High-end ones are needed more for large model training and unoptimized tuning, less so for inference alone. Most users would need just inference, at least in the beginning.