Training a model means defining your own parameters based on data you supply. As such that would require a massive dataset and also 100s of H100 Nvidia GPUs to churn that data into model parameters.
...
Maybe he meant specializing a model, instead of doing a full training from scratch.