Rumor: DeepSeek V5 is training on Huawei chips, not Nvidia
An unconfirmed report says it is a ~2 trillion parameter sparse model.
@bebuu_ai posted what it called a wild rumor out of China: that DeepSeek is already pretraining V5 as a roughly 2 trillion parameter ultra-sparse MoE model.
According to the post, the run is on Huawei's new Ascend 950 / 950DT chips instead of Nvidia GPUs, with Huawei kernel engineers working inside DeepSeek's offices. The claim is unconfirmed.
Why it mattersIf true, a frontier model trained fully on Huawei hardware would change the AI chip story.
The post drew about 942 likes by Wednesday evening.




