It's not such a problem, except if you want to train from scratch a large model (NLP or CV), not if you want to fine-tune it for a related task. So one trained model can be reused many times. In general training data is scarce, only in a few situations it is abundant.