One-sentence summary
FAQ on DeepSeek V5's planned Ascend-based training: reported by CLS citing industry sources and not officially confirmed; the new step is moving from inference to training.
Is the report confirmed?
CLS reported on November 3 that multiple supply-chain sources said DeepSeek V5 is planned to be trained on Ascend; no official DeepSeek announcement had been seen as of writing.
Why Ascend?
Ascend has been validated at scale for DeepSeek's inference workloads, and domestic chips offer supply security and autonomy advantages.
How is this different from before?
Ascend previously served inference only; the plan would bring Ascend into training, which demands far more from cluster interconnect, framework adaptation and operator optimization.
What are the main challenges?
Ecosystem migration cost on the training side (framework migration, operator development, cluster scheduling and stability) is the key variable.