They could certainly pick a previous-generation GPU architecture, and get it integrated their way with Rust drivers and whatever else their technology fetishism is, it wouldn't matter because ultimately people want to run software, and right now all the AI software is targeting NVIDIA SDK as a first-class citizen, and AMD SDK as a distant second-class citizen.

So their story is partially correct: get our systems for general purpose and buy GPUs for the rest. It's just that it kind of self-defeats their entire raison d'etre, which is to have a single highly-integrated vendor. Now turns out you still need to deal with the messy world of GPU computing.