Not sure I understand exactly which trade-off you're referring to, but on systems without the GPU-handicapping (e.g. IBM Power; and also when you put link up many GPUs together with NVLink) - there is still a significant design and implementation challenge to produce a full-fledged analytic DBMS, competitive vis-a-vis the state-of-the-art CPU-based systems.

There are also other considerations such as: The desire to combine analytics and transactions; performance-per-Watt rather than per-processor; performance-per-cubic-meter; existing deployed cluster hardware; vendor lock-in risk; etc.