Reducing dependence on NVIDIA
According to Wccftech, DeepSeek is preparing to train its upcoming V5 model from scratch on Huawei's Atlas SuperPoD/Supernode systems built around Ascend 950DT chips. The company reportedly ordered around 160,000 Ascend 950DT chips, with deliveries expected to ramp up gradually.
Why it matters
For Chinese AI labs, the availability of NVIDIA hardware has been volatile due to export controls. DeepSeek's move is seen as a systematic effort to reduce reliance on US chips.
The domestic stack claim
The report says the training run will be supported by 'TileLang', a CUDA-like kernel compiler. If successful, it would be China's first AI training run completed on a fully domestic hardware and software stack.
A cautious reading
The information rests on a single outlet's 'reportedly' framing; there is no official confirmation from DeepSeek or Huawei. The chip count, delivery timeline and scope of the training run have not been independently verified.
