Foreign media: Chinese AI startup DeepSeek is independently developing AI chips designed specifically for inference rather than model training, a project initiated about a year ago and currently in its early stages. The company is now negotiating partnerships with external chip design, manufacturing, and memory storage firms, and has quietly increased hiring of chip design engineers recently.
If successful, this move would mark a significant strategic shift for DeepSeek, reducing its reliance on chips from NVIDIA and Huawei, while simultaneously intensifying competition with Huawei in China's AI chip market.
Although Huawei holds roughly half of China’s $50 billion domestic AI chip market due to U.S. sanctions, it is now facing increasing challenges from rivals such as Alibaba and Baidu. DeepSeek was previously renowned for breakthroughs in model efficiency; its V4 model has already been adapted to Huawei’s Ascend chips. The company is also planning a fundraising round of $7 billion, with an estimated valuation ranging between $52 billion and $59 billion.
On a global scale, companies like OpenAI and Anthropic are also advancing their own custom chip development, but DeepSeek faces manufacturing bottlenecks, including restricted access to advanced process node fabrication and high-bandwidth memory.
Original article: toutiao.com/article/1870064520197120/
Disclaimer: The views expressed in this article are solely those of the author.