DeepSeek and Huawei forge a deep partnership to build “Nvidia-free” AI chips
According to The Information, DeepSeek is working with Huawei in a partnership that goes far beyond a typical customer relationship. It is directly involved in design decisions for next-generation AI chips, with a clear goal: break free from reliance on Nvidia and train the most advanced AI models using domestically made chips.
Huawei engineers have been stationed at DeepSeek’s offices for an extended period, tailoring chip designs to the model computation and data-transfer requirements specified by DeepSeek. The resulting chips are intended for use in AI chips launched in 2027 and beyond.
Under the current plan, DeepSeek could start using Huawei’s Ascend 950DT chips for model pre-training as soon as the end of this year, though testing is still underway. More significantly, Huawei plans to allocate most of the supply of its Ascend 960DT chips, scheduled for launch in 2027, to DeepSeek rather than distributing it evenly among other Chinese AI companies. DeepSeek will use the chips both to train and run its models.
On the software side, DeepSeek has also adopted the open-source programming tool TileLang to reduce its reliance on Nvidia’s CUDA software ecosystem. These moves suggest that the two companies are working to build a fully independent AI computing pathway, from hardware to software.
#FollowForFollow #FollowBack
According to The Information, DeepSeek is working with Huawei in a partnership that goes far beyond a typical customer relationship. It is directly involved in design decisions for next-generation AI chips, with a clear goal: break free from reliance on Nvidia and train the most advanced AI models using domestically made chips.
Huawei engineers have been stationed at DeepSeek’s offices for an extended period, tailoring chip designs to the model computation and data-transfer requirements specified by DeepSeek. The resulting chips are intended for use in AI chips launched in 2027 and beyond.
Under the current plan, DeepSeek could start using Huawei’s Ascend 950DT chips for model pre-training as soon as the end of this year, though testing is still underway. More significantly, Huawei plans to allocate most of the supply of its Ascend 960DT chips, scheduled for launch in 2027, to DeepSeek rather than distributing it evenly among other Chinese AI companies. DeepSeek will use the chips both to train and run its models.
On the software side, DeepSeek has also adopted the open-source programming tool TileLang to reduce its reliance on Nvidia’s CUDA software ecosystem. These moves suggest that the two companies are working to build a fully independent AI computing pathway, from hardware to software.
#FollowForFollow #FollowBack