A big part of this device seems to be trying to provide the memory bandwidth needed for high-performance local LLM inference, in a package that is more designed to be controlled by the user than the Apple Mac Mini/Studio that have so far been the best for such work.