Whether domestic compute can carry large-model training has been the key threshold from "usable" to "good enough." Market reports say DeepSeek V5 will be pretrained on Huawei Ascend chips — and training-grade workloads demand far more than inference in chip performance, interconnect, and cluster scale, so a top model placing pretraining on domestic silicon crosses the hardest barrier.
Why Training-Grade Compute Is Harder Than Inference
Inference can run on single cards or small clusters, but pretraining needs tens of thousands of stable interconnected accelerators, low-latency collective communication, and long error-free runs; any weakness stretches the training cycle. If Ascend truly hosts a top model's pretraining, it signals that cluster scheduling, operator compatibility, and stability are approaching the usable threshold, with servers, optical modules, liquid cooling, and interconnect chips all benefiting.
A Step-Change Channel for the Domestic Compute Chain
A training-grade breakthrough is not just a one-off order but a volume logic for the whole chain: from chips and boards to rack-level delivery, from software stacks to model adaptation, domestic substitution moves from edge inference to core training. Meanwhile AMD's data-center sales nearly doubled in Q3 and Microsoft and Meta held high capex guidance, with global AI-compute momentum continuously validated by earnings, supporting domestic assets through overseas mapping.
What to watch: whether the DeepSeek V5–Ascend tie-up is officially confirmed, the measured throughput and availability of domestic training clusters, and whether supporting interconnect and advanced-packaging capacity can keep up with training-grade scaling.