Training a model and deploying it to consumers to infer from it are two different things. NVidia will remain in demand for training, while deployment will use c

Training a model and deploying it to consumers to infer from it are two different things. NVidia will remain in demand for training, while deployment will use cheaper hardware on the final model to serve up to users. What am I missing?