He's saying there is a niche where models are too large to practically run on 5090, but are still runnable on DGX Spark.

The larger memory also allows for pre-training / finetuning models, hence why it's aimed at developers.