The amount of data is only low for inference, not for training, and AFAIK DGX spark is supposed to be a researcher's machine that can do small-scale training.