One-sentence summary
Using DeepSeek V5's planned Ascend-based training as an example: understand the gap between inference and training demands on chips, and ecosystem migration cost as the key variable.
The Difference Between Inference and Training
Inference answers with a trained model, sensitive to per-card compute and low latency; training learns a model from data, demanding far more from cluster interconnect, memory bandwidth and framework optimization.
Why Ascend Was Used for Inference First
Inference tasks are more standardized with controllable migration cost; training needs mature framework adaptation, operator libraries and cluster scheduling, where ecosystem maturity is the entry barrier.
What Ecosystem Migration Cost Includes
Framework migration, operator development and tuning, cluster interconnect and stability, and operational tooling. These costs set the ceiling for moving from inference to training.