With these in-memory DBs, scaling beyond 1TB becomes prohibitively expensive for most companies.
Note that MapD is not fully in-memory. It can pull from CPU Memory and SSD as necessary. So its essentially as fast as the fastest disk-based database at its worst and likely two orders-of-magnitude or more faster in many common scenarios where certain subsets of the data are repeatedly queries (for example, BI workloads or when repeatedly pulling out subsets of a dataset to train a model)
The catch is GPU memory is more expensive than CPU memory. You only want to use it if: 1) Your workload is compute heavy - think machine learning OR 2) You care about speed i.e., you want queries to be interactive.