI GPU that can be legally exported to China, and even then, shipments require US export licenses.
The H20 was specifically designed to comply with US export controls, but last April, it was made subject to an indefinite licensing requirement before the US later agreed to approve licenses for eligible exports.
More powerful AI GPUs, including the H100, H200, B200, GB200, and other Blackwell products, remain under current US export control rules.
The development comes as the Chinese AI company seeks its first external capital. DeepSeek recently closed a $7 billion funding round, valuing the company at over $50 billion and marking a departure from its long-standing policy of rejecting outside investment.
DeepSeek’s hardware ambitions also reflect China’s broader push to develop domestic AI infrastructure as US export controls continue to restrict access to NVIDIA’s most advanced processors.
Huawei has emerged as the country’s largest AI chip supplier with nearly 50% market share, while Alibaba, through its T-Head unit, and Baidu are developing proprietary AI processors for their own cloud platforms.
A new generation of independent chip companies, including Biren Technology, Iluvatar CoreX, Moore Threads, and MetaX, has also expanded aggressively through IPOs and fresh fundraising, targeting AI training and inference workloads as US export controls accelerate demand for domestic alternatives.
Your reaction
Discussion
What's your take on this story?
No comments yet — be the first to share your view.