The RAM choice was because I have never regretted buying more RAM - it's practically always a better trade than a slightly faster CPU - and 96GB DIMMs were at a sweet spot compared to 128GB DIMMs.
That, and the ability to have big LLMs in memory, for some local inference, even if it's slow mixed CPU/GPU inference, or paged on demand. And if not for big LLMs, then to keep models cached for quick swapping.